Upload 19 files

.gitattributes

@@ -6,3 +6,4 @@ assets/visual_results/bsr6.png filter=lfs diff=lfs merge=lfs -text
 assets/visual_results/tiled_sampling.png filter=lfs diff=lfs merge=lfs -text
 assets/visual_results/whole_image1.png filter=lfs diff=lfs merge=lfs -text
 assets/visual_results/whole_image2.png filter=lfs diff=lfs merge=lfs -text
+model/open_clip/bpe_simple_vocab_16e6.txt.gz filter=lfs diff=lfs merge=lfs -text

model/__init__.py

@@ -0,0 +1,12 @@
+from . import config
+
+from .controlnet import ControlledUnetModel, ControlNet
+from .vae import AutoencoderKL
+from .clip import FrozenOpenCLIPEmbedder
+
+from .cldm import ControlLDM
+from .gaussian_diffusion import Diffusion
+
+from .swinir import SwinIR
+from .bsrnet import RRDBNet
+from .scunet import SCUNet

model/attention.py

@@ -0,0 +1,298 @@
+from packaging import version
+import torch
+import torch.nn.functional as F
+from torch import nn, einsum
+from einops import rearrange, repeat
+from typing import Optional, Any
+
+from model.util import (
+    checkpoint, zero_module, exists, default
+)
+from model.config import Config, AttnMode
+
+
+# CrossAttn precision handling
+import os
+_ATTN_PRECISION = os.environ.get("ATTN_PRECISION", "fp32")
+
+
+# feedforward
+class GEGLU(nn.Module):
+    def __init__(self, dim_in, dim_out):
+        super().__init__()
+        self.proj = nn.Linear(dim_in, dim_out * 2)
+
+    def forward(self, x):
+        x, gate = self.proj(x).chunk(2, dim=-1)
+        return x * F.gelu(gate)
+
+
+class FeedForward(nn.Module):
+    def __init__(self, dim, dim_out=None, mult=4, glu=False, dropout=0.):
+        super().__init__()
+        inner_dim = int(dim * mult)
+        dim_out = default(dim_out, dim)
+        project_in = nn.Sequential(
+            nn.Linear(dim, inner_dim),
+            nn.GELU()
+        ) if not glu else GEGLU(dim, inner_dim)
+
+        self.net = nn.Sequential(
+            project_in,
+            nn.Dropout(dropout),
+            nn.Linear(inner_dim, dim_out)
+        )
+
+    def forward(self, x):
+        return self.net(x)
+
+
+def Normalize(in_channels):
+    return torch.nn.GroupNorm(num_groups=32, num_channels=in_channels, eps=1e-6, affine=True)
+
+
+class CrossAttention(nn.Module):
+    def __init__(self, query_dim, context_dim=None, heads=8, dim_head=64, dropout=0.):
+        super().__init__()
+        print(f"Setting up {self.__class__.__name__} (vanilla). Query dim is {query_dim}, context_dim is {context_dim} and using "
+              f"{heads} heads.")
+        inner_dim = dim_head * heads
+        context_dim = default(context_dim, query_dim)
+
+        self.scale = dim_head ** -0.5
+        self.heads = heads
+
+        self.to_q = nn.Linear(query_dim, inner_dim, bias=False)
+        self.to_k = nn.Linear(context_dim, inner_dim, bias=False)
+        self.to_v = nn.Linear(context_dim, inner_dim, bias=False)
+
+        self.to_out = nn.Sequential(
+            nn.Linear(inner_dim, query_dim),
+            nn.Dropout(dropout)
+        )
+
+    def forward(self, x, context=None, mask=None):
+        h = self.heads
+
+        q = self.to_q(x)
+        context = default(context, x)
+        k = self.to_k(context)
+        v = self.to_v(context)
+
+        q, k, v = map(lambda t: rearrange(t, 'b n (h d) -> (b h) n d', h=h), (q, k, v))
+
+        # force cast to fp32 to avoid overflowing
+        if _ATTN_PRECISION =="fp32":
+            # with torch.autocast(enabled=False, device_type = 'cuda'):
+            with torch.autocast(enabled=False, device_type="cuda" if str(x.device).startswith("cuda") else "cpu"):
+                q, k = q.float(), k.float()
+                sim = einsum('b i d, b j d -> b i j', q, k) * self.scale
+        else:
+            sim = einsum('b i d, b j d -> b i j', q, k) * self.scale
+        
+        del q, k
+    
+        if exists(mask):
+            mask = rearrange(mask, 'b ... -> b (...)')
+            max_neg_value = -torch.finfo(sim.dtype).max
+            mask = repeat(mask, 'b j -> (b h) () j', h=h)
+            sim.masked_fill_(~mask, max_neg_value)
+
+        # attention, what we cannot get enough of
+        sim = sim.softmax(dim=-1)
+
+        out = einsum('b i j, b j d -> b i d', sim, v)
+        out = rearrange(out, '(b h) n d -> b n (h d)', h=h)
+        return self.to_out(out)
+
+
+class MemoryEfficientCrossAttention(nn.Module):
+    # https://github.com/MatthieuTPHR/diffusers/blob/d80b531ff8060ec1ea982b65a1b8df70f73aa67c/src/diffusers/models/attention.py#L223
+    def __init__(self, query_dim, context_dim=None, heads=8, dim_head=64, dropout=0.0):
+        super().__init__()
+        print(f"Setting up {self.__class__.__name__} (xformers). Query dim is {query_dim}, context_dim is {context_dim} and using "
+              f"{heads} heads.")
+        inner_dim = dim_head * heads
+        context_dim = default(context_dim, query_dim)
+
+        self.heads = heads
+        self.dim_head = dim_head
+
+        self.to_q = nn.Linear(query_dim, inner_dim, bias=False)
+        self.to_k = nn.Linear(context_dim, inner_dim, bias=False)
+        self.to_v = nn.Linear(context_dim, inner_dim, bias=False)
+
+        self.to_out = nn.Sequential(nn.Linear(inner_dim, query_dim), nn.Dropout(dropout))
+        self.attention_op: Optional[Any] = None
+
+    def forward(self, x, context=None, mask=None):
+        q = self.to_q(x)
+        context = default(context, x)
+        k = self.to_k(context)
+        v = self.to_v(context)
+
+        b, _, _ = q.shape
+        q, k, v = map(
+            lambda t: t.unsqueeze(3)
+            .reshape(b, t.shape[1], self.heads, self.dim_head)
+            .permute(0, 2, 1, 3)
+            .reshape(b * self.heads, t.shape[1], self.dim_head)
+            .contiguous(),
+            (q, k, v),
+        )
+
+        # actually compute the attention, what we cannot get enough of
+        out = Config.xformers.ops.memory_efficient_attention(q, k, v, attn_bias=None, op=self.attention_op)
+        
+        if exists(mask):
+            raise NotImplementedError
+        out = (
+            out.unsqueeze(0)
+            .reshape(b, self.heads, out.shape[1], self.dim_head)
+            .permute(0, 2, 1, 3)
+            .reshape(b, out.shape[1], self.heads * self.dim_head)
+        )
+        return self.to_out(out)
+
+
+class SDPCrossAttention(nn.Module):
+    def __init__(self, query_dim, context_dim=None, heads=8, dim_head=64, dropout=0.0):
+        super().__init__()
+        print(f"Setting up {self.__class__.__name__} (sdp). Query dim is {query_dim}, context_dim is {context_dim} and using "
+              f"{heads} heads.")
+        inner_dim = dim_head * heads
+        context_dim = default(context_dim, query_dim)
+
+        self.heads = heads
+        self.dim_head = dim_head
+
+        self.to_q = nn.Linear(query_dim, inner_dim, bias=False)
+        self.to_k = nn.Linear(context_dim, inner_dim, bias=False)
+        self.to_v = nn.Linear(context_dim, inner_dim, bias=False)
+
+        self.to_out = nn.Sequential(nn.Linear(inner_dim, query_dim), nn.Dropout(dropout))
+
+    def forward(self, x, context=None, mask=None):
+        q = self.to_q(x)
+        context = default(context, x)
+        k = self.to_k(context)
+        v = self.to_v(context)
+
+        b, _, _ = q.shape
+        q, k, v = map(
+            lambda t: t.unsqueeze(3)
+            .reshape(b, t.shape[1], self.heads, self.dim_head)
+            .permute(0, 2, 1, 3)
+            .reshape(b * self.heads, t.shape[1], self.dim_head)
+            .contiguous(),
+            (q, k, v),
+        )
+
+        # actually compute the attention, what we cannot get enough of
+        out = F.scaled_dot_product_attention(q, k, v)
+        
+        if exists(mask):
+            raise NotImplementedError
+        out = (
+            out.unsqueeze(0)
+            .reshape(b, self.heads, out.shape[1], self.dim_head)
+            .permute(0, 2, 1, 3)
+            .reshape(b, out.shape[1], self.heads * self.dim_head)
+        )
+        return self.to_out(out)
+
+
+class BasicTransformerBlock(nn.Module):
+    ATTENTION_MODES = {
+        AttnMode.VANILLA: CrossAttention,  # vanilla attention
+        AttnMode.XFORMERS: MemoryEfficientCrossAttention,
+        AttnMode.SDP: SDPCrossAttention
+    }
+    def __init__(self, dim, n_heads, d_head, dropout=0., context_dim=None, gated_ff=True, checkpoint=True,
+                 disable_self_attn=False):
+        super().__init__()
+        attn_cls = self.ATTENTION_MODES[Config.attn_mode]
+        self.disable_self_attn = disable_self_attn
+        self.attn1 = attn_cls(query_dim=dim, heads=n_heads, dim_head=d_head, dropout=dropout,
+                              context_dim=context_dim if self.disable_self_attn else None)  # is a self-attention if not self.disable_self_attn
+        self.ff = FeedForward(dim, dropout=dropout, glu=gated_ff)
+        self.attn2 = attn_cls(query_dim=dim, context_dim=context_dim,
+                              heads=n_heads, dim_head=d_head, dropout=dropout)  # is self-attn if context is none
+        self.norm1 = nn.LayerNorm(dim)
+        self.norm2 = nn.LayerNorm(dim)
+        self.norm3 = nn.LayerNorm(dim)
+        self.checkpoint = checkpoint
+
+    def forward(self, x, context=None):
+        return checkpoint(self._forward, (x, context), self.parameters(), self.checkpoint)
+
+    def _forward(self, x, context=None):
+        x = self.attn1(self.norm1(x), context=context if self.disable_self_attn else None) + x
+        x = self.attn2(self.norm2(x), context=context) + x
+        x = self.ff(self.norm3(x)) + x
+        return x
+
+
+class SpatialTransformer(nn.Module):
+    """
+    Transformer block for image-like data.
+    First, project the input (aka embedding)
+    and reshape to b, t, d.
+    Then apply standard transformer action.
+    Finally, reshape to image
+    NEW: use_linear for more efficiency instead of the 1x1 convs
+    """
+    def __init__(self, in_channels, n_heads, d_head,
+                 depth=1, dropout=0., context_dim=None,
+                 disable_self_attn=False, use_linear=False,
+                 use_checkpoint=True):
+        super().__init__()
+        if exists(context_dim) and not isinstance(context_dim, list):
+            context_dim = [context_dim]
+        self.in_channels = in_channels
+        inner_dim = n_heads * d_head
+        self.norm = Normalize(in_channels)
+        if not use_linear:
+            self.proj_in = nn.Conv2d(in_channels,
+                                     inner_dim,
+                                     kernel_size=1,
+                                     stride=1,
+                                     padding=0)
+        else:
+            self.proj_in = nn.Linear(in_channels, inner_dim)
+
+        self.transformer_blocks = nn.ModuleList(
+            [BasicTransformerBlock(inner_dim, n_heads, d_head, dropout=dropout, context_dim=context_dim[d],
+                                   disable_self_attn=disable_self_attn, checkpoint=use_checkpoint)
+                for d in range(depth)]
+        )
+        if not use_linear:
+            self.proj_out = zero_module(nn.Conv2d(inner_dim,
+                                                  in_channels,
+                                                  kernel_size=1,
+                                                  stride=1,
+                                                  padding=0))
+        else:
+            self.proj_out = zero_module(nn.Linear(in_channels, inner_dim))
+        self.use_linear = use_linear
+
+    def forward(self, x, context=None):
+        # note: if no context is given, cross-attention defaults to self-attention
+        if not isinstance(context, list):
+            context = [context]
+        b, c, h, w = x.shape
+        x_in = x
+        x = self.norm(x)
+        if not self.use_linear:
+            x = self.proj_in(x)
+        x = rearrange(x, 'b c h w -> b (h w) c').contiguous()
+        if self.use_linear:
+            x = self.proj_in(x)
+        for i, block in enumerate(self.transformer_blocks):
+            x = block(x, context=context[i])
+        if self.use_linear:
+            x = self.proj_out(x)
+        x = rearrange(x, 'b (h w) c -> b c h w', h=h, w=w).contiguous()
+        if not self.use_linear:
+            x = self.proj_out(x)
+        return x + x_in

model/bsrnet.py

@@ -0,0 +1,104 @@
+# From BSRGAN
+import functools
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+import torch.nn.init as init
+
+
+def initialize_weights(net_l, scale=1):
+    if not isinstance(net_l, list):
+        net_l = [net_l]
+    for net in net_l:
+        for m in net.modules():
+            if isinstance(m, nn.Conv2d):
+                init.kaiming_normal_(m.weight, a=0, mode='fan_in')
+                m.weight.data *= scale  # for residual block
+                if m.bias is not None:
+                    m.bias.data.zero_()
+            elif isinstance(m, nn.Linear):
+                init.kaiming_normal_(m.weight, a=0, mode='fan_in')
+                m.weight.data *= scale
+                if m.bias is not None:
+                    m.bias.data.zero_()
+            elif isinstance(m, nn.BatchNorm2d):
+                init.constant_(m.weight, 1)
+                init.constant_(m.bias.data, 0.0)
+
+
+def make_layer(block, n_layers):
+    layers = []
+    for _ in range(n_layers):
+        layers.append(block())
+    return nn.Sequential(*layers)
+
+
+class ResidualDenseBlock_5C(nn.Module):
+    def __init__(self, nf=64, gc=32, bias=True):
+        super(ResidualDenseBlock_5C, self).__init__()
+        # gc: growth channel, i.e. intermediate channels
+        self.conv1 = nn.Conv2d(nf, gc, 3, 1, 1, bias=bias)
+        self.conv2 = nn.Conv2d(nf + gc, gc, 3, 1, 1, bias=bias)
+        self.conv3 = nn.Conv2d(nf + 2 * gc, gc, 3, 1, 1, bias=bias)
+        self.conv4 = nn.Conv2d(nf + 3 * gc, gc, 3, 1, 1, bias=bias)
+        self.conv5 = nn.Conv2d(nf + 4 * gc, nf, 3, 1, 1, bias=bias)
+        self.lrelu = nn.LeakyReLU(negative_slope=0.2, inplace=True)
+
+        # initialization
+        initialize_weights([self.conv1, self.conv2, self.conv3, self.conv4, self.conv5], 0.1)
+
+    def forward(self, x):
+        x1 = self.lrelu(self.conv1(x))
+        x2 = self.lrelu(self.conv2(torch.cat((x, x1), 1)))
+        x3 = self.lrelu(self.conv3(torch.cat((x, x1, x2), 1)))
+        x4 = self.lrelu(self.conv4(torch.cat((x, x1, x2, x3), 1)))
+        x5 = self.conv5(torch.cat((x, x1, x2, x3, x4), 1))
+        return x5 * 0.2 + x
+
+
+class RRDB(nn.Module):
+    '''Residual in Residual Dense Block'''
+
+    def __init__(self, nf, gc=32):
+        super(RRDB, self).__init__()
+        self.RDB1 = ResidualDenseBlock_5C(nf, gc)
+        self.RDB2 = ResidualDenseBlock_5C(nf, gc)
+        self.RDB3 = ResidualDenseBlock_5C(nf, gc)
+
+    def forward(self, x):
+        out = self.RDB1(x)
+        out = self.RDB2(out)
+        out = self.RDB3(out)
+        return out * 0.2 + x
+
+
+class RRDBNet(nn.Module):
+    def __init__(self, in_nc=3, out_nc=3, nf=64, nb=23, gc=32, sf=4):
+        super(RRDBNet, self).__init__()
+        RRDB_block_f = functools.partial(RRDB, nf=nf, gc=gc)
+        self.sf = sf
+        print([in_nc, out_nc, nf, nb, gc, sf])
+
+        self.conv_first = nn.Conv2d(in_nc, nf, 3, 1, 1, bias=True)
+        self.RRDB_trunk = make_layer(RRDB_block_f, nb)
+        self.trunk_conv = nn.Conv2d(nf, nf, 3, 1, 1, bias=True)
+        #### upsampling
+        self.upconv1 = nn.Conv2d(nf, nf, 3, 1, 1, bias=True)
+        if self.sf==4:
+            self.upconv2 = nn.Conv2d(nf, nf, 3, 1, 1, bias=True)
+        self.HRconv = nn.Conv2d(nf, nf, 3, 1, 1, bias=True)
+        self.conv_last = nn.Conv2d(nf, out_nc, 3, 1, 1, bias=True)
+
+        self.lrelu = nn.LeakyReLU(negative_slope=0.2, inplace=True)
+
+    def forward(self, x):
+        fea = self.conv_first(x)
+        trunk = self.trunk_conv(self.RRDB_trunk(fea))
+        fea = fea + trunk
+
+        fea = self.lrelu(self.upconv1(F.interpolate(fea, scale_factor=2, mode='nearest')))
+        if self.sf==4:
+            fea = self.lrelu(self.upconv2(F.interpolate(fea, scale_factor=2, mode='nearest')))
+        out = self.conv_last(self.lrelu(self.HRconv(fea)))
+
+        return out

model/cldm.py

@@ -0,0 +1,155 @@
+from typing import Tuple, Set, List, Dict
+
+import torch
+from torch import nn
+
+from model import (
+    ControlledUnetModel, ControlNet,
+    AutoencoderKL, FrozenOpenCLIPEmbedder
+)
+from utils.common import sliding_windows, count_vram_usage, gaussian_weights
+
+
+def disabled_train(self: nn.Module) -> nn.Module:
+    """Overwrite model.train with this function to make sure train/eval mode
+    does not change anymore."""
+    return self
+
+
+class ControlLDM(nn.Module):
+
+    def __init__(
+        self,
+        unet_cfg,
+        vae_cfg,
+        clip_cfg,
+        controlnet_cfg,
+        latent_scale_factor
+    ):
+        super().__init__()
+        self.unet = ControlledUnetModel(**unet_cfg)
+        self.vae = AutoencoderKL(**vae_cfg)
+        self.clip = FrozenOpenCLIPEmbedder(**clip_cfg)
+        self.controlnet = ControlNet(**controlnet_cfg)
+        self.scale_factor = latent_scale_factor
+        self.control_scales = [1.0] * 13
+
+    @torch.no_grad()
+    def load_pretrained_sd(self, sd: Dict[str, torch.Tensor]) -> Set[str]:
+        module_map = {
+            "unet": "model.diffusion_model",
+            "vae": "first_stage_model",
+            "clip": "cond_stage_model",
+        }
+        modules = [("unet", self.unet), ("vae", self.vae), ("clip", self.clip)]
+        used = set()
+        for name, module in modules:
+            init_sd = {}
+            scratch_sd = module.state_dict()
+            for key in scratch_sd:
+                target_key = ".".join([module_map[name], key])
+                init_sd[key] = sd[target_key].clone()
+                used.add(target_key)
+            module.load_state_dict(init_sd, strict=True)
+        unused = set(sd.keys()) - used
+        # NOTE: this is slightly different from previous version, which haven't switched
+        # the UNet to eval mode and disabled the requires_grad flag.
+        for module in [self.vae, self.clip, self.unet]:
+            module.eval()
+            module.train = disabled_train
+            for p in module.parameters():
+                p.requires_grad = False
+        return unused
+    
+    @torch.no_grad()
+    def load_controlnet_from_ckpt(self, sd: Dict[str, torch.Tensor]) -> None:
+        self.controlnet.load_state_dict(sd, strict=True)
+
+    @torch.no_grad()
+    def load_controlnet_from_unet(self) -> Tuple[Set[str]]:
+        unet_sd = self.unet.state_dict()
+        scratch_sd = self.controlnet.state_dict()
+        init_sd = {}
+        init_with_new_zero = set()
+        init_with_scratch = set()
+        for key in scratch_sd:
+            if key in unet_sd:
+                this, target = scratch_sd[key], unet_sd[key]
+                if this.size() == target.size():
+                    init_sd[key] = target.clone()
+                else:
+                    d_ic = this.size(1) - target.size(1)
+                    oc, _, h, w = this.size()
+                    zeros = torch.zeros((oc, d_ic, h, w), dtype=target.dtype)
+                    init_sd[key] = torch.cat((target, zeros), dim=1)
+                    init_with_new_zero.add(key)
+            else:
+                init_sd[key] = scratch_sd[key].clone()
+                init_with_scratch.add(key)
+        self.controlnet.load_state_dict(init_sd, strict=True)
+        return init_with_new_zero, init_with_scratch
+    
+    def vae_encode(self, image: torch.Tensor, sample: bool=True) -> torch.Tensor:
+        if sample:
+            return self.vae.encode(image).sample() * self.scale_factor
+        else:
+            return self.vae.encode(image).mode() * self.scale_factor
+    
+    def vae_encode_tiled(self, image: torch.Tensor, tile_size: int, tile_stride: int, sample: bool=True) -> torch.Tensor:
+        bs, _, h, w = image.shape
+        z = torch.zeros((bs, 4, h // 8, w // 8), dtype=torch.float32, device=image.device)
+        count = torch.zeros_like(z, dtype=torch.float32)
+        weights = gaussian_weights(tile_size // 8, tile_size // 8)[None, None]
+        weights = torch.tensor(weights, dtype=torch.float32, device=image.device)
+        tiles = sliding_windows(h // 8, w // 8, tile_size // 8, tile_stride // 8)
+        for hi, hi_end, wi, wi_end in tiles:
+            tile_image = image[:, :, hi * 8:hi_end * 8, wi * 8:wi_end * 8]
+            z[:, :, hi:hi_end, wi:wi_end] += self.vae_encode(tile_image, sample=sample) * weights
+            count[:, :, hi:hi_end, wi:wi_end] += weights
+        z.div_(count)
+        return z
+    
+    def vae_decode(self, z: torch.Tensor) -> torch.Tensor:
+        return self.vae.decode(z / self.scale_factor)
+    
+    @count_vram_usage
+    def vae_decode_tiled(self, z: torch.Tensor, tile_size: int, tile_stride: int) -> torch.Tensor:
+        bs, _, h, w = z.shape
+        image = torch.zeros((bs, 3, h * 8, w * 8), dtype=torch.float32, device=z.device)
+        count = torch.zeros_like(image, dtype=torch.float32)
+        weights = gaussian_weights(tile_size * 8, tile_size * 8)[None, None]
+        weights = torch.tensor(weights, dtype=torch.float32, device=z.device)
+        tiles = sliding_windows(h, w, tile_size, tile_stride)
+        for hi, hi_end, wi, wi_end in tiles:
+            tile_z = z[:, :, hi:hi_end, wi:wi_end]
+            image[:, :, hi * 8:hi_end * 8, wi * 8:wi_end * 8] += self.vae_decode(tile_z) * weights
+            count[:, :, hi * 8:hi_end * 8, wi * 8:wi_end * 8] += weights
+        image.div_(count)
+        return image
+
+    def prepare_condition(self, clean: torch.Tensor, txt: List[str]) -> Dict[str, torch.Tensor]:
+        return dict(
+            c_txt=self.clip.encode(txt),
+            c_img=self.vae_encode(clean * 2 - 1, sample=False)
+        )
+    
+    @count_vram_usage
+    def prepare_condition_tiled(self, clean: torch.Tensor, txt: List[str], tile_size: int, tile_stride: int) -> Dict[str, torch.Tensor]:
+        return dict(
+            c_txt=self.clip.encode(txt),
+            c_img=self.vae_encode_tiled(clean * 2 - 1, tile_size, tile_stride, sample=False)
+        )
+
+    def forward(self, x_noisy, t, cond):
+        c_txt = cond["c_txt"]
+        c_img = cond["c_img"]
+        control = self.controlnet(
+            x=x_noisy, hint=c_img,
+            timesteps=t, context=c_txt
+        )
+        control = [c * scale for c, scale in zip(control, self.control_scales)]
+        eps = self.unet(
+            x=x_noisy, timesteps=t,
+            context=c_txt, control=control, only_mid_control=False
+        )
+        return eps

model/clip.py

@@ -0,0 +1,65 @@
+from typing import List
+import torch
+import torch.nn as nn
+from torch.utils.checkpoint import checkpoint
+from model.open_clip import CLIP, tokenize
+
+### pretrained model path
+# _VITH14 = dict(
+#     laion2b_s32b_b79k=_pcfg(hf_hub='laion/CLIP-ViT-H-14-laion2B-s32B-b79K/'),
+# )
+
+class FrozenOpenCLIPEmbedder(nn.Module):
+    """
+    Uses the OpenCLIP transformer encoder for text
+    """
+    LAYERS = [
+        #"pooled",
+        "last",
+        "penultimate"
+    ]
+    def __init__(self, embed_dim, vision_cfg, text_cfg, layer="last"):
+        super().__init__()
+        assert layer in self.LAYERS
+        # model, _, _ = open_clip.create_model_and_transforms(arch, device=torch.device('cpu'), pretrained=version)
+        model = CLIP(embed_dim, dict(vision_cfg), dict(text_cfg))
+        del model.visual
+        self.model = model
+        
+        self.layer = layer
+        if self.layer == "last":
+            self.layer_idx = 0
+        elif self.layer == "penultimate":
+            self.layer_idx = 1
+        else:
+            raise NotImplementedError()
+
+    def forward(self, tokens):
+        z = self.encode_with_transformer(tokens)
+        return z
+
+    def encode_with_transformer(self, text):
+        x = self.model.token_embedding(text)  # [batch_size, n_ctx, d_model]
+        x = x + self.model.positional_embedding
+        x = x.permute(1, 0, 2)  # NLD -> LND
+        x = self.text_transformer_forward(x, attn_mask=self.model.attn_mask)
+        x = x.permute(1, 0, 2)  # LND -> NLD
+        x = self.model.ln_final(x)
+        return x
+
+    def text_transformer_forward(self, x: torch.Tensor, attn_mask = None):
+        for i, r in enumerate(self.model.transformer.resblocks):
+            if i == len(self.model.transformer.resblocks) - self.layer_idx:
+                break
+            if self.model.transformer.grad_checkpointing and not torch.jit.is_scripting():
+                x = checkpoint(r, x, attn_mask)
+            else:
+                x = r(x, attn_mask=attn_mask)
+        return x
+
+    def encode(self, text: List[str]) -> torch.Tensor:
+        # convert a batch of text to tensor
+        tokens = tokenize(text)
+        # move tensor to model device
+        tokens = tokens.to(next(self.model.parameters()).device)
+        return self(tokens)

model/config.py

@@ -0,0 +1,62 @@
+import os
+from typing import Optional, Literal
+from types import ModuleType
+import enum
+from packaging import version
+
+import torch
+
+# collect system information
+if version.parse(torch.__version__) >= version.parse("2.0.0"):
+    SDP_IS_AVAILABLE = True
+else:
+    SDP_IS_AVAILABLE = False
+
+try:
+    import xformers
+    import xformers.ops
+    XFORMERS_IS_AVAILBLE = True
+except:
+    XFORMERS_IS_AVAILBLE = False
+
+
+class AttnMode(enum.Enum):
+    SDP = 0
+    XFORMERS = 1
+    VANILLA = 2
+
+
+class Config:
+    xformers: Optional[ModuleType] = None
+    attn_mode: AttnMode = AttnMode.VANILLA
+
+
+# initialize attention mode
+if SDP_IS_AVAILABLE:
+    Config.attn_mode = AttnMode.SDP
+    print(f"use sdp attention as default")
+elif XFORMERS_IS_AVAILBLE:
+    Config.attn_mode = AttnMode.XFORMERS
+    print(f"use xformers attention as default")
+else:
+    print(f"both sdp attention and xformers are not available, use vanilla attention (very expensive) as default")
+
+if XFORMERS_IS_AVAILBLE:
+    Config.xformers = xformers
+
+
+# user-specified attention mode
+ATTN_MODE = os.environ.get("ATTN_MODE", None)
+if ATTN_MODE is not None:
+    assert ATTN_MODE in ["vanilla", "sdp", "xformers"]
+    if ATTN_MODE == "sdp":
+        assert SDP_IS_AVAILABLE
+        Config.attn_mode = AttnMode.SDP
+    elif ATTN_MODE == "xformers":
+        assert XFORMERS_IS_AVAILBLE
+        Config.attn_mode = AttnMode.XFORMERS
+    else:
+        Config.attn_mode = AttnMode.VANILLA
+    print(f"set attention mode to {ATTN_MODE}")
+else:
+    print("keep default attention mode")

model/controlnet.py

@@ -0,0 +1,277 @@
+import torch
+import torch as th
+import torch.nn as nn
+
+from model.util import (
+    conv_nd,
+    linear,
+    zero_module,
+    timestep_embedding,
+    exists
+)
+from model.attention import SpatialTransformer
+from model.unet import (
+    TimestepEmbedSequential, ResBlock, Downsample, AttentionBlock, UNetModel
+)
+
+
+class ControlledUnetModel(UNetModel):
+
+    def forward(self, x, timesteps=None, context=None, control=None, only_mid_control=False, **kwargs):
+        hs = []
+        t_emb = timestep_embedding(timesteps, self.model_channels, repeat_only=False)
+        emb = self.time_embed(t_emb)
+        h = x.type(self.dtype)
+        for module in self.input_blocks:
+            h = module(h, emb, context)
+            hs.append(h)
+        h = self.middle_block(h, emb, context)
+
+        if control is not None:
+            h += control.pop()
+
+        for i, module in enumerate(self.output_blocks):
+            if only_mid_control or control is None:
+                h = torch.cat([h, hs.pop()], dim=1)
+            else:
+                h = torch.cat([h, hs.pop() + control.pop()], dim=1)
+            h = module(h, emb, context)
+
+        h = h.type(x.dtype)
+        return self.out(h)
+
+
+class ControlNet(nn.Module):
+
+    def __init__(
+        self,
+        image_size,
+        in_channels,
+        model_channels,
+        hint_channels,
+        num_res_blocks,
+        attention_resolutions,
+        dropout=0,
+        channel_mult=(1, 2, 4, 8),
+        conv_resample=True,
+        dims=2,
+        use_checkpoint=False,
+        use_fp16=False,
+        num_heads=-1,
+        num_head_channels=-1,
+        num_heads_upsample=-1,
+        use_scale_shift_norm=False,
+        resblock_updown=False,
+        use_new_attention_order=False,
+        use_spatial_transformer=False,  # custom transformer support
+        transformer_depth=1,  # custom transformer support
+        context_dim=None,  # custom transformer support
+        n_embed=None,  # custom support for prediction of discrete ids into codebook of first stage vq model
+        legacy=True,
+        disable_self_attentions=None,
+        num_attention_blocks=None,
+        disable_middle_self_attn=False,
+        use_linear_in_transformer=False,
+    ):
+        super().__init__()
+        if use_spatial_transformer:
+            assert context_dim is not None, 'Fool!! You forgot to include the dimension of your cross-attention conditioning...'
+
+        if context_dim is not None:
+            assert use_spatial_transformer, 'Fool!! You forgot to use the spatial transformer for your cross-attention conditioning...'
+            from omegaconf.listconfig import ListConfig
+            if type(context_dim) == ListConfig:
+                context_dim = list(context_dim)
+
+        if num_heads_upsample == -1:
+            num_heads_upsample = num_heads
+
+        if num_heads == -1:
+            assert num_head_channels != -1, 'Either num_heads or num_head_channels has to be set'
+
+        if num_head_channels == -1:
+            assert num_heads != -1, 'Either num_heads or num_head_channels has to be set'
+
+        self.dims = dims
+        self.image_size = image_size
+        self.in_channels = in_channels
+        self.model_channels = model_channels
+        if isinstance(num_res_blocks, int):
+            self.num_res_blocks = len(channel_mult) * [num_res_blocks]
+        else:
+            if len(num_res_blocks) != len(channel_mult):
+                raise ValueError("provide num_res_blocks either as an int (globally constant) or "
+                                 "as a list/tuple (per-level) with the same length as channel_mult")
+            self.num_res_blocks = num_res_blocks
+        if disable_self_attentions is not None:
+            # should be a list of booleans, indicating whether to disable self-attention in TransformerBlocks or not
+            assert len(disable_self_attentions) == len(channel_mult)
+        if num_attention_blocks is not None:
+            assert len(num_attention_blocks) == len(self.num_res_blocks)
+            assert all(map(lambda i: self.num_res_blocks[i] >= num_attention_blocks[i], range(len(num_attention_blocks))))
+            print(f"Constructor of UNetModel received num_attention_blocks={num_attention_blocks}. "
+                  f"This option has LESS priority than attention_resolutions {attention_resolutions}, "
+                  f"i.e., in cases where num_attention_blocks[i] > 0 but 2**i not in attention_resolutions, "
+                  f"attention will still not be set.")
+
+        self.attention_resolutions = attention_resolutions
+        self.dropout = dropout
+        self.channel_mult = channel_mult
+        self.conv_resample = conv_resample
+        self.use_checkpoint = use_checkpoint
+        self.dtype = th.float16 if use_fp16 else th.float32
+        self.num_heads = num_heads
+        self.num_head_channels = num_head_channels
+        self.num_heads_upsample = num_heads_upsample
+        self.predict_codebook_ids = n_embed is not None
+
+        time_embed_dim = model_channels * 4
+        self.time_embed = nn.Sequential(
+            linear(model_channels, time_embed_dim),
+            nn.SiLU(),
+            linear(time_embed_dim, time_embed_dim),
+        )
+
+        self.input_blocks = nn.ModuleList(
+            [
+                TimestepEmbedSequential(
+                    conv_nd(dims, in_channels + hint_channels, model_channels, 3, padding=1)
+                )
+            ]
+        )
+        self.zero_convs = nn.ModuleList([self.make_zero_conv(model_channels)])
+
+        self._feature_size = model_channels
+        input_block_chans = [model_channels]
+        ch = model_channels
+        ds = 1
+        for level, mult in enumerate(channel_mult):
+            for nr in range(self.num_res_blocks[level]):
+                layers = [
+                    ResBlock(
+                        ch,
+                        time_embed_dim,
+                        dropout,
+                        out_channels=mult * model_channels,
+                        dims=dims,
+                        use_checkpoint=use_checkpoint,
+                        use_scale_shift_norm=use_scale_shift_norm,
+                    )
+                ]
+                ch = mult * model_channels
+                if ds in attention_resolutions:
+                    if num_head_channels == -1:
+                        dim_head = ch // num_heads
+                    else:
+                        num_heads = ch // num_head_channels
+                        dim_head = num_head_channels
+                    if legacy:
+                        # num_heads = 1
+                        dim_head = ch // num_heads if use_spatial_transformer else num_head_channels
+                    if exists(disable_self_attentions):
+                        disabled_sa = disable_self_attentions[level]
+                    else:
+                        disabled_sa = False
+
+                    if not exists(num_attention_blocks) or nr < num_attention_blocks[level]:
+                        layers.append(
+                            AttentionBlock(
+                                ch,
+                                use_checkpoint=use_checkpoint,
+                                num_heads=num_heads,
+                                num_head_channels=dim_head,
+                                use_new_attention_order=use_new_attention_order,
+                            ) if not use_spatial_transformer else SpatialTransformer(
+                                ch, num_heads, dim_head, depth=transformer_depth, context_dim=context_dim,
+                                disable_self_attn=disabled_sa, use_linear=use_linear_in_transformer,
+                                use_checkpoint=use_checkpoint
+                            )
+                        )
+                self.input_blocks.append(TimestepEmbedSequential(*layers))
+                self.zero_convs.append(self.make_zero_conv(ch))
+                self._feature_size += ch
+                input_block_chans.append(ch)
+            if level != len(channel_mult) - 1:
+                out_ch = ch
+                self.input_blocks.append(
+                    TimestepEmbedSequential(
+                        ResBlock(
+                            ch,
+                            time_embed_dim,
+                            dropout,
+                            out_channels=out_ch,
+                            dims=dims,
+                            use_checkpoint=use_checkpoint,
+                            use_scale_shift_norm=use_scale_shift_norm,
+                            down=True,
+                        )
+                        if resblock_updown
+                        else Downsample(
+                            ch, conv_resample, dims=dims, out_channels=out_ch
+                        )
+                    )
+                )
+                ch = out_ch
+                input_block_chans.append(ch)
+                self.zero_convs.append(self.make_zero_conv(ch))
+                ds *= 2
+                self._feature_size += ch
+
+        if num_head_channels == -1:
+            dim_head = ch // num_heads
+        else:
+            num_heads = ch // num_head_channels
+            dim_head = num_head_channels
+        if legacy:
+            # num_heads = 1
+            dim_head = ch // num_heads if use_spatial_transformer else num_head_channels
+        self.middle_block = TimestepEmbedSequential(
+            ResBlock(
+                ch,
+                time_embed_dim,
+                dropout,
+                dims=dims,
+                use_checkpoint=use_checkpoint,
+                use_scale_shift_norm=use_scale_shift_norm,
+            ),
+            AttentionBlock(
+                ch,
+                use_checkpoint=use_checkpoint,
+                num_heads=num_heads,
+                num_head_channels=dim_head,
+                use_new_attention_order=use_new_attention_order,
+            ) if not use_spatial_transformer else SpatialTransformer(  # always uses a self-attn
+                ch, num_heads, dim_head, depth=transformer_depth, context_dim=context_dim,
+                disable_self_attn=disable_middle_self_attn, use_linear=use_linear_in_transformer,
+                use_checkpoint=use_checkpoint
+            ),
+            ResBlock(
+                ch,
+                time_embed_dim,
+                dropout,
+                dims=dims,
+                use_checkpoint=use_checkpoint,
+                use_scale_shift_norm=use_scale_shift_norm,
+            ),
+        )
+        self.middle_block_out = self.make_zero_conv(ch)
+        self._feature_size += ch
+
+    def make_zero_conv(self, channels):
+        return TimestepEmbedSequential(zero_module(conv_nd(self.dims, channels, channels, 1, padding=0)))
+
+    def forward(self, x, hint, timesteps, context, **kwargs):
+        t_emb = timestep_embedding(timesteps, self.model_channels, repeat_only=False)
+        emb = self.time_embed(t_emb)
+        x = torch.cat((x, hint), dim=1)
+        outs = []
+
+        h = x.type(self.dtype)
+        for module, zero_conv in zip(self.input_blocks, self.zero_convs):
+            h = module(h, emb, context)
+            outs.append(zero_conv(h, emb, context))
+
+        h = self.middle_block(h, emb, context)
+        outs.append(self.middle_block_out(h, emb, context))
+
+        return outs

model/distributions.py

@@ -0,0 +1,92 @@
+import torch
+import numpy as np
+
+
+class AbstractDistribution:
+    def sample(self):
+        raise NotImplementedError()
+
+    def mode(self):
+        raise NotImplementedError()
+
+
+class DiracDistribution(AbstractDistribution):
+    def __init__(self, value):
+        self.value = value
+
+    def sample(self):
+        return self.value
+
+    def mode(self):
+        return self.value
+
+
+class DiagonalGaussianDistribution(object):
+    def __init__(self, parameters, deterministic=False):
+        self.parameters = parameters
+        self.mean, self.logvar = torch.chunk(parameters, 2, dim=1)
+        self.logvar = torch.clamp(self.logvar, -30.0, 20.0)
+        self.deterministic = deterministic
+        self.std = torch.exp(0.5 * self.logvar)
+        self.var = torch.exp(self.logvar)
+        if self.deterministic:
+            self.var = self.std = torch.zeros_like(self.mean).to(device=self.parameters.device)
+
+    def sample(self):
+        x = self.mean + self.std * torch.randn(self.mean.shape).to(device=self.parameters.device)
+        return x
+
+    def kl(self, other=None):
+        if self.deterministic:
+            return torch.Tensor([0.])
+        else:
+            if other is None:
+                return 0.5 * torch.sum(torch.pow(self.mean, 2)
+                                       + self.var - 1.0 - self.logvar,
+                                       dim=[1, 2, 3])
+            else:
+                return 0.5 * torch.sum(
+                    torch.pow(self.mean - other.mean, 2) / other.var
+                    + self.var / other.var - 1.0 - self.logvar + other.logvar,
+                    dim=[1, 2, 3])
+
+    def nll(self, sample, dims=[1,2,3]):
+        if self.deterministic:
+            return torch.Tensor([0.])
+        logtwopi = np.log(2.0 * np.pi)
+        return 0.5 * torch.sum(
+            logtwopi + self.logvar + torch.pow(sample - self.mean, 2) / self.var,
+            dim=dims)
+
+    def mode(self):
+        return self.mean
+
+
+def normal_kl(mean1, logvar1, mean2, logvar2):
+    """
+    source: https://github.com/openai/guided-diffusion/blob/27c20a8fab9cb472df5d6bdd6c8d11c8f430b924/guided_diffusion/losses.py#L12
+    Compute the KL divergence between two gaussians.
+    Shapes are automatically broadcasted, so batches can be compared to
+    scalars, among other use cases.
+    """
+    tensor = None
+    for obj in (mean1, logvar1, mean2, logvar2):
+        if isinstance(obj, torch.Tensor):
+            tensor = obj
+            break
+    assert tensor is not None, "at least one argument must be a Tensor"
+
+    # Force variances to be Tensors. Broadcasting helps convert scalars to
+    # Tensors, but it does not work for torch.exp().
+    logvar1, logvar2 = [
+        x if isinstance(x, torch.Tensor) else torch.tensor(x).to(tensor)
+        for x in (logvar1, logvar2)
+    ]
+
+    return 0.5 * (
+        -1.0
+        + logvar2
+        - logvar1
+        + torch.exp(logvar1 - logvar2)
+        + ((mean1 - mean2) ** 2) * torch.exp(-logvar2)
+    )

model/gaussian_diffusion.py

@@ -0,0 +1,118 @@
+from functools import partial
+from typing import Tuple
+
+import torch
+from torch import nn
+import numpy as np
+
+
+def make_beta_schedule(schedule, n_timestep, linear_start=1e-4, linear_end=2e-2, cosine_s=8e-3):
+    if schedule == "linear":
+        betas = (
+            np.linspace(linear_start ** 0.5, linear_end ** 0.5, n_timestep, dtype=np.float64) ** 2
+        )
+
+    elif schedule == "cosine":
+        timesteps = (
+            np.arange(n_timestep + 1, dtype=np.float64) / n_timestep + cosine_s
+        )
+        alphas = timesteps / (1 + cosine_s) * np.pi / 2
+        alphas = np.cos(alphas).pow(2)
+        alphas = alphas / alphas[0]
+        betas = 1 - alphas[1:] / alphas[:-1]
+        betas = np.clip(betas, a_min=0, a_max=0.999)
+
+    elif schedule == "sqrt_linear":
+        betas = np.linspace(linear_start, linear_end, n_timestep, dtype=np.float64)
+    elif schedule == "sqrt":
+        betas = np.linspace(linear_start, linear_end, n_timestep, dtype=np.float64) ** 0.5
+    else:
+        raise ValueError(f"schedule '{schedule}' unknown.")
+    return betas
+
+
+def extract_into_tensor(a: torch.Tensor, t: torch.Tensor, x_shape: Tuple[int]) -> torch.Tensor:
+    b, *_ = t.shape
+    out = a.gather(-1, t)
+    return out.reshape(b, *((1,) * (len(x_shape) - 1)))
+
+
+class Diffusion(nn.Module):
+
+    def __init__(
+        self,
+        timesteps=1000,
+        beta_schedule="linear",
+        loss_type="l2",
+        linear_start=1e-4,
+        linear_end=2e-2,
+        cosine_s=8e-3,
+        parameterization="eps"
+    ):
+        super().__init__()
+        self.num_timesteps = timesteps
+        self.beta_schedule = beta_schedule
+        self.linear_start = linear_start
+        self.linear_end = linear_end
+        self.cosine_s = cosine_s
+        assert parameterization in ["eps", "x0", "v"], "currently only supporting 'eps' and 'x0' and 'v'"
+        self.parameterization = parameterization
+        self.loss_type = loss_type
+        
+        betas = make_beta_schedule(beta_schedule, timesteps, linear_start=linear_start, linear_end=linear_end,
+                                   cosine_s=cosine_s)
+        alphas = 1. - betas
+        alphas_cumprod = np.cumprod(alphas, axis=0)
+        sqrt_alphas_cumprod = np.sqrt(alphas_cumprod)
+        sqrt_one_minus_alphas_cumprod = np.sqrt(1. - alphas_cumprod)
+
+        self.betas = betas
+        self.register("sqrt_alphas_cumprod", sqrt_alphas_cumprod)
+        self.register("sqrt_one_minus_alphas_cumprod", sqrt_one_minus_alphas_cumprod)
+    
+    def register(self, name: str, value: np.ndarray) -> None:
+        self.register_buffer(name, torch.tensor(value, dtype=torch.float32))
+
+    def q_sample(self, x_start, t, noise):
+        return (
+            extract_into_tensor(self.sqrt_alphas_cumprod, t, x_start.shape) * x_start +
+            extract_into_tensor(self.sqrt_one_minus_alphas_cumprod, t, x_start.shape) * noise
+        )
+
+    def get_v(self, x, noise, t):
+        return (
+            extract_into_tensor(self.sqrt_alphas_cumprod, t, x.shape) * noise -
+            extract_into_tensor(self.sqrt_one_minus_alphas_cumprod, t, x.shape) * x
+        )
+
+    def get_loss(self, pred, target, mean=True):
+        if self.loss_type == 'l1':
+            loss = (target - pred).abs()
+            if mean:
+                loss = loss.mean()
+        elif self.loss_type == 'l2':
+            if mean:
+                loss = torch.nn.functional.mse_loss(target, pred)
+            else:
+                loss = torch.nn.functional.mse_loss(target, pred, reduction='none')
+        else:
+            raise NotImplementedError("unknown loss type '{loss_type}'")
+
+        return loss
+
+    def p_losses(self, model, x_start, t, cond):
+        noise = torch.randn_like(x_start)
+        x_noisy = self.q_sample(x_start=x_start, t=t, noise=noise)
+        model_output = model(x_noisy, t, cond)
+
+        if self.parameterization == "x0":
+            target = x_start
+        elif self.parameterization == "eps":
+            target = noise
+        elif self.parameterization == "v":
+            target = self.get_v(x_start, noise, t)
+        else:
+            raise NotImplementedError()
+
+        loss_simple = self.get_loss(model_output, target, mean=False).mean()
+        return loss_simple

model/open_clip/__init__.py

@@ -0,0 +1,4 @@
+from .model import CLIP
+from .tokenizer import tokenize
+
+__all__ = ["CLIP", "tokenize"]

model/open_clip/bpe_simple_vocab_16e6.txt.gz

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:924691ac288e54409236115652ad4aa250f48203de50a9e4722a6ecd48d6804a
+size 1356917

model/open_clip/model.py

@@ -0,0 +1,206 @@
+""" CLIP Model
+
+Adapted from https://github.com/openai/CLIP. Originally MIT License, Copyright (c) 2021 OpenAI.
+"""
+from dataclasses import dataclass
+from typing import Optional, Tuple, Union
+
+import numpy as np
+import torch
+import torch.nn.functional as F
+from torch import nn
+
+from .transformer import LayerNormFp32, LayerNorm, QuickGELU, VisionTransformer, TextTransformer
+
+
+@dataclass
+class CLIPVisionCfg:
+    layers: Union[Tuple[int, int, int, int], int] = 12
+    width: int = 768
+    head_width: int = 64
+    mlp_ratio: float = 4.0
+    patch_size: int = 16
+    image_size: Union[Tuple[int, int], int] = 224
+
+    ls_init_value: Optional[float] = None  # layer scale initial value
+    patch_dropout: float = 0.  # what fraction of patches to dropout during training (0 would mean disabled and no patches dropped) - 0.5 to 0.75 recommended in the paper for optimal results
+    input_patchnorm: bool = False  # whether to use dual patchnorm - would only apply the input layernorm on each patch, as post-layernorm already exist in original clip vit design
+    global_average_pool: bool = False  # whether to global average pool the last embedding layer, instead of using CLS token (https://arxiv.org/abs/2205.01580)
+    attentional_pool: bool = False  # whether to use attentional pooler in the last embedding layer
+    n_queries: int = 256  # n_queries for attentional pooler
+    attn_pooler_heads: int = 8  # n heads for attentional_pooling
+    output_tokens: bool = False
+
+    timm_model_name: str = None  # a valid model name overrides layers, width, patch_size
+    timm_model_pretrained: bool = False  # use (imagenet) pretrained weights for named model
+    timm_pool: str = 'avg'  # feature pooling for timm model ('abs_attn', 'rot_attn', 'avg', '')
+    timm_proj: str = 'linear'  # linear projection for timm model output ('linear', 'mlp', '')
+    timm_proj_bias: bool = False  # enable bias final projection
+    timm_drop: float = 0.  # head dropout
+    timm_drop_path: Optional[float] = None  # backbone stochastic depth
+
+
+@dataclass
+class CLIPTextCfg:
+    context_length: int = 77
+    vocab_size: int = 49408
+    width: int = 512
+    heads: int = 8
+    layers: int = 12
+    ls_init_value: Optional[float] = None  # layer scale initial value
+    hf_model_name: str = None
+    hf_tokenizer_name: str = None
+    hf_model_pretrained: bool = True
+    proj: str = 'mlp'
+    pooler_type: str = 'mean_pooler'
+    embed_cls: bool = False
+    pad_id: int = 0
+    output_tokens: bool = False
+
+
+def get_cast_dtype(precision: str):
+    cast_dtype = None
+    if precision == 'bf16':
+        cast_dtype = torch.bfloat16
+    elif precision == 'fp16':
+        cast_dtype = torch.float16
+    return cast_dtype
+
+
+def _build_vision_tower(
+        embed_dim: int,
+        vision_cfg: CLIPVisionCfg,
+        quick_gelu: bool = False,
+        cast_dtype: Optional[torch.dtype] = None
+):
+    if isinstance(vision_cfg, dict):
+        vision_cfg = CLIPVisionCfg(**vision_cfg)
+
+    # OpenAI models are pretrained w/ QuickGELU but native nn.GELU is both faster and more
+    # memory efficient in recent PyTorch releases (>= 1.10).
+    # NOTE: timm models always use native GELU regardless of quick_gelu flag.
+    act_layer = QuickGELU if quick_gelu else nn.GELU
+
+    vision_heads = vision_cfg.width // vision_cfg.head_width
+    norm_layer = LayerNormFp32 if cast_dtype in (torch.float16, torch.bfloat16) else LayerNorm
+    visual = VisionTransformer(
+        image_size=vision_cfg.image_size,
+        patch_size=vision_cfg.patch_size,
+        width=vision_cfg.width,
+        layers=vision_cfg.layers,
+        heads=vision_heads,
+        mlp_ratio=vision_cfg.mlp_ratio,
+        ls_init_value=vision_cfg.ls_init_value,
+        patch_dropout=vision_cfg.patch_dropout,
+        input_patchnorm=vision_cfg.input_patchnorm,
+        global_average_pool=vision_cfg.global_average_pool,
+        attentional_pool=vision_cfg.attentional_pool,
+        n_queries=vision_cfg.n_queries,
+        attn_pooler_heads=vision_cfg.attn_pooler_heads,
+        output_tokens=vision_cfg.output_tokens,
+        output_dim=embed_dim,
+        act_layer=act_layer,
+        norm_layer=norm_layer,
+    )
+
+    return visual
+
+
+def _build_text_tower(
+        embed_dim: int,
+        text_cfg: CLIPTextCfg,
+        quick_gelu: bool = False,
+        cast_dtype: Optional[torch.dtype] = None,
+):
+    if isinstance(text_cfg, dict):
+        text_cfg = CLIPTextCfg(**text_cfg)
+
+    act_layer = QuickGELU if quick_gelu else nn.GELU
+    norm_layer = LayerNormFp32 if cast_dtype in (torch.float16, torch.bfloat16) else LayerNorm
+
+    text = TextTransformer(
+        context_length=text_cfg.context_length,
+        vocab_size=text_cfg.vocab_size,
+        width=text_cfg.width,
+        heads=text_cfg.heads,
+        layers=text_cfg.layers,
+        ls_init_value=text_cfg.ls_init_value,
+        output_dim=embed_dim,
+        embed_cls=text_cfg.embed_cls,
+        output_tokens=text_cfg.output_tokens,
+        pad_id=text_cfg.pad_id,
+        act_layer=act_layer,
+        norm_layer=norm_layer,
+    )
+    return text
+
+
+class CLIP(nn.Module):
+    output_dict: torch.jit.Final[bool]
+
+    def __init__(
+            self,
+            embed_dim: int,
+            vision_cfg: CLIPVisionCfg,
+            text_cfg: CLIPTextCfg,
+            quick_gelu: bool = False,
+            cast_dtype: Optional[torch.dtype] = None,
+            output_dict: bool = False,
+    ):
+        super().__init__()
+        self.output_dict = output_dict
+        self.visual = _build_vision_tower(embed_dim, vision_cfg, quick_gelu, cast_dtype)
+
+        text = _build_text_tower(embed_dim, text_cfg, quick_gelu, cast_dtype)
+        self.transformer = text.transformer
+        self.context_length = text.context_length
+        self.vocab_size = text.vocab_size
+        self.token_embedding = text.token_embedding
+        self.positional_embedding = text.positional_embedding
+        self.ln_final = text.ln_final
+        self.text_projection = text.text_projection
+        self.register_buffer('attn_mask', text.attn_mask, persistent=False)
+
+        self.logit_scale = nn.Parameter(torch.ones([]) * np.log(1 / 0.07))
+
+    def lock_image_tower(self, unlocked_groups=0, freeze_bn_stats=False):
+        # lock image tower as per LiT - https://arxiv.org/abs/2111.07991
+        self.visual.lock(unlocked_groups=unlocked_groups, freeze_bn_stats=freeze_bn_stats)
+
+    @torch.jit.ignore
+    def set_grad_checkpointing(self, enable=True):
+        self.visual.set_grad_checkpointing(enable)
+        self.transformer.grad_checkpointing = enable
+
+    def encode_image(self, image, normalize: bool = False):
+        features = self.visual(image)
+        return F.normalize(features, dim=-1) if normalize else features
+
+    def encode_text(self, text, normalize: bool = False):
+        cast_dtype = self.transformer.get_cast_dtype()
+
+        x = self.token_embedding(text).to(cast_dtype)  # [batch_size, n_ctx, d_model]
+
+        x = x + self.positional_embedding.to(cast_dtype)
+        x = x.permute(1, 0, 2)  # NLD -> LND
+        x = self.transformer(x, attn_mask=self.attn_mask)
+        x = x.permute(1, 0, 2)  # LND -> NLD
+        x = self.ln_final(x)  # [batch_size, n_ctx, transformer.width]
+        # take features from the eot embedding (eot_token is the highest number in each sequence)
+        x = x[torch.arange(x.shape[0]), text.argmax(dim=-1)] @ self.text_projection
+        return F.normalize(x, dim=-1) if normalize else x
+
+    def forward(
+            self,
+            image: Optional[torch.Tensor] = None,
+            text: Optional[torch.Tensor] = None,
+    ):
+        image_features = self.encode_image(image, normalize=True) if image is not None else None
+        text_features = self.encode_text(text, normalize=True) if text is not None else None
+        if self.output_dict:
+            return {
+                "image_features": image_features,
+                "text_features": text_features,
+                "logit_scale": self.logit_scale.exp()
+            }
+        return image_features, text_features, self.logit_scale.exp()

model/open_clip/tokenizer.py

@@ -0,0 +1,214 @@
+""" CLIP tokenizer
+
+Copied from https://github.com/openai/CLIP. Originally MIT License, Copyright (c) 2021 OpenAI.
+"""
+import gzip
+import html
+import os
+from functools import lru_cache
+from typing import Union, List
+
+import ftfy
+import regex as re
+import torch
+
+# https://stackoverflow.com/q/62691279
+import os
+os.environ["TOKENIZERS_PARALLELISM"] = "false"
+
+
+@lru_cache()
+def default_bpe():
+    return os.path.join(os.path.dirname(os.path.abspath(__file__)), "bpe_simple_vocab_16e6.txt.gz")
+
+
+@lru_cache()
+def bytes_to_unicode():
+    """
+    Returns list of utf-8 byte and a corresponding list of unicode strings.
+    The reversible bpe codes work on unicode strings.
+    This means you need a large # of unicode characters in your vocab if you want to avoid UNKs.
+    When you're at something like a 10B token dataset you end up needing around 5K for decent coverage.
+    This is a significant percentage of your normal, say, 32K bpe vocab.
+    To avoid that, we want lookup tables between utf-8 bytes and unicode strings.
+    And avoids mapping to whitespace/control characters the bpe code barfs on.
+    """
+    bs = list(range(ord("!"), ord("~")+1))+list(range(ord("¡"), ord("¬")+1))+list(range(ord("®"), ord("ÿ")+1))
+    cs = bs[:]
+    n = 0
+    for b in range(2**8):
+        if b not in bs:
+            bs.append(b)
+            cs.append(2**8+n)
+            n += 1
+    cs = [chr(n) for n in cs]
+    return dict(zip(bs, cs))
+
+
+def get_pairs(word):
+    """Return set of symbol pairs in a word.
+    Word is represented as tuple of symbols (symbols being variable-length strings).
+    """
+    pairs = set()
+    prev_char = word[0]
+    for char in word[1:]:
+        pairs.add((prev_char, char))
+        prev_char = char
+    return pairs
+
+
+def basic_clean(text):
+    text = ftfy.fix_text(text)
+    text = html.unescape(html.unescape(text))
+    return text.strip()
+
+
+def whitespace_clean(text):
+    text = re.sub(r'\s+', ' ', text)
+    text = text.strip()
+    return text
+
+
+class SimpleTokenizer(object):
+    def __init__(self, bpe_path: str = default_bpe(), special_tokens=None):
+        self.byte_encoder = bytes_to_unicode()
+        self.byte_decoder = {v: k for k, v in self.byte_encoder.items()}
+        merges = gzip.open(bpe_path).read().decode("utf-8").split('\n')
+        merges = merges[1:49152-256-2+1]
+        merges = [tuple(merge.split()) for merge in merges]
+        vocab = list(bytes_to_unicode().values())
+        vocab = vocab + [v+'</w>' for v in vocab]
+        for merge in merges:
+            vocab.append(''.join(merge))
+        if not special_tokens:
+            special_tokens = ['<start_of_text>', '<end_of_text>']
+        else:
+            special_tokens = ['<start_of_text>', '<end_of_text>'] + special_tokens
+        vocab.extend(special_tokens)
+        self.encoder = dict(zip(vocab, range(len(vocab))))
+        self.decoder = {v: k for k, v in self.encoder.items()}
+        self.bpe_ranks = dict(zip(merges, range(len(merges))))
+        self.cache = {t:t for t in special_tokens}
+        special = "|".join(special_tokens)
+        self.pat = re.compile(special + r"""|'s|'t|'re|'ve|'m|'ll|'d|[\p{L}]+|[\p{N}]|[^\s\p{L}\p{N}]+""", re.IGNORECASE)
+
+        self.vocab_size = len(self.encoder)
+        self.all_special_ids = [self.encoder[t] for t in special_tokens]
+
+    def bpe(self, token):
+        if token in self.cache:
+            return self.cache[token]
+        word = tuple(token[:-1]) + ( token[-1] + '</w>',)
+        pairs = get_pairs(word)
+
+        if not pairs:
+            return token+'</w>'
+
+        while True:
+            bigram = min(pairs, key = lambda pair: self.bpe_ranks.get(pair, float('inf')))
+            if bigram not in self.bpe_ranks:
+                break
+            first, second = bigram
+            new_word = []
+            i = 0
+            while i < len(word):
+                try:
+                    j = word.index(first, i)
+                    new_word.extend(word[i:j])
+                    i = j
+                except:
+                    new_word.extend(word[i:])
+                    break
+
+                if word[i] == first and i < len(word)-1 and word[i+1] == second:
+                    new_word.append(first+second)
+                    i += 2
+                else:
+                    new_word.append(word[i])
+                    i += 1
+            new_word = tuple(new_word)
+            word = new_word
+            if len(word) == 1:
+                break
+            else:
+                pairs = get_pairs(word)
+        word = ' '.join(word)
+        self.cache[token] = word
+        return word
+
+    def encode(self, text):
+        bpe_tokens = []
+        text = whitespace_clean(basic_clean(text)).lower()
+        for token in re.findall(self.pat, text):
+            token = ''.join(self.byte_encoder[b] for b in token.encode('utf-8'))
+            bpe_tokens.extend(self.encoder[bpe_token] for bpe_token in self.bpe(token).split(' '))
+        return bpe_tokens
+
+    def decode(self, tokens):
+        text = ''.join([self.decoder[token] for token in tokens])
+        text = bytearray([self.byte_decoder[c] for c in text]).decode('utf-8', errors="replace").replace('</w>', ' ')
+        return text
+
+
+_tokenizer = SimpleTokenizer()
+
+def decode(output_ids: torch.Tensor):
+    output_ids = output_ids.cpu().numpy()
+    return _tokenizer.decode(output_ids)
+
+def tokenize(texts: Union[str, List[str]], context_length: int = 77) -> torch.LongTensor:
+    """
+    Returns the tokenized representation of given input string(s)
+
+    Parameters
+    ----------
+    texts : Union[str, List[str]]
+        An input string or a list of input strings to tokenize
+    context_length : int
+        The context length to use; all CLIP models use 77 as the context length
+
+    Returns
+    -------
+    A two-dimensional tensor containing the resulting tokens, shape = [number of input strings, context_length]
+    """
+    if isinstance(texts, str):
+        texts = [texts]
+
+    sot_token = _tokenizer.encoder["<start_of_text>"]
+    eot_token = _tokenizer.encoder["<end_of_text>"]
+    all_tokens = [[sot_token] + _tokenizer.encode(text) + [eot_token] for text in texts]
+    result = torch.zeros(len(all_tokens), context_length, dtype=torch.long)
+
+    for i, tokens in enumerate(all_tokens):
+        if len(tokens) > context_length:
+            tokens = tokens[:context_length]  # Truncate
+            tokens[-1] = eot_token
+        result[i, :len(tokens)] = torch.tensor(tokens)
+
+    return result
+
+
+class HFTokenizer:
+    """HuggingFace tokenizer wrapper"""
+
+    def __init__(self, tokenizer_name: str):
+        from transformers import AutoTokenizer
+        self.tokenizer = AutoTokenizer.from_pretrained(tokenizer_name)
+
+    def save_pretrained(self, dest):
+        self.tokenizer.save_pretrained(dest)
+
+    def __call__(self, texts: Union[str, List[str]], context_length: int = 77) -> torch.Tensor:
+        # same cleaning as for default tokenizer, except lowercasing
+        # adding lower (for case-sensitive tokenizers) will make it more robust but less sensitive to nuance
+        if isinstance(texts, str):
+            texts = [texts]
+        texts = [whitespace_clean(basic_clean(text)) for text in texts]
+        input_ids = self.tokenizer(
+            texts,
+            return_tensors='pt',
+            max_length=context_length,
+            padding='max_length',
+            truncation=True,
+        ).input_ids
+        return input_ids

model/open_clip/transformer.py

@@ -0,0 +1,736 @@
+import collections
+from collections import OrderedDict
+import math
+from typing import Callable, Optional, Sequence, Tuple
+from itertools import repeat
+
+import torch
+from torch import nn
+from torch.nn import functional as F
+from torch.utils.checkpoint import checkpoint
+
+# From PyTorch internals
+def _ntuple(n):
+    def parse(x):
+        if isinstance(x, collections.abc.Iterable):
+            return x
+        return tuple(repeat(x, n))
+    return parse
+
+to_2tuple = _ntuple(2)
+
+
+class LayerNormFp32(nn.LayerNorm):
+    """Subclass torch's LayerNorm to handle fp16 (by casting to float32 and back)."""
+
+    def forward(self, x: torch.Tensor):
+        orig_type = x.dtype
+        x = F.layer_norm(x.to(torch.float32), self.normalized_shape, self.weight, self.bias, self.eps)
+        return x.to(orig_type)
+
+
+class LayerNorm(nn.LayerNorm):
+    """Subclass torch's LayerNorm (with cast back to input dtype)."""
+
+    def forward(self, x: torch.Tensor):
+        orig_type = x.dtype
+        x = F.layer_norm(x, self.normalized_shape, self.weight, self.bias, self.eps)
+        return x.to(orig_type)
+
+
+class QuickGELU(nn.Module):
+    # NOTE This is slower than nn.GELU or nn.SiLU and uses more GPU memory
+    def forward(self, x: torch.Tensor):
+        return x * torch.sigmoid(1.702 * x)
+
+
+class LayerScale(nn.Module):
+    def __init__(self, dim, init_values=1e-5, inplace=False):
+        super().__init__()
+        self.inplace = inplace
+        self.gamma = nn.Parameter(init_values * torch.ones(dim))
+
+    def forward(self, x):
+        return x.mul_(self.gamma) if self.inplace else x * self.gamma
+
+
+class PatchDropout(nn.Module):
+    """
+    https://arxiv.org/abs/2212.00794
+    """
+
+    def __init__(self, prob, exclude_first_token=True):
+        super().__init__()
+        assert 0 <= prob < 1.
+        self.prob = prob
+        self.exclude_first_token = exclude_first_token  # exclude CLS token
+
+    def forward(self, x):
+        if not self.training or self.prob == 0.:
+            return x
+
+        if self.exclude_first_token:
+            cls_tokens, x = x[:, :1], x[:, 1:]
+        else:
+            cls_tokens = torch.jit.annotate(torch.Tensor, x[:, :1])
+
+        batch = x.size()[0]
+        num_tokens = x.size()[1]
+
+        batch_indices = torch.arange(batch)
+        batch_indices = batch_indices[..., None]
+
+        keep_prob = 1 - self.prob
+        num_patches_keep = max(1, int(num_tokens * keep_prob))
+
+        rand = torch.randn(batch, num_tokens)
+        patch_indices_keep = rand.topk(num_patches_keep, dim=-1).indices
+
+        x = x[batch_indices, patch_indices_keep]
+
+        if self.exclude_first_token:
+            x = torch.cat((cls_tokens, x), dim=1)
+
+        return x
+
+
+class Attention(nn.Module):
+    def __init__(
+            self,
+            dim,
+            num_heads=8,
+            qkv_bias=True,
+            scaled_cosine=False,
+            scale_heads=False,
+            logit_scale_max=math.log(1. / 0.01),
+            attn_drop=0.,
+            proj_drop=0.
+    ):
+        super().__init__()
+        self.scaled_cosine = scaled_cosine
+        self.scale_heads = scale_heads
+        assert dim % num_heads == 0, 'dim should be divisible by num_heads'
+        self.num_heads = num_heads
+        self.head_dim = dim // num_heads
+        self.scale = self.head_dim ** -0.5
+        self.logit_scale_max = logit_scale_max
+
+        # keeping in_proj in this form (instead of nn.Linear) to match weight scheme of original
+        self.in_proj_weight = nn.Parameter(torch.randn((dim * 3, dim)) * self.scale)
+        if qkv_bias:
+            self.in_proj_bias = nn.Parameter(torch.zeros(dim * 3))
+        else:
+            self.in_proj_bias = None
+
+        if self.scaled_cosine:
+            self.logit_scale = nn.Parameter(torch.log(10 * torch.ones((num_heads, 1, 1))))
+        else:
+            self.logit_scale = None
+        self.attn_drop = nn.Dropout(attn_drop)
+        if self.scale_heads:
+            self.head_scale = nn.Parameter(torch.ones((num_heads, 1, 1)))
+        else:
+            self.head_scale = None
+        self.out_proj = nn.Linear(dim, dim)
+        self.out_drop = nn.Dropout(proj_drop)
+
+    def forward(self, x, attn_mask: Optional[torch.Tensor] = None):
+        L, N, C = x.shape
+        q, k, v = F.linear(x, self.in_proj_weight, self.in_proj_bias).chunk(3, dim=-1)
+        q = q.contiguous().view(L, N * self.num_heads, -1).transpose(0, 1)
+        k = k.contiguous().view(L, N * self.num_heads, -1).transpose(0, 1)
+        v = v.contiguous().view(L, N * self.num_heads, -1).transpose(0, 1)
+
+        if self.logit_scale is not None:
+            attn = torch.bmm(F.normalize(q, dim=-1), F.normalize(k, dim=-1).transpose(-1, -2))
+            logit_scale = torch.clamp(self.logit_scale, max=self.logit_scale_max).exp()
+            attn = attn.view(N, self.num_heads, L, L) * logit_scale
+            attn = attn.view(-1, L, L)
+        else:
+            q = q * self.scale
+            attn = torch.bmm(q, k.transpose(-1, -2))
+
+        if attn_mask is not None:
+            if attn_mask.dtype == torch.bool:
+                new_attn_mask = torch.zeros_like(attn_mask, dtype=q.dtype)
+                new_attn_mask.masked_fill_(attn_mask, float("-inf"))
+                attn_mask = new_attn_mask
+            attn += attn_mask
+
+        attn = attn.softmax(dim=-1)
+        attn = self.attn_drop(attn)
+
+        x = torch.bmm(attn, v)
+        if self.head_scale is not None:
+            x = x.view(N, self.num_heads, L, C) * self.head_scale
+            x = x.view(-1, L, C)
+        x = x.transpose(0, 1).reshape(L, N, C)
+        x = self.out_proj(x)
+        x = self.out_drop(x)
+        return x
+
+
+class AttentionalPooler(nn.Module):
+    def __init__(
+            self,
+            d_model: int,
+            context_dim: int,
+            n_head: int = 8,
+            n_queries: int = 256,
+            norm_layer: Callable = LayerNorm
+    ):
+        super().__init__()
+        self.query = nn.Parameter(torch.randn(n_queries, d_model))
+        self.attn = nn.MultiheadAttention(d_model, n_head, kdim=context_dim, vdim=context_dim)
+        self.ln_q = norm_layer(d_model)
+        self.ln_k = norm_layer(context_dim)
+
+    def forward(self, x: torch.Tensor):
+        x = self.ln_k(x).permute(1, 0, 2)  # NLD -> LND
+        N = x.shape[1]
+        q = self.ln_q(self.query)
+        out = self.attn(self._repeat(q, N), x, x, need_weights=False)[0]
+        return out.permute(1, 0, 2)  # LND -> NLD
+
+    def _repeat(self, query, N: int):
+        return query.unsqueeze(1).repeat(1, N, 1)
+
+
+class ResidualAttentionBlock(nn.Module):
+    def __init__(
+            self,
+            d_model: int,
+            n_head: int,
+            mlp_ratio: float = 4.0,
+            ls_init_value: float = None,
+            act_layer: Callable = nn.GELU,
+            norm_layer: Callable = LayerNorm,
+            is_cross_attention: bool = False,
+    ):
+        super().__init__()
+
+        self.ln_1 = norm_layer(d_model)
+        self.attn = nn.MultiheadAttention(d_model, n_head)
+        self.ls_1 = LayerScale(d_model, ls_init_value) if ls_init_value is not None else nn.Identity()
+        if is_cross_attention:
+            self.ln_1_kv = norm_layer(d_model)
+
+        self.ln_2 = norm_layer(d_model)
+        mlp_width = int(d_model * mlp_ratio)
+        self.mlp = nn.Sequential(OrderedDict([
+            ("c_fc", nn.Linear(d_model, mlp_width)),
+            ("gelu", act_layer()),
+            ("c_proj", nn.Linear(mlp_width, d_model))
+        ]))
+        self.ls_2 = LayerScale(d_model, ls_init_value) if ls_init_value is not None else nn.Identity()
+
+    def attention(
+            self,
+            q_x: torch.Tensor,
+            k_x: Optional[torch.Tensor] = None,
+            v_x: Optional[torch.Tensor] = None,
+            attn_mask: Optional[torch.Tensor] = None,
+    ):
+        k_x = k_x if k_x is not None else q_x
+        v_x = v_x if v_x is not None else q_x
+
+        attn_mask = attn_mask.to(q_x.dtype) if attn_mask is not None else None
+        return self.attn(
+            q_x, k_x, v_x, need_weights=False, attn_mask=attn_mask
+        )[0]
+
+    def forward(
+            self,
+            q_x: torch.Tensor,
+            k_x: Optional[torch.Tensor] = None,
+            v_x: Optional[torch.Tensor] = None,
+            attn_mask: Optional[torch.Tensor] = None,
+    ):
+        k_x = self.ln_1_kv(k_x) if hasattr(self, "ln_1_kv") and k_x is not None else None
+        v_x = self.ln_1_kv(v_x) if hasattr(self, "ln_1_kv") and v_x is not None else None
+
+        x = q_x + self.ls_1(self.attention(q_x=self.ln_1(q_x), k_x=k_x, v_x=v_x, attn_mask=attn_mask))
+        x = x + self.ls_2(self.mlp(self.ln_2(x)))
+        return x
+
+
+class CustomResidualAttentionBlock(nn.Module):
+    def __init__(
+            self,
+            d_model: int,
+            n_head: int,
+            mlp_ratio: float = 4.0,
+            ls_init_value: float = None,
+            act_layer: Callable = nn.GELU,
+            norm_layer: Callable = LayerNorm,
+            scale_cosine_attn: bool = False,
+            scale_heads: bool = False,
+            scale_attn: bool = False,
+            scale_fc: bool = False,
+    ):
+        super().__init__()
+
+        self.ln_1 = norm_layer(d_model)
+        self.attn = Attention(
+            d_model, n_head,
+            scaled_cosine=scale_cosine_attn,
+            scale_heads=scale_heads,
+        )
+        self.ln_attn = norm_layer(d_model) if scale_attn else nn.Identity()
+        self.ls_1 = LayerScale(d_model, ls_init_value) if ls_init_value is not None else nn.Identity()
+
+        self.ln_2 = norm_layer(d_model)
+        mlp_width = int(d_model * mlp_ratio)
+        self.mlp = nn.Sequential(OrderedDict([
+            ("c_fc", nn.Linear(d_model, mlp_width)),
+            ('ln', norm_layer(mlp_width) if scale_fc else nn.Identity()),
+            ("gelu", act_layer()),
+            ("c_proj", nn.Linear(mlp_width, d_model))
+        ]))
+        self.ls_2 = LayerScale(d_model, ls_init_value) if ls_init_value is not None else nn.Identity()
+
+    def forward(self, x: torch.Tensor, attn_mask: Optional[torch.Tensor] = None):
+        x = x + self.ls_1(self.ln_attn(self.attn(self.ln_1(x), attn_mask=attn_mask)))
+        x = x + self.ls_2(self.mlp(self.ln_2(x)))
+        return x
+
+
+class Transformer(nn.Module):
+    def __init__(
+            self,
+            width: int,
+            layers: int,
+            heads: int,
+            mlp_ratio: float = 4.0,
+            ls_init_value: float = None,
+            act_layer: Callable = nn.GELU,
+            norm_layer: Callable = LayerNorm,
+    ):
+        super().__init__()
+        self.width = width
+        self.layers = layers
+        self.grad_checkpointing = False
+
+        self.resblocks = nn.ModuleList([
+            ResidualAttentionBlock(
+                width, heads, mlp_ratio, ls_init_value=ls_init_value, act_layer=act_layer, norm_layer=norm_layer)
+            for _ in range(layers)
+        ])
+
+    def get_cast_dtype(self) -> torch.dtype:
+        if hasattr(self.resblocks[0].mlp.c_fc, 'int8_original_dtype'):
+            return self.resblocks[0].mlp.c_fc.int8_original_dtype
+        return self.resblocks[0].mlp.c_fc.weight.dtype
+
+    def forward(self, x: torch.Tensor, attn_mask: Optional[torch.Tensor] = None):
+        for r in self.resblocks:
+            if self.grad_checkpointing and not torch.jit.is_scripting():
+                # TODO: handle kwargs https://github.com/pytorch/pytorch/issues/79887#issuecomment-1161758372
+                x = checkpoint(r, x, None, None, attn_mask)
+            else:
+                x = r(x, attn_mask=attn_mask)
+        return x
+
+
+class VisionTransformer(nn.Module):
+    output_tokens: torch.jit.Final[bool]
+
+    def __init__(
+            self,
+            image_size: int,
+            patch_size: int,
+            width: int,
+            layers: int,
+            heads: int,
+            mlp_ratio: float,
+            ls_init_value: float = None,
+            global_average_pool: bool = False,
+            attentional_pool: bool = False,
+            n_queries: int = 256,
+            attn_pooler_heads: int = 8,
+            output_dim: int = 512,
+            patch_dropout: float = 0.,
+            input_patchnorm: bool = False,
+            act_layer: Callable = nn.GELU,
+            norm_layer: Callable = LayerNorm,
+            output_tokens: bool = False
+    ):
+        super().__init__()
+        self.output_tokens = output_tokens
+        image_height, image_width = self.image_size = to_2tuple(image_size)
+        patch_height, patch_width = self.patch_size = to_2tuple(patch_size)
+        self.grid_size = (image_height // patch_height, image_width // patch_width)
+        self.output_dim = output_dim
+
+        # whether to layernorm each patch, as done in dual patchnorm paper - https://arxiv.org/abs/2302.01327v1
+        self.input_patchnorm = input_patchnorm
+
+        if input_patchnorm:
+            patch_input_dim = patch_height * patch_width * 3
+            self.patchnorm_pre_ln = LayerNorm(patch_input_dim)
+            self.conv1 = nn.Linear(patch_input_dim, width)
+        else:
+            self.patchnorm_pre_ln = nn.Identity()
+            self.conv1 = nn.Conv2d(in_channels=3, out_channels=width, kernel_size=patch_size, stride=patch_size, bias=False)
+
+        # class embeddings and positional embeddings
+        scale = width ** -0.5
+        self.class_embedding = nn.Parameter(scale * torch.randn(width))
+        self.positional_embedding = nn.Parameter(scale * torch.randn(self.grid_size[0] * self.grid_size[1] + 1, width))
+
+        # setting a patch_dropout of 0. would mean it is disabled and this function would be the identity fn
+        self.patch_dropout = PatchDropout(patch_dropout) if patch_dropout > 0. else nn.Identity()
+
+        self.ln_pre = norm_layer(width)
+        self.transformer = Transformer(
+            width,
+            layers,
+            heads,
+            mlp_ratio,
+            ls_init_value=ls_init_value,
+            act_layer=act_layer,
+            norm_layer=norm_layer,
+        )
+
+        self.global_average_pool = global_average_pool
+        if attentional_pool:
+            self.attn_pool = AttentionalPooler(output_dim, width, n_head=attn_pooler_heads, n_queries=n_queries)
+            self.ln_post = norm_layer(output_dim)
+            self.proj = nn.Parameter(scale * torch.randn(output_dim, output_dim))
+        else:
+            self.attn_pool = None
+            self.ln_post = norm_layer(width)
+            self.proj = nn.Parameter(scale * torch.randn(width, output_dim))
+
+        self.init_parameters()
+
+    def lock(self, unlocked_groups=0, freeze_bn_stats=False):
+        for param in self.parameters():
+            param.requires_grad = False
+
+        if unlocked_groups != 0:
+            groups = [
+                [
+                    self.conv1,
+                    self.class_embedding,
+                    self.positional_embedding,
+                    self.ln_pre,
+                ],
+                *self.transformer.resblocks[:-1],
+                [
+                    self.transformer.resblocks[-1],
+                    self.ln_post,
+                ],
+                self.proj,
+            ]
+
+            def _unlock(x):
+                if isinstance(x, Sequence):
+                    for g in x:
+                        _unlock(g)
+                else:
+                    if isinstance(x, torch.nn.Parameter):
+                        x.requires_grad = True
+                    else:
+                        for p in x.parameters():
+                            p.requires_grad = True
+
+            _unlock(groups[-unlocked_groups:])
+
+    def init_parameters(self):
+        # FIXME OpenAI CLIP did not define an init for the VisualTransformer
+        # TODO experiment if default PyTorch init, below, or alternate init is best.
+
+        # nn.init.normal_(self.class_embedding, std=self.scale)
+        # nn.init.normal_(self.positional_embedding, std=self.scale)
+        #
+        # proj_std = (self.transformer.width ** -0.5) * ((2 * self.transformer.layers) ** -0.5)
+        # attn_std = self.transformer.width ** -0.5
+        # fc_std = (2 * self.transformer.width) ** -0.5
+        # for block in self.transformer.resblocks:
+        #     nn.init.normal_(block.attn.in_proj_weight, std=attn_std)
+        #     nn.init.normal_(block.attn.out_proj.weight, std=proj_std)
+        #     nn.init.normal_(block.mlp.c_fc.weight, std=fc_std)
+        #     nn.init.normal_(block.mlp.c_proj.weight, std=proj_std)
+        #
+        # if self.text_projection is not None:
+        #     nn.init.normal_(self.text_projection, std=self.scale)
+        pass
+
+    @torch.jit.ignore
+    def set_grad_checkpointing(self, enable=True):
+        self.transformer.grad_checkpointing = enable
+
+    def _global_pool(self, x: torch.Tensor) -> Tuple[torch.Tensor, torch.Tensor]:
+        if self.global_average_pool:
+            return x.mean(dim=1), x
+        else:
+            return x[:, 0], x[:, 1:]
+
+    def forward(self, x: torch.Tensor):
+
+        # to patches - whether to use dual patchnorm - https://arxiv.org/abs/2302.01327v1
+        if self.input_patchnorm:
+            # einops - rearrange(x, 'b c (h p1) (w p2) -> b (h w) (c p1 p2)')
+            x = x.reshape(x.shape[0], x.shape[1], self.grid_size[0], self.patch_size[0], self.grid_size[1], self.patch_size[1])
+            x = x.permute(0, 2, 4, 1, 3, 5)
+            x = x.reshape(x.shape[0], self.grid_size[0] * self.grid_size[1], -1)
+            x = self.patchnorm_pre_ln(x)
+            x = self.conv1(x)
+        else:
+            x = self.conv1(x)  # shape = [*, width, grid, grid]
+            x = x.reshape(x.shape[0], x.shape[1], -1)  # shape = [*, width, grid ** 2]
+            x = x.permute(0, 2, 1)  # shape = [*, grid ** 2, width]
+
+        # class embeddings and positional embeddings
+        x = torch.cat(
+            [self.class_embedding.to(x.dtype) + torch.zeros(x.shape[0], 1, x.shape[-1], dtype=x.dtype, device=x.device),
+             x], dim=1)  # shape = [*, grid ** 2 + 1, width]
+        x = x + self.positional_embedding.to(x.dtype)
+
+        # a patch_dropout of 0. would mean it is disabled and this function would do nothing but return what was passed in
+        x = self.patch_dropout(x)
+        x = self.ln_pre(x)
+
+        x = x.permute(1, 0, 2)  # NLD -> LND
+        x = self.transformer(x)
+        x = x.permute(1, 0, 2)  # LND -> NLD
+
+        if self.attn_pool is not None:
+            x = self.attn_pool(x)
+            x = self.ln_post(x)
+            pooled, tokens = self._global_pool(x)
+        else:
+            pooled, tokens = self._global_pool(x)
+            pooled = self.ln_post(pooled)
+
+        if self.proj is not None:
+            pooled = pooled @ self.proj
+
+        if self.output_tokens:
+            return pooled, tokens
+        
+        return pooled
+
+
+class TextTransformer(nn.Module):
+    output_tokens: torch.jit.Final[bool]
+
+    def __init__(
+            self,
+            context_length: int = 77,
+            vocab_size: int = 49408,
+            width: int = 512,
+            heads: int = 8,
+            layers: int = 12,
+            ls_init_value: float = None,
+            output_dim: int = 512,
+            act_layer: Callable = nn.GELU,
+            norm_layer: Callable = LayerNorm,
+            embed_cls: bool = False,
+            pad_id: int = 0,
+            output_tokens: bool = False,
+    ):
+        super().__init__()
+        self.output_tokens = output_tokens
+        self.num_pos = self.context_length = context_length
+        self.vocab_size = vocab_size
+        self.width = width
+        self.output_dim = output_dim
+        self.heads = heads
+        self.pad_id = pad_id
+
+        self.text_projection = nn.Parameter(torch.empty(width, output_dim))
+
+        if embed_cls:
+            self.cls_emb = nn.Parameter(torch.empty(width))
+            self.num_pos += 1
+        else:
+            self.cls_emb = None
+
+        self.token_embedding = nn.Embedding(vocab_size, width)
+        self.positional_embedding = nn.Parameter(torch.empty(self.num_pos, width))
+        self.transformer = Transformer(
+            width=width,
+            layers=layers,
+            heads=heads,
+            ls_init_value=ls_init_value,
+            act_layer=act_layer,
+            norm_layer=norm_layer,
+        )
+        self.ln_final = norm_layer(width)
+
+        self.register_buffer('attn_mask', self.build_attention_mask(), persistent=False)
+
+        self.init_parameters()
+
+    def init_parameters(self):
+        nn.init.normal_(self.token_embedding.weight, std=0.02)
+        nn.init.normal_(self.positional_embedding, std=0.01)
+        if self.cls_emb is not None:
+            nn.init.normal_(self.cls_emb, std=0.01)
+
+        proj_std = (self.transformer.width ** -0.5) * ((2 * self.transformer.layers) ** -0.5)
+        attn_std = self.transformer.width ** -0.5
+        fc_std = (2 * self.transformer.width) ** -0.5
+        for block in self.transformer.resblocks:
+            nn.init.normal_(block.attn.in_proj_weight, std=attn_std)
+            nn.init.normal_(block.attn.out_proj.weight, std=proj_std)
+            nn.init.normal_(block.mlp.c_fc.weight, std=fc_std)
+            nn.init.normal_(block.mlp.c_proj.weight, std=proj_std)
+
+        if self.text_projection is not None:
+            nn.init.normal_(self.text_projection, std=self.transformer.width ** -0.5)
+
+    @torch.jit.ignore
+    def set_grad_checkpointing(self, enable=True):
+        self.transformer.grad_checkpointing = enable
+
+    def build_attention_mask(self):
+        # lazily create causal attention mask, with full attention between the tokens
+        # pytorch uses additive attention mask; fill with -inf
+        mask = torch.empty(self.num_pos, self.num_pos)
+        mask.fill_(float("-inf"))
+        mask.triu_(1)  # zero out the lower diagonal
+        return mask
+
+    def build_cls_mask(self, text, cast_dtype: torch.dtype):
+        cls_mask = (text != self.pad_id).unsqueeze(1)
+        cls_mask = F.pad(cls_mask, (1, 0, cls_mask.shape[2], 0), value=1.0)
+        additive_mask = torch.empty(cls_mask.shape, dtype=cast_dtype, device=cls_mask.device)
+        additive_mask.fill_(0)
+        additive_mask.masked_fill_(~cls_mask, float("-inf"))
+        additive_mask = torch.repeat_interleave(additive_mask, self.heads, 0)
+        return additive_mask
+
+    def _repeat(self, t, N: int):
+        return t.reshape(1, 1, -1).repeat(N, 1, 1)
+
+    def forward(self, text):
+        cast_dtype = self.transformer.get_cast_dtype()
+        seq_len = text.shape[1]
+
+        x = self.token_embedding(text).to(cast_dtype)  # [batch_size, n_ctx, d_model]
+        attn_mask = self.attn_mask
+        if self.cls_emb is not None:
+            seq_len += 1
+            x = torch.cat([x, self._repeat(self.cls_emb, x.shape[0])], dim=1)
+            cls_mask = self.build_cls_mask(text, cast_dtype)
+            attn_mask = attn_mask[None, :seq_len, :seq_len] + cls_mask[:, :seq_len, :seq_len]
+
+        x = x + self.positional_embedding[:seq_len].to(cast_dtype)
+        x = x.permute(1, 0, 2)  # NLD -> LND
+        x = self.transformer(x, attn_mask=attn_mask)
+        x = x.permute(1, 0, 2)  # LND -> NLD
+
+        # x.shape = [batch_size, n_ctx, transformer.width]
+        # take features from the eot embedding (eot_token is the highest number in each sequence)
+        if self.cls_emb is not None:
+            pooled, tokens = x[:, -1], x[:, :-1]
+            pooled = self.ln_final(pooled)
+        else:
+            x = self.ln_final(x)
+            pooled, tokens = x[torch.arange(x.shape[0]), text.argmax(dim=-1)], x
+
+        if self.text_projection is not None:
+            pooled = pooled @ self.text_projection
+
+        if self.output_tokens:
+            return pooled, tokens
+
+        return pooled
+
+
+class MultimodalTransformer(Transformer):
+    def __init__(
+            self,
+            width: int,
+            layers: int,
+            heads: int,
+            context_length: int = 77,
+            mlp_ratio: float = 4.0,
+            ls_init_value: float = None,
+            act_layer: Callable = nn.GELU,
+            norm_layer: Callable = LayerNorm,
+            output_dim: int = 512,
+    ):
+
+        super().__init__(
+            width=width,
+            layers=layers,
+            heads=heads,
+            mlp_ratio=mlp_ratio,
+            ls_init_value=ls_init_value,
+            act_layer=act_layer,
+            norm_layer=norm_layer,
+        )
+        self.context_length = context_length
+        self.cross_attn = nn.ModuleList([
+            ResidualAttentionBlock(
+                width,
+                heads,
+                mlp_ratio,
+                ls_init_value=ls_init_value,
+                act_layer=act_layer,
+                norm_layer=norm_layer,
+                is_cross_attention=True,
+            )
+            for _ in range(layers)
+        ])
+
+        self.register_buffer('attn_mask', self.build_attention_mask(), persistent=False)
+
+        self.ln_final = norm_layer(width)
+        self.text_projection = nn.Parameter(torch.empty(width, output_dim))
+
+    def init_parameters(self):
+        proj_std = (self.transformer.width ** -0.5) * ((2 * self.transformer.layers) ** -0.5)
+        attn_std = self.transformer.width ** -0.5
+        fc_std = (2 * self.transformer.width) ** -0.5
+        for block in self.transformer.resblocks:
+            nn.init.normal_(block.attn.in_proj_weight, std=attn_std)
+            nn.init.normal_(block.attn.out_proj.weight, std=proj_std)
+            nn.init.normal_(block.mlp.c_fc.weight, std=fc_std)
+            nn.init.normal_(block.mlp.c_proj.weight, std=proj_std)
+        for block in self.transformer.cross_attn:
+            nn.init.normal_(block.attn.in_proj_weight, std=attn_std)
+            nn.init.normal_(block.attn.out_proj.weight, std=proj_std)
+            nn.init.normal_(block.mlp.c_fc.weight, std=fc_std)
+            nn.init.normal_(block.mlp.c_proj.weight, std=proj_std)
+
+        if self.text_projection is not None:
+            nn.init.normal_(self.text_projection, std=self.transformer.width ** -0.5)
+
+    def build_attention_mask(self):
+        # lazily create causal attention mask, with full attention between the tokens
+        # pytorch uses additive attention mask; fill with -inf
+        mask = torch.empty(self.context_length, self.context_length)
+        mask.fill_(float("-inf"))
+        mask.triu_(1)  # zero out the lower diagonal
+        return mask
+
+    def forward(self, image_embs, text_embs):
+        text_embs = text_embs.permute(1, 0, 2)  # NLD -> LNDsq
+        image_embs = image_embs.permute(1, 0, 2)  # NLD -> LND
+        seq_len = text_embs.shape[0]
+
+        for resblock, cross_attn in zip(self.resblocks, self.cross_attn):
+            if self.grad_checkpointing and not torch.jit.is_scripting():
+                # TODO: handle kwargs https://github.com/pytorch/pytorch/issues/79887#issuecomment-1161758372
+                text_embs = checkpoint(resblock, text_embs, None, None, self.attn_mask[:seq_len, :seq_len])
+                text_embs = checkpoint(cross_attn, text_embs, image_embs, image_embs, None)
+            else:
+                text_embs = resblock(text_embs, attn_mask=self.attn_mask[:seq_len, :seq_len])
+                text_embs = cross_attn(text_embs, k_x=image_embs, v_x=image_embs)
+
+        x = text_embs.permute(1, 0, 2)  # LND -> NLD
+        x = self.ln_final(x)
+
+        if self.text_projection is not None:
+            x = x @ self.text_projection
+
+        return x
+
+    @torch.jit.ignore
+    def set_grad_checkpointing(self, enable=True):
+        self.grad_checkpointing = enable

model/scunet.py

@@ -0,0 +1,264 @@
+import torch
+import torch.nn as nn
+import numpy as np
+from einops import rearrange 
+from einops.layers.torch import Rearrange
+from timm.models.layers import trunc_normal_, DropPath
+
+
+class WMSA(nn.Module):
+    """ Self-attention module in Swin Transformer
+    """
+
+    def __init__(self, input_dim, output_dim, head_dim, window_size, type):
+        super(WMSA, self).__init__()
+        self.input_dim = input_dim
+        self.output_dim = output_dim
+        self.head_dim = head_dim 
+        self.scale = self.head_dim ** -0.5
+        self.n_heads = input_dim//head_dim
+        self.window_size = window_size
+        self.type=type
+        self.embedding_layer = nn.Linear(self.input_dim, 3*self.input_dim, bias=True)
+
+        # TODO recover
+        # self.relative_position_params = nn.Parameter(torch.zeros(self.n_heads, 2 * window_size - 1, 2 * window_size -1))
+        self.relative_position_params = nn.Parameter(torch.zeros((2 * window_size - 1)*(2 * window_size -1), self.n_heads))
+
+        self.linear = nn.Linear(self.input_dim, self.output_dim)
+
+        trunc_normal_(self.relative_position_params, std=.02)
+        self.relative_position_params = torch.nn.Parameter(self.relative_position_params.view(2*window_size-1, 2*window_size-1, self.n_heads).transpose(1,2).transpose(0,1))
+
+    def generate_mask(self, h, w, p, shift):
+        """ generating the mask of SW-MSA
+        Args:
+            shift: shift parameters in CyclicShift.
+        Returns:
+            attn_mask: should be (1 1 w p p),
+        """
+        # supporting sqaure.
+        attn_mask = torch.zeros(h, w, p, p, p, p, dtype=torch.bool, device=self.relative_position_params.device)
+        if self.type == 'W':
+            return attn_mask
+
+        s = p - shift
+        attn_mask[-1, :, :s, :, s:, :] = True
+        attn_mask[-1, :, s:, :, :s, :] = True
+        attn_mask[:, -1, :, :s, :, s:] = True
+        attn_mask[:, -1, :, s:, :, :s] = True
+        attn_mask = rearrange(attn_mask, 'w1 w2 p1 p2 p3 p4 -> 1 1 (w1 w2) (p1 p2) (p3 p4)')
+        return attn_mask
+
+    def forward(self, x):
+        """ Forward pass of Window Multi-head Self-attention module.
+        Args:
+            x: input tensor with shape of [b h w c];
+            attn_mask: attention mask, fill -inf where the value is True; 
+        Returns:
+            output: tensor shape [b h w c]
+        """
+        if self.type!='W': x = torch.roll(x, shifts=(-(self.window_size//2), -(self.window_size//2)), dims=(1,2))
+        x = rearrange(x, 'b (w1 p1) (w2 p2) c -> b w1 w2 p1 p2 c', p1=self.window_size, p2=self.window_size)
+        h_windows = x.size(1)
+        w_windows = x.size(2)
+        # sqaure validation
+        # assert h_windows == w_windows
+
+        x = rearrange(x, 'b w1 w2 p1 p2 c -> b (w1 w2) (p1 p2) c', p1=self.window_size, p2=self.window_size)
+        qkv = self.embedding_layer(x)
+        q, k, v = rearrange(qkv, 'b nw np (threeh c) -> threeh b nw np c', c=self.head_dim).chunk(3, dim=0)
+        sim = torch.einsum('hbwpc,hbwqc->hbwpq', q, k) * self.scale
+        # Adding learnable relative embedding
+        sim = sim + rearrange(self.relative_embedding(), 'h p q -> h 1 1 p q')
+        # Using Attn Mask to distinguish different subwindows.
+        if self.type != 'W':
+            attn_mask = self.generate_mask(h_windows, w_windows, self.window_size, shift=self.window_size//2)
+            sim = sim.masked_fill_(attn_mask, float("-inf"))
+
+        probs = nn.functional.softmax(sim, dim=-1)
+        output = torch.einsum('hbwij,hbwjc->hbwic', probs, v)
+        output = rearrange(output, 'h b w p c -> b w p (h c)')
+        output = self.linear(output)
+        output = rearrange(output, 'b (w1 w2) (p1 p2) c -> b (w1 p1) (w2 p2) c', w1=h_windows, p1=self.window_size)
+
+        if self.type!='W': output = torch.roll(output, shifts=(self.window_size//2, self.window_size//2), dims=(1,2))
+        return output
+
+    def relative_embedding(self):
+        cord = torch.tensor(np.array([[i, j] for i in range(self.window_size) for j in range(self.window_size)]))
+        relation = cord[:, None, :] - cord[None, :, :] + self.window_size -1
+        # negative is allowed
+        return self.relative_position_params[:, relation[:,:,0].long(), relation[:,:,1].long()]
+
+
+class Block(nn.Module):
+    def __init__(self, input_dim, output_dim, head_dim, window_size, drop_path, type='W', input_resolution=None):
+        """ SwinTransformer Block
+        """
+        super(Block, self).__init__()
+        self.input_dim = input_dim
+        self.output_dim = output_dim
+        assert type in ['W', 'SW']
+        self.type = type
+        if input_resolution <= window_size:
+            self.type = 'W'
+
+        print("Block Initial Type: {}, drop_path_rate:{:.6f}".format(self.type, drop_path))
+        self.ln1 = nn.LayerNorm(input_dim)
+        self.msa = WMSA(input_dim, input_dim, head_dim, window_size, self.type)
+        self.drop_path = DropPath(drop_path) if drop_path > 0. else nn.Identity()
+        self.ln2 = nn.LayerNorm(input_dim)
+        self.mlp = nn.Sequential(
+            nn.Linear(input_dim, 4 * input_dim),
+            nn.GELU(),
+            nn.Linear(4 * input_dim, output_dim),
+        )
+
+    def forward(self, x):
+        x = x + self.drop_path(self.msa(self.ln1(x)))
+        x = x + self.drop_path(self.mlp(self.ln2(x)))
+        return x
+
+
+class ConvTransBlock(nn.Module):
+    def __init__(self, conv_dim, trans_dim, head_dim, window_size, drop_path, type='W', input_resolution=None):
+        """ SwinTransformer and Conv Block
+        """
+        super(ConvTransBlock, self).__init__()
+        self.conv_dim = conv_dim
+        self.trans_dim = trans_dim
+        self.head_dim = head_dim
+        self.window_size = window_size
+        self.drop_path = drop_path
+        self.type = type
+        self.input_resolution = input_resolution
+
+        assert self.type in ['W', 'SW']
+        if self.input_resolution <= self.window_size:
+            self.type = 'W'
+
+        self.trans_block = Block(self.trans_dim, self.trans_dim, self.head_dim, self.window_size, self.drop_path, self.type, self.input_resolution)
+        self.conv1_1 = nn.Conv2d(self.conv_dim+self.trans_dim, self.conv_dim+self.trans_dim, 1, 1, 0, bias=True)
+        self.conv1_2 = nn.Conv2d(self.conv_dim+self.trans_dim, self.conv_dim+self.trans_dim, 1, 1, 0, bias=True)
+
+        self.conv_block = nn.Sequential(
+                nn.Conv2d(self.conv_dim, self.conv_dim, 3, 1, 1, bias=False),
+                nn.ReLU(True),
+                nn.Conv2d(self.conv_dim, self.conv_dim, 3, 1, 1, bias=False)
+                )
+
+    def forward(self, x):
+        conv_x, trans_x = torch.split(self.conv1_1(x), (self.conv_dim, self.trans_dim), dim=1)
+        conv_x = self.conv_block(conv_x) + conv_x
+        trans_x = Rearrange('b c h w -> b h w c')(trans_x)
+        trans_x = self.trans_block(trans_x)
+        trans_x = Rearrange('b h w c -> b c h w')(trans_x)
+        res = self.conv1_2(torch.cat((conv_x, trans_x), dim=1))
+        x = x + res
+
+        return x
+
+
+class SCUNet(nn.Module):
+
+    def __init__(self, in_nc=3, config=[2,2,2,2,2,2,2], dim=64, drop_path_rate=0.0, input_resolution=256):
+        super(SCUNet, self).__init__()
+        self.config = config
+        self.dim = dim
+        self.head_dim = 32
+        self.window_size = 8
+
+        # drop path rate for each layer
+        dpr = [x.item() for x in torch.linspace(0, drop_path_rate, sum(config))]
+
+        self.m_head = [nn.Conv2d(in_nc, dim, 3, 1, 1, bias=False)]
+
+        begin = 0
+        self.m_down1 = [ConvTransBlock(dim//2, dim//2, self.head_dim, self.window_size, dpr[i+begin], 'W' if not i%2 else 'SW', input_resolution) 
+                      for i in range(config[0])] + \
+                      [nn.Conv2d(dim, 2*dim, 2, 2, 0, bias=False)]
+
+        begin += config[0]
+        self.m_down2 = [ConvTransBlock(dim, dim, self.head_dim, self.window_size, dpr[i+begin], 'W' if not i%2 else 'SW', input_resolution//2)
+                      for i in range(config[1])] + \
+                      [nn.Conv2d(2*dim, 4*dim, 2, 2, 0, bias=False)]
+
+        begin += config[1]
+        self.m_down3 = [ConvTransBlock(2*dim, 2*dim, self.head_dim, self.window_size, dpr[i+begin], 'W' if not i%2 else 'SW',input_resolution//4)
+                      for i in range(config[2])] + \
+                      [nn.Conv2d(4*dim, 8*dim, 2, 2, 0, bias=False)]
+
+        begin += config[2]
+        self.m_body = [ConvTransBlock(4*dim, 4*dim, self.head_dim, self.window_size, dpr[i+begin], 'W' if not i%2 else 'SW', input_resolution//8)
+                    for i in range(config[3])]
+
+        begin += config[3]
+        self.m_up3 = [nn.ConvTranspose2d(8*dim, 4*dim, 2, 2, 0, bias=False),] + \
+                      [ConvTransBlock(2*dim, 2*dim, self.head_dim, self.window_size, dpr[i+begin], 'W' if not i%2 else 'SW',input_resolution//4)
+                      for i in range(config[4])]
+                      
+        begin += config[4]
+        self.m_up2 = [nn.ConvTranspose2d(4*dim, 2*dim, 2, 2, 0, bias=False),] + \
+                      [ConvTransBlock(dim, dim, self.head_dim, self.window_size, dpr[i+begin], 'W' if not i%2 else 'SW', input_resolution//2)
+                      for i in range(config[5])]
+                      
+        begin += config[5]
+        self.m_up1 = [nn.ConvTranspose2d(2*dim, dim, 2, 2, 0, bias=False),] + \
+                    [ConvTransBlock(dim//2, dim//2, self.head_dim, self.window_size, dpr[i+begin], 'W' if not i%2 else 'SW', input_resolution) 
+                      for i in range(config[6])]
+
+        self.m_tail = [nn.Conv2d(dim, in_nc, 3, 1, 1, bias=False)]
+
+        self.m_head = nn.Sequential(*self.m_head)
+        self.m_down1 = nn.Sequential(*self.m_down1)
+        self.m_down2 = nn.Sequential(*self.m_down2)
+        self.m_down3 = nn.Sequential(*self.m_down3)
+        self.m_body = nn.Sequential(*self.m_body)
+        self.m_up3 = nn.Sequential(*self.m_up3)
+        self.m_up2 = nn.Sequential(*self.m_up2)
+        self.m_up1 = nn.Sequential(*self.m_up1)
+        self.m_tail = nn.Sequential(*self.m_tail)  
+        #self.apply(self._init_weights)
+
+    def forward(self, x0):
+
+        h, w = x0.size()[-2:]
+        paddingBottom = int(np.ceil(h/64)*64-h)
+        paddingRight = int(np.ceil(w/64)*64-w)
+        x0 = nn.ReplicationPad2d((0, paddingRight, 0, paddingBottom))(x0)
+
+        x1 = self.m_head(x0)
+        x2 = self.m_down1(x1)
+        x3 = self.m_down2(x2)
+        x4 = self.m_down3(x3)
+        x = self.m_body(x4)
+        x = self.m_up3(x+x4)
+        x = self.m_up2(x+x3)
+        x = self.m_up1(x+x2)
+        x = self.m_tail(x+x1)
+
+        x = x[..., :h, :w]
+        
+        return x
+
+
+    def _init_weights(self, m):
+        if isinstance(m, nn.Linear):
+            trunc_normal_(m.weight, std=.02)
+            if m.bias is not None:
+                nn.init.constant_(m.bias, 0)
+        elif isinstance(m, nn.LayerNorm):
+            nn.init.constant_(m.bias, 0)
+            nn.init.constant_(m.weight, 1.0)
+
+
+
+if __name__ == '__main__':
+
+    # torch.cuda.empty_cache()
+    net = SCUNet()
+
+    x = torch.randn((2, 3, 64, 128))
+    x = net(x)
+    print(x.shape)

model/swinir.py

@@ -0,0 +1,905 @@
+# -----------------------------------------------------------------------------------
+# SwinIR: Image Restoration Using Swin Transformer, https://arxiv.org/abs/2108.10257
+# Originally Written by Ze Liu, Modified by Jingyun Liang.
+# -----------------------------------------------------------------------------------
+
+# Originally borrowed from DifFace (https://github.com/zsyOAOA/DifFace/blob/master/models/swinir.py)
+
+import math
+from typing import Set
+
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+import torch.utils.checkpoint as checkpoint
+from timm.models.layers import DropPath, to_2tuple, trunc_normal_
+
+
+class Mlp(nn.Module):
+    def __init__(self, in_features, hidden_features=None, out_features=None, act_layer=nn.GELU, drop=0.):
+        super().__init__()
+        out_features = out_features or in_features
+        hidden_features = hidden_features or in_features
+        self.fc1 = nn.Linear(in_features, hidden_features)
+        self.act = act_layer()
+        self.fc2 = nn.Linear(hidden_features, out_features)
+        self.drop = nn.Dropout(drop)
+
+    def forward(self, x):
+        x = self.fc1(x)
+        x = self.act(x)
+        x = self.drop(x)
+        x = self.fc2(x)
+        x = self.drop(x)
+        return x
+
+
+def window_partition(x, window_size):
+    """
+    Args:
+        x: (B, H, W, C)
+        window_size (int): window size
+
+    Returns:
+        windows: (num_windows*B, window_size, window_size, C)
+    """
+    B, H, W, C = x.shape
+    x = x.view(B, H // window_size, window_size, W // window_size, window_size, C)
+    windows = x.permute(0, 1, 3, 2, 4, 5).contiguous().view(-1, window_size, window_size, C)
+    return windows
+
+
+def window_reverse(windows, window_size, H, W):
+    """
+    Args:
+        windows: (num_windows*B, window_size, window_size, C)
+        window_size (int): Window size
+        H (int): Height of image
+        W (int): Width of image
+
+    Returns:
+        x: (B, H, W, C)
+    """
+    B = int(windows.shape[0] / (H * W / window_size / window_size))
+    x = windows.view(B, H // window_size, W // window_size, window_size, window_size, -1)
+    x = x.permute(0, 1, 3, 2, 4, 5).contiguous().view(B, H, W, -1)
+    return x
+
+
+class WindowAttention(nn.Module):
+    r""" Window based multi-head self attention (W-MSA) module with relative position bias.
+    It supports both of shifted and non-shifted window.
+
+    Args:
+        dim (int): Number of input channels.
+        window_size (tuple[int]): The height and width of the window.
+        num_heads (int): Number of attention heads.
+        qkv_bias (bool, optional):  If True, add a learnable bias to query, key, value. Default: True
+        qk_scale (float | None, optional): Override default qk scale of head_dim ** -0.5 if set
+        attn_drop (float, optional): Dropout ratio of attention weight. Default: 0.0
+        proj_drop (float, optional): Dropout ratio of output. Default: 0.0
+    """
+
+    def __init__(self, dim, window_size, num_heads, qkv_bias=True, qk_scale=None, attn_drop=0., proj_drop=0.):
+
+        super().__init__()
+        self.dim = dim
+        self.window_size = window_size  # Wh, Ww
+        self.num_heads = num_heads
+        head_dim = dim // num_heads
+        self.scale = qk_scale or head_dim ** -0.5
+
+        # define a parameter table of relative position bias
+        self.relative_position_bias_table = nn.Parameter(
+            torch.zeros((2 * window_size[0] - 1) * (2 * window_size[1] - 1), num_heads))  # 2*Wh-1 * 2*Ww-1, nH
+
+        # get pair-wise relative position index for each token inside the window
+        coords_h = torch.arange(self.window_size[0])
+        coords_w = torch.arange(self.window_size[1])
+        # coords = torch.stack(torch.meshgrid([coords_h, coords_w]))  # 2, Wh, Ww
+        # Fix: Pass indexing="ij" to avoid warning
+        coords = torch.stack(torch.meshgrid([coords_h, coords_w], indexing="ij"))  # 2, Wh, Ww
+        coords_flatten = torch.flatten(coords, 1)  # 2, Wh*Ww
+        relative_coords = coords_flatten[:, :, None] - coords_flatten[:, None, :]  # 2, Wh*Ww, Wh*Ww
+        relative_coords = relative_coords.permute(1, 2, 0).contiguous()  # Wh*Ww, Wh*Ww, 2
+        relative_coords[:, :, 0] += self.window_size[0] - 1  # shift to start from 0
+        relative_coords[:, :, 1] += self.window_size[1] - 1
+        relative_coords[:, :, 0] *= 2 * self.window_size[1] - 1
+        relative_position_index = relative_coords.sum(-1)  # Wh*Ww, Wh*Ww
+        self.register_buffer("relative_position_index", relative_position_index)
+
+        self.qkv = nn.Linear(dim, dim * 3, bias=qkv_bias)
+        self.attn_drop = nn.Dropout(attn_drop)
+        self.proj = nn.Linear(dim, dim)
+
+        self.proj_drop = nn.Dropout(proj_drop)
+
+        trunc_normal_(self.relative_position_bias_table, std=.02)
+        self.softmax = nn.Softmax(dim=-1)
+
+    def forward(self, x, mask=None):
+        """
+        Args:
+            x: input features with shape of (num_windows*B, N, C)
+            mask: (0/-inf) mask with shape of (num_windows, Wh*Ww, Wh*Ww) or None
+        """
+        B_, N, C = x.shape
+        qkv = self.qkv(x).reshape(B_, N, 3, self.num_heads, C // self.num_heads).permute(2, 0, 3, 1, 4)
+        q, k, v = qkv[0], qkv[1], qkv[2]  # make torchscript happy (cannot use tensor as tuple)
+
+        q = q * self.scale
+        attn = (q @ k.transpose(-2, -1))
+
+        relative_position_bias = self.relative_position_bias_table[self.relative_position_index.view(-1)].view(
+            self.window_size[0] * self.window_size[1], self.window_size[0] * self.window_size[1], -1)  # Wh*Ww,Wh*Ww,nH
+        relative_position_bias = relative_position_bias.permute(2, 0, 1).contiguous()  # nH, Wh*Ww, Wh*Ww
+        attn = attn + relative_position_bias.unsqueeze(0)
+
+        if mask is not None:
+            nW = mask.shape[0]
+            attn = attn.view(B_ // nW, nW, self.num_heads, N, N) + mask.unsqueeze(1).unsqueeze(0)
+            attn = attn.view(-1, self.num_heads, N, N)
+            attn = self.softmax(attn)
+        else:
+            attn = self.softmax(attn)
+
+        attn = self.attn_drop(attn)
+
+        x = (attn @ v).transpose(1, 2).reshape(B_, N, C)
+        x = self.proj(x)
+        x = self.proj_drop(x)
+        return x
+
+    def extra_repr(self) -> str:
+        return f'dim={self.dim}, window_size={self.window_size}, num_heads={self.num_heads}'
+
+    def flops(self, N):
+        # calculate flops for 1 window with token length of N
+        flops = 0
+        # qkv = self.qkv(x)
+        flops += N * self.dim * 3 * self.dim
+        # attn = (q @ k.transpose(-2, -1))
+        flops += self.num_heads * N * (self.dim // self.num_heads) * N
+        #  x = (attn @ v)
+        flops += self.num_heads * N * N * (self.dim // self.num_heads)
+        # x = self.proj(x)
+        flops += N * self.dim * self.dim
+        return flops
+
+
+class SwinTransformerBlock(nn.Module):
+    r""" Swin Transformer Block.
+
+    Args:
+        dim (int): Number of input channels.
+        input_resolution (tuple[int]): Input resulotion.
+        num_heads (int): Number of attention heads.
+        window_size (int): Window size.
+        shift_size (int): Shift size for SW-MSA.
+        mlp_ratio (float): Ratio of mlp hidden dim to embedding dim.
+        qkv_bias (bool, optional): If True, add a learnable bias to query, key, value. Default: True
+        qk_scale (float | None, optional): Override default qk scale of head_dim ** -0.5 if set.
+        drop (float, optional): Dropout rate. Default: 0.0
+        attn_drop (float, optional): Attention dropout rate. Default: 0.0
+        drop_path (float, optional): Stochastic depth rate. Default: 0.0
+        act_layer (nn.Module, optional): Activation layer. Default: nn.GELU
+        norm_layer (nn.Module, optional): Normalization layer.  Default: nn.LayerNorm
+    """
+
+    def __init__(self, dim, input_resolution, num_heads, window_size=7, shift_size=0,
+                 mlp_ratio=4., qkv_bias=True, qk_scale=None, drop=0., attn_drop=0., drop_path=0.,
+                 act_layer=nn.GELU, norm_layer=nn.LayerNorm):
+        super().__init__()
+        self.dim = dim
+        self.input_resolution = input_resolution
+        self.num_heads = num_heads
+        self.window_size = window_size
+        self.shift_size = shift_size
+        self.mlp_ratio = mlp_ratio
+        if min(self.input_resolution) <= self.window_size:
+            # if window size is larger than input resolution, we don't partition windows
+            self.shift_size = 0
+            self.window_size = min(self.input_resolution)
+        assert 0 <= self.shift_size < self.window_size, "shift_size must in 0-window_size"
+
+        self.norm1 = norm_layer(dim)
+        self.attn = WindowAttention(
+            dim, window_size=to_2tuple(self.window_size), num_heads=num_heads,
+            qkv_bias=qkv_bias, qk_scale=qk_scale, attn_drop=attn_drop, proj_drop=drop)
+
+        self.drop_path = DropPath(drop_path) if drop_path > 0. else nn.Identity()
+        self.norm2 = norm_layer(dim)
+        mlp_hidden_dim = int(dim * mlp_ratio)
+        self.mlp = Mlp(in_features=dim, hidden_features=mlp_hidden_dim, act_layer=act_layer, drop=drop)
+
+        if self.shift_size > 0:
+            attn_mask = self.calculate_mask(self.input_resolution)
+        else:
+            attn_mask = None
+
+        self.register_buffer("attn_mask", attn_mask)
+
+    def calculate_mask(self, x_size):
+        # calculate attention mask for SW-MSA
+        H, W = x_size
+        img_mask = torch.zeros((1, H, W, 1))  # 1 H W 1
+        h_slices = (slice(0, -self.window_size),
+                    slice(-self.window_size, -self.shift_size),
+                    slice(-self.shift_size, None))
+        w_slices = (slice(0, -self.window_size),
+                    slice(-self.window_size, -self.shift_size),
+                    slice(-self.shift_size, None))
+        cnt = 0
+        for h in h_slices:
+            for w in w_slices:
+                img_mask[:, h, w, :] = cnt
+                cnt += 1
+
+        mask_windows = window_partition(img_mask, self.window_size)  # nW, window_size, window_size, 1
+        mask_windows = mask_windows.view(-1, self.window_size * self.window_size)
+        attn_mask = mask_windows.unsqueeze(1) - mask_windows.unsqueeze(2)
+        attn_mask = attn_mask.masked_fill(attn_mask != 0, float(-100.0)).masked_fill(attn_mask == 0, float(0.0))
+
+        return attn_mask
+
+    def forward(self, x, x_size):
+        H, W = x_size
+        B, L, C = x.shape
+        # assert L == H * W, "input feature has wrong size"
+
+        shortcut = x
+        x = self.norm1(x)
+        x = x.view(B, H, W, C)
+
+        # cyclic shift
+        if self.shift_size > 0:
+            shifted_x = torch.roll(x, shifts=(-self.shift_size, -self.shift_size), dims=(1, 2))
+        else:
+            shifted_x = x
+
+        # partition windows
+        x_windows = window_partition(shifted_x, self.window_size)  # nW*B, window_size, window_size, C
+        x_windows = x_windows.view(-1, self.window_size * self.window_size, C)  # nW*B, window_size*window_size, C
+
+        # W-MSA/SW-MSA (to be compatible for testing on images whose shapes are the multiple of window size
+        if self.input_resolution == x_size:
+            attn_windows = self.attn(x_windows, mask=self.attn_mask)  # nW*B, window_size*window_size, C
+        else:
+            attn_windows = self.attn(x_windows, mask=self.calculate_mask(x_size).to(x.device))
+
+        # merge windows
+        attn_windows = attn_windows.view(-1, self.window_size, self.window_size, C)
+        shifted_x = window_reverse(attn_windows, self.window_size, H, W)  # B H' W' C
+
+        # reverse cyclic shift
+        if self.shift_size > 0:
+            x = torch.roll(shifted_x, shifts=(self.shift_size, self.shift_size), dims=(1, 2))
+        else:
+            x = shifted_x
+        x = x.view(B, H * W, C)
+
+        # FFN
+        x = shortcut + self.drop_path(x)
+        x = x + self.drop_path(self.mlp(self.norm2(x)))
+
+        return x
+
+    def extra_repr(self) -> str:
+        return f"dim={self.dim}, input_resolution={self.input_resolution}, num_heads={self.num_heads}, " \
+               f"window_size={self.window_size}, shift_size={self.shift_size}, mlp_ratio={self.mlp_ratio}"
+
+    def flops(self):
+        flops = 0
+        H, W = self.input_resolution
+        # norm1
+        flops += self.dim * H * W
+        # W-MSA/SW-MSA
+        nW = H * W / self.window_size / self.window_size
+        flops += nW * self.attn.flops(self.window_size * self.window_size)
+        # mlp
+        flops += 2 * H * W * self.dim * self.dim * self.mlp_ratio
+        # norm2
+        flops += self.dim * H * W
+        return flops
+
+
+class PatchMerging(nn.Module):
+    r""" Patch Merging Layer.
+
+    Args:
+        input_resolution (tuple[int]): Resolution of input feature.
+        dim (int): Number of input channels.
+        norm_layer (nn.Module, optional): Normalization layer.  Default: nn.LayerNorm
+    """
+
+    def __init__(self, input_resolution, dim, norm_layer=nn.LayerNorm):
+        super().__init__()
+        self.input_resolution = input_resolution
+        self.dim = dim
+        self.reduction = nn.Linear(4 * dim, 2 * dim, bias=False)
+        self.norm = norm_layer(4 * dim)
+
+    def forward(self, x):
+        """
+        x: B, H*W, C
+        """
+        H, W = self.input_resolution
+        B, L, C = x.shape
+        assert L == H * W, "input feature has wrong size"
+        assert H % 2 == 0 and W % 2 == 0, f"x size ({H}*{W}) are not even."
+
+        x = x.view(B, H, W, C)
+
+        x0 = x[:, 0::2, 0::2, :]  # B H/2 W/2 C
+        x1 = x[:, 1::2, 0::2, :]  # B H/2 W/2 C
+        x2 = x[:, 0::2, 1::2, :]  # B H/2 W/2 C
+        x3 = x[:, 1::2, 1::2, :]  # B H/2 W/2 C
+        x = torch.cat([x0, x1, x2, x3], -1)  # B H/2 W/2 4*C
+        x = x.view(B, -1, 4 * C)  # B H/2*W/2 4*C
+
+        x = self.norm(x)
+        x = self.reduction(x)
+
+        return x
+
+    def extra_repr(self) -> str:
+        return f"input_resolution={self.input_resolution}, dim={self.dim}"
+
+    def flops(self):
+        H, W = self.input_resolution
+        flops = H * W * self.dim
+        flops += (H // 2) * (W // 2) * 4 * self.dim * 2 * self.dim
+        return flops
+
+
+class BasicLayer(nn.Module):
+    """ A basic Swin Transformer layer for one stage.
+
+    Args:
+        dim (int): Number of input channels.
+        input_resolution (tuple[int]): Input resolution.
+        depth (int): Number of blocks.
+        num_heads (int): Number of attention heads.
+        window_size (int): Local window size.
+        mlp_ratio (float): Ratio of mlp hidden dim to embedding dim.
+        qkv_bias (bool, optional): If True, add a learnable bias to query, key, value. Default: True
+        qk_scale (float | None, optional): Override default qk scale of head_dim ** -0.5 if set.
+        drop (float, optional): Dropout rate. Default: 0.0
+        attn_drop (float, optional): Attention dropout rate. Default: 0.0
+        drop_path (float | tuple[float], optional): Stochastic depth rate. Default: 0.0
+        norm_layer (nn.Module, optional): Normalization layer. Default: nn.LayerNorm
+        downsample (nn.Module | None, optional): Downsample layer at the end of the layer. Default: None
+        use_checkpoint (bool): Whether to use checkpointing to save memory. Default: False.
+    """
+
+    def __init__(self, dim, input_resolution, depth, num_heads, window_size,
+                 mlp_ratio=4., qkv_bias=True, qk_scale=None, drop=0., attn_drop=0.,
+                 drop_path=0., norm_layer=nn.LayerNorm, downsample=None, use_checkpoint=False):
+
+        super().__init__()
+        self.dim = dim
+        self.input_resolution = input_resolution
+        self.depth = depth
+        self.use_checkpoint = use_checkpoint
+
+        # build blocks
+        self.blocks = nn.ModuleList([
+            SwinTransformerBlock(dim=dim, input_resolution=input_resolution,
+                                 num_heads=num_heads, window_size=window_size,
+                                 shift_size=0 if (i % 2 == 0) else window_size // 2,
+                                 mlp_ratio=mlp_ratio,
+                                 qkv_bias=qkv_bias, qk_scale=qk_scale,
+                                 drop=drop, attn_drop=attn_drop,
+                                 drop_path=drop_path[i] if isinstance(drop_path, list) else drop_path,
+                                 norm_layer=norm_layer)
+            for i in range(depth)])
+
+        # patch merging layer
+        if downsample is not None:
+            self.downsample = downsample(input_resolution, dim=dim, norm_layer=norm_layer)
+        else:
+            self.downsample = None
+
+    def forward(self, x, x_size):
+        for blk in self.blocks:
+            if self.use_checkpoint:
+                x = checkpoint.checkpoint(blk, x, x_size)
+            else:
+                x = blk(x, x_size)
+        if self.downsample is not None:
+            x = self.downsample(x)
+        return x
+
+    def extra_repr(self) -> str:
+        return f"dim={self.dim}, input_resolution={self.input_resolution}, depth={self.depth}"
+
+    def flops(self):
+        flops = 0
+        for blk in self.blocks:
+            flops += blk.flops()
+        if self.downsample is not None:
+            flops += self.downsample.flops()
+        return flops
+
+
+class RSTB(nn.Module):
+    """Residual Swin Transformer Block (RSTB).
+
+    Args:
+        dim (int): Number of input channels.
+        input_resolution (tuple[int]): Input resolution.
+        depth (int): Number of blocks.
+        num_heads (int): Number of attention heads.
+        window_size (int): Local window size.
+        mlp_ratio (float): Ratio of mlp hidden dim to embedding dim.
+        qkv_bias (bool, optional): If True, add a learnable bias to query, key, value. Default: True
+        qk_scale (float | None, optional): Override default qk scale of head_dim ** -0.5 if set.
+        drop (float, optional): Dropout rate. Default: 0.0
+        attn_drop (float, optional): Attention dropout rate. Default: 0.0
+        drop_path (float | tuple[float], optional): Stochastic depth rate. Default: 0.0
+        norm_layer (nn.Module, optional): Normalization layer. Default: nn.LayerNorm
+        downsample (nn.Module | None, optional): Downsample layer at the end of the layer. Default: None
+        use_checkpoint (bool): Whether to use checkpointing to save memory. Default: False.
+        img_size: Input image size.
+        patch_size: Patch size.
+        resi_connection: The convolutional block before residual connection.
+    """
+
+    def __init__(self, dim, input_resolution, depth, num_heads, window_size,
+                 mlp_ratio=4., qkv_bias=True, qk_scale=None, drop=0., attn_drop=0.,
+                 drop_path=0., norm_layer=nn.LayerNorm, downsample=None, use_checkpoint=False,
+                 img_size=224, patch_size=4, resi_connection='1conv'):
+        super(RSTB, self).__init__()
+
+        self.dim = dim
+        self.input_resolution = input_resolution
+
+        self.residual_group = BasicLayer(dim=dim,
+                                         input_resolution=input_resolution,
+                                         depth=depth,
+                                         num_heads=num_heads,
+                                         window_size=window_size,
+                                         mlp_ratio=mlp_ratio,
+                                         qkv_bias=qkv_bias, qk_scale=qk_scale,
+                                         drop=drop, attn_drop=attn_drop,
+                                         drop_path=drop_path,
+                                         norm_layer=norm_layer,
+                                         downsample=downsample,
+                                         use_checkpoint=use_checkpoint)
+
+        if resi_connection == '1conv':
+            self.conv = nn.Conv2d(dim, dim, 3, 1, 1)
+        elif resi_connection == '3conv':
+            # to save parameters and memory
+            self.conv = nn.Sequential(nn.Conv2d(dim, dim // 4, 3, 1, 1), nn.LeakyReLU(negative_slope=0.2, inplace=True),
+                                      nn.Conv2d(dim // 4, dim // 4, 1, 1, 0),
+                                      nn.LeakyReLU(negative_slope=0.2, inplace=True),
+                                      nn.Conv2d(dim // 4, dim, 3, 1, 1))
+
+        self.patch_embed = PatchEmbed(
+            img_size=img_size, patch_size=patch_size, in_chans=0, embed_dim=dim,
+            norm_layer=None)
+
+        self.patch_unembed = PatchUnEmbed(
+            img_size=img_size, patch_size=patch_size, in_chans=0, embed_dim=dim,
+            norm_layer=None)
+
+    def forward(self, x, x_size):
+        return self.patch_embed(self.conv(self.patch_unembed(self.residual_group(x, x_size), x_size))) + x
+
+    def flops(self):
+        flops = 0
+        flops += self.residual_group.flops()
+        H, W = self.input_resolution
+        flops += H * W * self.dim * self.dim * 9
+        flops += self.patch_embed.flops()
+        flops += self.patch_unembed.flops()
+
+        return flops
+
+
+class PatchEmbed(nn.Module):
+    r""" Image to Patch Embedding
+
+    Args:
+        img_size (int): Image size.  Default: 224.
+        patch_size (int): Patch token size. Default: 4.
+        in_chans (int): Number of input image channels. Default: 3.
+        embed_dim (int): Number of linear projection output channels. Default: 96.
+        norm_layer (nn.Module, optional): Normalization layer. Default: None
+    """
+
+    def __init__(self, img_size=224, patch_size=4, in_chans=3, embed_dim=96, norm_layer=None):
+        super().__init__()
+        img_size = to_2tuple(img_size)
+        patch_size = to_2tuple(patch_size)
+        patches_resolution = [img_size[0] // patch_size[0], img_size[1] // patch_size[1]]
+        self.img_size = img_size
+        self.patch_size = patch_size
+        self.patches_resolution = patches_resolution
+        self.num_patches = patches_resolution[0] * patches_resolution[1]
+
+        self.in_chans = in_chans
+        self.embed_dim = embed_dim
+
+        if norm_layer is not None:
+            self.norm = norm_layer(embed_dim)
+        else:
+            self.norm = None
+
+    def forward(self, x):
+        x = x.flatten(2).transpose(1, 2)  # B Ph*Pw C
+        if self.norm is not None:
+            x = self.norm(x)
+        return x
+
+    def flops(self):
+        flops = 0
+        H, W = self.img_size
+        if self.norm is not None:
+            flops += H * W * self.embed_dim
+        return flops
+
+
+class PatchUnEmbed(nn.Module):
+    r""" Image to Patch Unembedding
+
+    Args:
+        img_size (int): Image size.  Default: 224.
+        patch_size (int): Patch token size. Default: 4.
+        in_chans (int): Number of input image channels. Default: 3.
+        embed_dim (int): Number of linear projection output channels. Default: 96.
+        norm_layer (nn.Module, optional): Normalization layer. Default: None
+    """
+
+    def __init__(self, img_size=224, patch_size=4, in_chans=3, embed_dim=96, norm_layer=None):
+        super().__init__()
+        img_size = to_2tuple(img_size)
+        patch_size = to_2tuple(patch_size)
+        patches_resolution = [img_size[0] // patch_size[0], img_size[1] // patch_size[1]]
+        self.img_size = img_size
+        self.patch_size = patch_size
+        self.patches_resolution = patches_resolution
+        self.num_patches = patches_resolution[0] * patches_resolution[1]
+
+        self.in_chans = in_chans
+        self.embed_dim = embed_dim
+
+    def forward(self, x, x_size):
+        B, HW, C = x.shape
+        x = x.transpose(1, 2).view(B, self.embed_dim, x_size[0], x_size[1])  # B Ph*Pw C
+        return x
+
+    def flops(self):
+        flops = 0
+        return flops
+
+
+class Upsample(nn.Sequential):
+    """Upsample module.
+
+    Args:
+        scale (int): Scale factor. Supported scales: 2^n and 3.
+        num_feat (int): Channel number of intermediate features.
+    """
+
+    def __init__(self, scale, num_feat):
+        m = []
+        if (scale & (scale - 1)) == 0:  # scale = 2^n
+            for _ in range(int(math.log(scale, 2))):
+                m.append(nn.Conv2d(num_feat, 4 * num_feat, 3, 1, 1))
+                m.append(nn.PixelShuffle(2))
+        elif scale == 3:
+            m.append(nn.Conv2d(num_feat, 9 * num_feat, 3, 1, 1))
+            m.append(nn.PixelShuffle(3))
+        else:
+            raise ValueError(f'scale {scale} is not supported. ' 'Supported scales: 2^n and 3.')
+        super(Upsample, self).__init__(*m)
+
+
+class UpsampleOneStep(nn.Sequential):
+    """UpsampleOneStep module (the difference with Upsample is that it always only has 1conv + 1pixelshuffle)
+       Used in lightweight SR to save parameters.
+
+    Args:
+        scale (int): Scale factor. Supported scales: 2^n and 3.
+        num_feat (int): Channel number of intermediate features.
+
+    """
+
+    def __init__(self, scale, num_feat, num_out_ch, input_resolution=None):
+        self.num_feat = num_feat
+        self.input_resolution = input_resolution
+        m = []
+        m.append(nn.Conv2d(num_feat, (scale ** 2) * num_out_ch, 3, 1, 1))
+        m.append(nn.PixelShuffle(scale))
+        super(UpsampleOneStep, self).__init__(*m)
+
+    def flops(self):
+        H, W = self.input_resolution
+        flops = H * W * self.num_feat * 3 * 9
+        return flops
+
+
+class SwinIR(nn.Module):
+    r""" SwinIR
+        A PyTorch impl of : `SwinIR: Image Restoration Using Swin Transformer`, based on Swin Transformer.
+
+    Args:
+        img_size (int | tuple(int)): Input image size. Default 64
+        patch_size (int | tuple(int)): Patch size. Default: 1
+        in_chans (int): Number of input image channels. Default: 3
+        embed_dim (int): Patch embedding dimension. Default: 96
+        depths (tuple(int)): Depth of each Swin Transformer layer.
+        num_heads (tuple(int)): Number of attention heads in different layers.
+        window_size (int): Window size. Default: 7
+        mlp_ratio (float): Ratio of mlp hidden dim to embedding dim. Default: 4
+        qkv_bias (bool): If True, add a learnable bias to query, key, value. Default: True
+        qk_scale (float): Override default qk scale of head_dim ** -0.5 if set. Default: None
+        drop_rate (float): Dropout rate. Default: 0
+        attn_drop_rate (float): Attention dropout rate. Default: 0
+        drop_path_rate (float): Stochastic depth rate. Default: 0.1
+        norm_layer (nn.Module): Normalization layer. Default: nn.LayerNorm.
+        ape (bool): If True, add absolute position embedding to the patch embedding. Default: False
+        patch_norm (bool): If True, add normalization after patch embedding. Default: True
+        use_checkpoint (bool): Whether to use checkpointing to save memory. Default: False
+        sf: Upscale factor. 2/3/4/8 for image SR, 1 for denoising and compress artifact reduction
+        img_range: Image range. 1. or 255.
+        upsampler: The reconstruction reconstruction module. 'pixelshuffle'/'pixelshuffledirect'/'nearest+conv'/None
+        resi_connection: The convolutional block before residual connection. '1conv'/'3conv'
+    """
+
+    def __init__(
+        self,
+        img_size=64,
+        patch_size=1,
+        in_chans=3,
+        embed_dim=96,
+        depths=[6, 6, 6, 6],
+        num_heads=[6, 6, 6, 6],
+        window_size=7,
+        mlp_ratio=4.,
+        qkv_bias=True,
+        qk_scale=None,
+        drop_rate=0.,
+        attn_drop_rate=0.,
+        drop_path_rate=0.1,
+        norm_layer=nn.LayerNorm,
+        ape=False,
+        patch_norm=True,
+        use_checkpoint=False,
+        sf=4,
+        img_range=1.,
+        upsampler='',
+        resi_connection='1conv',
+        unshuffle=False,
+        unshuffle_scale=None,
+        hq_key: str="jpg",
+        lq_key: str="hint",
+        learning_rate: float=None,
+        weight_decay: float=None
+    ) -> "SwinIR":
+        super(SwinIR, self).__init__()
+        num_in_ch = in_chans * (unshuffle_scale**2) if unshuffle else in_chans
+        num_out_ch = in_chans
+        num_feat = 64
+        self.img_range = img_range
+        if in_chans == 3:
+            rgb_mean = (0.4488, 0.4371, 0.4040)
+            self.mean = torch.Tensor(rgb_mean).view(1, 3, 1, 1)
+        else:
+            self.mean = torch.zeros(1, 1, 1, 1)
+        self.upscale = sf
+        self.upsampler = upsampler
+        self.window_size = window_size
+        self.unshuffle_scale = unshuffle_scale
+        self.unshuffle = unshuffle
+
+        #####################################################################################################
+        ################################### 1, shallow feature extraction ###################################
+        if unshuffle:
+            assert unshuffle_scale is not None
+            self.conv_first = nn.Sequential(
+                nn.PixelUnshuffle(sf),
+                nn.Conv2d(num_in_ch, embed_dim, 3, 1, 1),
+            )
+        else:
+            self.conv_first = nn.Conv2d(num_in_ch, embed_dim, 3, 1, 1)
+
+        #####################################################################################################
+        ################################### 2, deep feature extraction ######################################
+        self.num_layers = len(depths)
+        self.embed_dim = embed_dim
+        self.ape = ape
+        self.patch_norm = patch_norm
+        self.num_features = embed_dim
+        self.mlp_ratio = mlp_ratio
+
+        # split image into non-overlapping patches
+        self.patch_embed = PatchEmbed(
+            img_size=img_size, patch_size=patch_size, in_chans=embed_dim, embed_dim=embed_dim,
+            norm_layer=norm_layer if self.patch_norm else None
+        )
+        num_patches = self.patch_embed.num_patches
+        patches_resolution = self.patch_embed.patches_resolution
+        self.patches_resolution = patches_resolution
+
+        # merge non-overlapping patches into image
+        self.patch_unembed = PatchUnEmbed(
+            img_size=img_size, patch_size=patch_size, in_chans=embed_dim, embed_dim=embed_dim,
+            norm_layer=norm_layer if self.patch_norm else None
+        )
+
+        # absolute position embedding
+        if self.ape:
+            self.absolute_pos_embed = nn.Parameter(torch.zeros(1, num_patches, embed_dim))
+            trunc_normal_(self.absolute_pos_embed, std=.02)
+
+        self.pos_drop = nn.Dropout(p=drop_rate)
+
+        # stochastic depth
+        dpr = [x.item() for x in torch.linspace(0, drop_path_rate, sum(depths))]  # stochastic depth decay rule
+
+        # build Residual Swin Transformer blocks (RSTB)
+        self.layers = nn.ModuleList()
+        for i_layer in range(self.num_layers):
+            layer = RSTB(
+                dim=embed_dim,
+                input_resolution=(patches_resolution[0], patches_resolution[1]),
+                depth=depths[i_layer],
+                num_heads=num_heads[i_layer],
+                window_size=window_size,
+                mlp_ratio=self.mlp_ratio,
+                qkv_bias=qkv_bias, qk_scale=qk_scale,
+                drop=drop_rate, attn_drop=attn_drop_rate,
+                drop_path=dpr[sum(depths[:i_layer]):sum(depths[:i_layer + 1])],  # no impact on SR results
+                norm_layer=norm_layer,
+                downsample=None,
+                use_checkpoint=use_checkpoint,
+                img_size=img_size,
+                patch_size=patch_size,
+                resi_connection=resi_connection
+            )
+            self.layers.append(layer)
+        self.norm = norm_layer(self.num_features)
+
+        # build the last conv layer in deep feature extraction
+        if resi_connection == '1conv':
+            self.conv_after_body = nn.Conv2d(embed_dim, embed_dim, 3, 1, 1)
+        elif resi_connection == '3conv':
+            # to save parameters and memory
+            self.conv_after_body = nn.Sequential(
+                nn.Conv2d(embed_dim, embed_dim // 4, 3, 1, 1),
+                nn.LeakyReLU(negative_slope=0.2, inplace=True),
+                nn.Conv2d(embed_dim // 4, embed_dim // 4, 1, 1, 0),
+                nn.LeakyReLU(negative_slope=0.2, inplace=True),
+                nn.Conv2d(embed_dim // 4, embed_dim, 3, 1, 1)
+            )
+
+        #####################################################################################################
+        ################################ 3, high quality image reconstruction ################################
+        if self.upsampler == 'pixelshuffle':
+            # for classical SR
+            self.conv_before_upsample = nn.Sequential(
+                nn.Conv2d(embed_dim, num_feat, 3, 1, 1),
+                nn.LeakyReLU(inplace=True)
+            )
+            self.upsample = Upsample(sf, num_feat)
+            self.conv_last = nn.Conv2d(num_feat, num_out_ch, 3, 1, 1)
+        elif self.upsampler == 'pixelshuffledirect':
+            # for lightweight SR (to save parameters)
+            self.upsample = UpsampleOneStep(
+                sf, embed_dim, num_out_ch,
+                (patches_resolution[0], patches_resolution[1])
+            )
+        elif self.upsampler == 'nearest+conv':
+            # for real-world SR (less artifacts)
+            self.conv_before_upsample = nn.Sequential(
+                nn.Conv2d(embed_dim, num_feat, 3, 1, 1),
+                nn.LeakyReLU(inplace=True)
+            )
+            self.conv_up1 = nn.Conv2d(num_feat, num_feat, 3, 1, 1)
+            if self.upscale == 4:
+                self.conv_up2 = nn.Conv2d(num_feat, num_feat, 3, 1, 1)
+            elif self.upscale == 8:
+                self.conv_up2 = nn.Conv2d(num_feat, num_feat, 3, 1, 1)
+                self.conv_up3 = nn.Conv2d(num_feat, num_feat, 3, 1, 1)
+            self.conv_hr = nn.Conv2d(num_feat, num_feat, 3, 1, 1)
+            self.conv_last = nn.Conv2d(num_feat, num_out_ch, 3, 1, 1)
+            self.lrelu = nn.LeakyReLU(negative_slope=0.2, inplace=True)
+        else:
+            # for image denoising and JPEG compression artifact reduction
+            self.conv_last = nn.Conv2d(embed_dim, num_out_ch, 3, 1, 1)
+
+        self.apply(self._init_weights)
+
+    def _init_weights(self, m: nn.Module) -> None:
+        if isinstance(m, nn.Linear):
+            trunc_normal_(m.weight, std=.02)
+            if isinstance(m, nn.Linear) and m.bias is not None:
+                nn.init.constant_(m.bias, 0)
+        elif isinstance(m, nn.LayerNorm):
+            nn.init.constant_(m.bias, 0)
+            nn.init.constant_(m.weight, 1.0)
+
+    # TODO: What's this ?
+    @torch.jit.ignore
+    def no_weight_decay(self) -> Set[str]:
+        return {'absolute_pos_embed'}
+
+    @torch.jit.ignore
+    def no_weight_decay_keywords(self) -> Set[str]:
+        return {'relative_position_bias_table'}
+
+    def check_image_size(self, x: torch.Tensor) -> torch.Tensor:
+        _, _, h, w = x.size()
+        mod_pad_h = (self.window_size - h % self.window_size) % self.window_size
+        mod_pad_w = (self.window_size - w % self.window_size) % self.window_size
+        x = F.pad(x, (0, mod_pad_w, 0, mod_pad_h), 'reflect')
+        return x
+
+    def forward_features(self, x: torch.Tensor) -> torch.Tensor:
+        x_size = (x.shape[2], x.shape[3])
+        x = self.patch_embed(x)
+        if self.ape:
+            x = x + self.absolute_pos_embed
+        x = self.pos_drop(x)
+
+        for layer in self.layers:
+            x = layer(x, x_size)
+
+        x = self.norm(x)  # B L C
+        x = self.patch_unembed(x, x_size)
+
+        return x
+
+    def forward(self, x: torch.Tensor) -> torch.Tensor:
+        H, W = x.shape[2:]
+        x = self.check_image_size(x)
+
+        self.mean = self.mean.type_as(x)
+        x = (x - self.mean) * self.img_range
+
+        if self.upsampler == 'pixelshuffle':
+            # for classical SR
+            x = self.conv_first(x)
+            x = self.conv_after_body(self.forward_features(x)) + x
+            x = self.conv_before_upsample(x)
+            x = self.conv_last(self.upsample(x))
+        elif self.upsampler == 'pixelshuffledirect':
+            # for lightweight SR
+            x = self.conv_first(x)
+            x = self.conv_after_body(self.forward_features(x)) + x
+            x = self.upsample(x)
+        elif self.upsampler == 'nearest+conv':
+            # for real-world SR
+            x = self.conv_first(x)
+            x = self.conv_after_body(self.forward_features(x)) + x
+            x = self.conv_before_upsample(x)
+            x = self.lrelu(self.conv_up1(torch.nn.functional.interpolate(x, scale_factor=2, mode='nearest')))
+            if self.upscale == 4:
+                x = self.lrelu(self.conv_up2(torch.nn.functional.interpolate(x, scale_factor=2, mode='nearest')))
+            elif self.upscale == 8:
+                x = self.lrelu(self.conv_up2(torch.nn.functional.interpolate(x, scale_factor=2, mode='nearest')))
+                x = self.lrelu(self.conv_up3(torch.nn.functional.interpolate(x, scale_factor=2, mode='nearest')))
+            x = self.conv_last(self.lrelu(self.conv_hr(x)))
+        else:
+            # for image denoising and JPEG compression artifact reduction
+            x_first = self.conv_first(x)
+            res = self.conv_after_body(self.forward_features(x_first)) + x_first
+            x = x + self.conv_last(res)
+
+        x = x / self.img_range + self.mean
+
+        return x[:, :, :H*self.upscale, :W*self.upscale]
+
+    def flops(self) -> int:
+        flops = 0
+        H, W = self.patches_resolution
+        flops += H * W * 3 * self.embed_dim * 9
+        flops += self.patch_embed.flops()
+        for i, layer in enumerate(self.layers):
+            flops += layer.flops()
+        flops += H * W * 3 * self.embed_dim * self.embed_dim
+        flops += self.upsample.flops()
+        return flops

model/unet.py

@@ -0,0 +1,719 @@
+from abc import abstractmethod
+import math
+
+import numpy as np
+import torch as th
+import torch.nn as nn
+import torch.nn.functional as F
+
+from model.util import (
+    checkpoint,
+    conv_nd,
+    linear,
+    avg_pool_nd,
+    zero_module,
+    normalization,
+    timestep_embedding,
+    exists
+)
+from model.attention import SpatialTransformer
+
+
+class TimestepBlock(nn.Module):
+    """
+    Any module where forward() takes timestep embeddings as a second argument.
+    """
+
+    @abstractmethod
+    def forward(self, x, emb):
+        """
+        Apply the module to `x` given `emb` timestep embeddings.
+        """
+
+
+class TimestepEmbedSequential(nn.Sequential, TimestepBlock):
+    """
+    A sequential module that passes timestep embeddings to the children that
+    support it as an extra input.
+    """
+
+    def forward(self, x, emb, context=None):
+        for layer in self:
+            if isinstance(layer, TimestepBlock):
+                x = layer(x, emb)
+            elif isinstance(layer, SpatialTransformer):
+                x = layer(x, context)
+            else:
+                x = layer(x)
+        return x
+
+
+class Upsample(nn.Module):
+    """
+    An upsampling layer with an optional convolution.
+    :param channels: channels in the inputs and outputs.
+    :param use_conv: a bool determining if a convolution is applied.
+    :param dims: determines if the signal is 1D, 2D, or 3D. If 3D, then
+                 upsampling occurs in the inner-two dimensions.
+    """
+
+    def __init__(self, channels, use_conv, dims=2, out_channels=None, padding=1):
+        super().__init__()
+        self.channels = channels
+        self.out_channels = out_channels or channels
+        self.use_conv = use_conv
+        self.dims = dims
+        if use_conv:
+            self.conv = conv_nd(dims, self.channels, self.out_channels, 3, padding=padding)
+
+    def forward(self, x):
+        assert x.shape[1] == self.channels
+        if self.dims == 3:
+            x = F.interpolate(
+                x, (x.shape[2], x.shape[3] * 2, x.shape[4] * 2), mode="nearest"
+            )
+        else:
+            x = F.interpolate(x, scale_factor=2, mode="nearest")
+        if self.use_conv:
+            x = self.conv(x)
+        return x
+
+
+class Downsample(nn.Module):
+    """
+    A downsampling layer with an optional convolution.
+    :param channels: channels in the inputs and outputs.
+    :param use_conv: a bool determining if a convolution is applied.
+    :param dims: determines if the signal is 1D, 2D, or 3D. If 3D, then
+                 downsampling occurs in the inner-two dimensions.
+    """
+
+    def __init__(self, channels, use_conv, dims=2, out_channels=None,padding=1):
+        super().__init__()
+        self.channels = channels
+        self.out_channels = out_channels or channels
+        self.use_conv = use_conv
+        self.dims = dims
+        stride = 2 if dims != 3 else (1, 2, 2)
+        if use_conv:
+            self.op = conv_nd(
+                dims, self.channels, self.out_channels, 3, stride=stride, padding=padding
+            )
+        else:
+            assert self.channels == self.out_channels
+            self.op = avg_pool_nd(dims, kernel_size=stride, stride=stride)
+
+    def forward(self, x):
+        assert x.shape[1] == self.channels
+        return self.op(x)
+
+
+class ResBlock(TimestepBlock):
+    """
+    A residual block that can optionally change the number of channels.
+    :param channels: the number of input channels.
+    :param emb_channels: the number of timestep embedding channels.
+    :param dropout: the rate of dropout.
+    :param out_channels: if specified, the number of out channels.
+    :param use_conv: if True and out_channels is specified, use a spatial
+        convolution instead of a smaller 1x1 convolution to change the
+        channels in the skip connection.
+    :param dims: determines if the signal is 1D, 2D, or 3D.
+    :param use_checkpoint: if True, use gradient checkpointing on this module.
+    :param up: if True, use this block for upsampling.
+    :param down: if True, use this block for downsampling.
+    """
+
+    def __init__(
+        self,
+        channels,
+        emb_channels,
+        dropout,
+        out_channels=None,
+        use_conv=False,
+        use_scale_shift_norm=False,
+        dims=2,
+        use_checkpoint=False,
+        up=False,
+        down=False,
+    ):
+        super().__init__()
+        self.channels = channels
+        self.emb_channels = emb_channels
+        self.dropout = dropout
+        self.out_channels = out_channels or channels
+        self.use_conv = use_conv
+        self.use_checkpoint = use_checkpoint
+        self.use_scale_shift_norm = use_scale_shift_norm
+
+        self.in_layers = nn.Sequential(
+            normalization(channels),
+            nn.SiLU(),
+            conv_nd(dims, channels, self.out_channels, 3, padding=1),
+        )
+
+        self.updown = up or down
+
+        if up:
+            self.h_upd = Upsample(channels, False, dims)
+            self.x_upd = Upsample(channels, False, dims)
+        elif down:
+            self.h_upd = Downsample(channels, False, dims)
+            self.x_upd = Downsample(channels, False, dims)
+        else:
+            self.h_upd = self.x_upd = nn.Identity()
+
+        self.emb_layers = nn.Sequential(
+            nn.SiLU(),
+            linear(
+                emb_channels,
+                2 * self.out_channels if use_scale_shift_norm else self.out_channels,
+            ),
+        )
+        self.out_layers = nn.Sequential(
+            normalization(self.out_channels),
+            nn.SiLU(),
+            nn.Dropout(p=dropout),
+            zero_module(
+                conv_nd(dims, self.out_channels, self.out_channels, 3, padding=1)
+            ),
+        )
+
+        if self.out_channels == channels:
+            self.skip_connection = nn.Identity()
+        elif use_conv:
+            self.skip_connection = conv_nd(
+                dims, channels, self.out_channels, 3, padding=1
+            )
+        else:
+            self.skip_connection = conv_nd(dims, channels, self.out_channels, 1)
+
+    def forward(self, x, emb):
+        """
+        Apply the block to a Tensor, conditioned on a timestep embedding.
+        :param x: an [N x C x ...] Tensor of features.
+        :param emb: an [N x emb_channels] Tensor of timestep embeddings.
+        :return: an [N x C x ...] Tensor of outputs.
+        """
+        return checkpoint(
+            self._forward, (x, emb), self.parameters(), self.use_checkpoint
+        )
+
+
+    def _forward(self, x, emb):
+        if self.updown:
+            in_rest, in_conv = self.in_layers[:-1], self.in_layers[-1]
+            h = in_rest(x)
+            h = self.h_upd(h)
+            x = self.x_upd(x)
+            h = in_conv(h)
+        else:
+            h = self.in_layers(x)
+        emb_out = self.emb_layers(emb).type(h.dtype)
+        while len(emb_out.shape) < len(h.shape):
+            emb_out = emb_out[..., None]
+        if self.use_scale_shift_norm:
+            out_norm, out_rest = self.out_layers[0], self.out_layers[1:]
+            scale, shift = th.chunk(emb_out, 2, dim=1)
+            h = out_norm(h) * (1 + scale) + shift
+            h = out_rest(h)
+        else:
+            h = h + emb_out
+            h = self.out_layers(h)
+        return self.skip_connection(x) + h
+
+
+class AttentionBlock(nn.Module):
+    """
+    An attention block that allows spatial positions to attend to each other.
+    Originally ported from here, but adapted to the N-d case.
+    https://github.com/hojonathanho/diffusion/blob/1e0dceb3b3495bbe19116a5e1b3596cd0706c543/diffusion_tf/models/unet.py#L66.
+    """
+
+    def __init__(
+        self,
+        channels,
+        num_heads=1,
+        num_head_channels=-1,
+        use_checkpoint=False,
+        use_new_attention_order=False,
+    ):
+        super().__init__()
+        self.channels = channels
+        if num_head_channels == -1:
+            self.num_heads = num_heads
+        else:
+            assert (
+                channels % num_head_channels == 0
+            ), f"q,k,v channels {channels} is not divisible by num_head_channels {num_head_channels}"
+            self.num_heads = channels // num_head_channels
+        self.use_checkpoint = use_checkpoint
+        self.norm = normalization(channels)
+        self.qkv = conv_nd(1, channels, channels * 3, 1)
+        if use_new_attention_order:
+            # split qkv before split heads
+            self.attention = QKVAttention(self.num_heads)
+        else:
+            # split heads before split qkv
+            self.attention = QKVAttentionLegacy(self.num_heads)
+
+        self.proj_out = zero_module(conv_nd(1, channels, channels, 1))
+
+    def forward(self, x):
+        return checkpoint(self._forward, (x,), self.parameters(), True)   # TODO: check checkpoint usage, is True # TODO: fix the .half call!!!
+        #return pt_checkpoint(self._forward, x)  # pytorch
+
+    def _forward(self, x):
+        b, c, *spatial = x.shape
+        x = x.reshape(b, c, -1)
+        qkv = self.qkv(self.norm(x))
+        h = self.attention(qkv)
+        h = self.proj_out(h)
+        return (x + h).reshape(b, c, *spatial)
+
+
+def count_flops_attn(model, _x, y):
+    """
+    A counter for the `thop` package to count the operations in an
+    attention operation.
+    Meant to be used like:
+        macs, params = thop.profile(
+            model,
+            inputs=(inputs, timestamps),
+            custom_ops={QKVAttention: QKVAttention.count_flops},
+        )
+    """
+    b, c, *spatial = y[0].shape
+    num_spatial = int(np.prod(spatial))
+    # We perform two matmuls with the same number of ops.
+    # The first computes the weight matrix, the second computes
+    # the combination of the value vectors.
+    matmul_ops = 2 * b * (num_spatial ** 2) * c
+    model.total_ops += th.DoubleTensor([matmul_ops])
+
+
+class QKVAttentionLegacy(nn.Module):
+    """
+    A module which performs QKV attention. Matches legacy QKVAttention + input/ouput heads shaping
+    """
+
+    def __init__(self, n_heads):
+        super().__init__()
+        self.n_heads = n_heads
+
+    def forward(self, qkv):
+        """
+        Apply QKV attention.
+        :param qkv: an [N x (H * 3 * C) x T] tensor of Qs, Ks, and Vs.
+        :return: an [N x (H * C) x T] tensor after attention.
+        """
+        bs, width, length = qkv.shape
+        assert width % (3 * self.n_heads) == 0
+        ch = width // (3 * self.n_heads)
+        q, k, v = qkv.reshape(bs * self.n_heads, ch * 3, length).split(ch, dim=1)
+        scale = 1 / math.sqrt(math.sqrt(ch))
+        weight = th.einsum(
+            "bct,bcs->bts", q * scale, k * scale
+        )  # More stable with f16 than dividing afterwards
+        weight = th.softmax(weight.float(), dim=-1).type(weight.dtype)
+        a = th.einsum("bts,bcs->bct", weight, v)
+        return a.reshape(bs, -1, length)
+
+    @staticmethod
+    def count_flops(model, _x, y):
+        return count_flops_attn(model, _x, y)
+
+
+class QKVAttention(nn.Module):
+    """
+    A module which performs QKV attention and splits in a different order.
+    """
+
+    def __init__(self, n_heads):
+        super().__init__()
+        self.n_heads = n_heads
+
+    def forward(self, qkv):
+        """
+        Apply QKV attention.
+        :param qkv: an [N x (3 * H * C) x T] tensor of Qs, Ks, and Vs.
+        :return: an [N x (H * C) x T] tensor after attention.
+        """
+        bs, width, length = qkv.shape
+        assert width % (3 * self.n_heads) == 0
+        ch = width // (3 * self.n_heads)
+        q, k, v = qkv.chunk(3, dim=1)
+        scale = 1 / math.sqrt(math.sqrt(ch))
+        weight = th.einsum(
+            "bct,bcs->bts",
+            (q * scale).view(bs * self.n_heads, ch, length),
+            (k * scale).view(bs * self.n_heads, ch, length),
+        )  # More stable with f16 than dividing afterwards
+        weight = th.softmax(weight.float(), dim=-1).type(weight.dtype)
+        a = th.einsum("bts,bcs->bct", weight, v.reshape(bs * self.n_heads, ch, length))
+        return a.reshape(bs, -1, length)
+
+    @staticmethod
+    def count_flops(model, _x, y):
+        return count_flops_attn(model, _x, y)
+
+
+class UNetModel(nn.Module):
+    """
+    The full UNet model with attention and timestep embedding.
+    :param in_channels: channels in the input Tensor.
+    :param model_channels: base channel count for the model.
+    :param out_channels: channels in the output Tensor.
+    :param num_res_blocks: number of residual blocks per downsample.
+    :param attention_resolutions: a collection of downsample rates at which
+        attention will take place. May be a set, list, or tuple.
+        For example, if this contains 4, then at 4x downsampling, attention
+        will be used.
+    :param dropout: the dropout probability.
+    :param channel_mult: channel multiplier for each level of the UNet.
+    :param conv_resample: if True, use learned convolutions for upsampling and
+        downsampling.
+    :param dims: determines if the signal is 1D, 2D, or 3D.
+    :param num_classes: if specified (as an int), then this model will be
+        class-conditional with `num_classes` classes.
+    :param use_checkpoint: use gradient checkpointing to reduce memory usage.
+    :param num_heads: the number of attention heads in each attention layer.
+    :param num_heads_channels: if specified, ignore num_heads and instead use
+                               a fixed channel width per attention head.
+    :param num_heads_upsample: works with num_heads to set a different number
+                               of heads for upsampling. Deprecated.
+    :param use_scale_shift_norm: use a FiLM-like conditioning mechanism.
+    :param resblock_updown: use residual blocks for up/downsampling.
+    :param use_new_attention_order: use a different attention pattern for potentially
+                                    increased efficiency.
+    """
+
+    def __init__(
+        self,
+        image_size,
+        in_channels,
+        model_channels,
+        out_channels,
+        num_res_blocks,
+        attention_resolutions,
+        dropout=0,
+        channel_mult=(1, 2, 4, 8),
+        conv_resample=True,
+        dims=2,
+        num_classes=None,
+        use_checkpoint=False,
+        use_fp16=False,
+        num_heads=-1,
+        num_head_channels=-1,
+        num_heads_upsample=-1,
+        use_scale_shift_norm=False,
+        resblock_updown=False,
+        use_new_attention_order=False,
+        use_spatial_transformer=False,    # custom transformer support
+        transformer_depth=1,              # custom transformer support
+        context_dim=None,                 # custom transformer support
+        n_embed=None,                     # custom support for prediction of discrete ids into codebook of first stage vq model
+        legacy=True,
+        disable_self_attentions=None,
+        num_attention_blocks=None,
+        disable_middle_self_attn=False,
+        use_linear_in_transformer=False,
+    ):
+        super().__init__()
+        if use_spatial_transformer:
+            assert context_dim is not None, 'Fool!! You forgot to include the dimension of your cross-attention conditioning...'
+
+        if context_dim is not None:
+            assert use_spatial_transformer, 'Fool!! You forgot to use the spatial transformer for your cross-attention conditioning...'
+            from omegaconf.listconfig import ListConfig
+            if type(context_dim) == ListConfig:
+                context_dim = list(context_dim)
+
+        if num_heads_upsample == -1:
+            num_heads_upsample = num_heads
+
+        if num_heads == -1:
+            assert num_head_channels != -1, 'Either num_heads or num_head_channels has to be set'
+
+        if num_head_channels == -1:
+            assert num_heads != -1, 'Either num_heads or num_head_channels has to be set'
+
+        self.image_size = image_size
+        self.in_channels = in_channels
+        self.model_channels = model_channels
+        self.out_channels = out_channels
+        if isinstance(num_res_blocks, int):
+            self.num_res_blocks = len(channel_mult) * [num_res_blocks]
+        else:
+            if len(num_res_blocks) != len(channel_mult):
+                raise ValueError("provide num_res_blocks either as an int (globally constant) or "
+                                 "as a list/tuple (per-level) with the same length as channel_mult")
+            self.num_res_blocks = num_res_blocks
+        if disable_self_attentions is not None:
+            # should be a list of booleans, indicating whether to disable self-attention in TransformerBlocks or not
+            assert len(disable_self_attentions) == len(channel_mult)
+        if num_attention_blocks is not None:
+            assert len(num_attention_blocks) == len(self.num_res_blocks)
+            assert all(map(lambda i: self.num_res_blocks[i] >= num_attention_blocks[i], range(len(num_attention_blocks))))
+            print(f"Constructor of UNetModel received num_attention_blocks={num_attention_blocks}. "
+                  f"This option has LESS priority than attention_resolutions {attention_resolutions}, "
+                  f"i.e., in cases where num_attention_blocks[i] > 0 but 2**i not in attention_resolutions, "
+                  f"attention will still not be set.")
+
+        self.attention_resolutions = attention_resolutions
+        self.dropout = dropout
+        self.channel_mult = channel_mult
+        self.conv_resample = conv_resample
+        self.num_classes = num_classes
+        self.use_checkpoint = use_checkpoint
+        self.dtype = th.float16 if use_fp16 else th.float32
+        self.num_heads = num_heads
+        self.num_head_channels = num_head_channels
+        self.num_heads_upsample = num_heads_upsample
+        self.predict_codebook_ids = n_embed is not None
+
+        time_embed_dim = model_channels * 4
+        self.time_embed = nn.Sequential(
+            linear(model_channels, time_embed_dim),
+            nn.SiLU(),
+            linear(time_embed_dim, time_embed_dim),
+        )
+
+        if self.num_classes is not None:
+            if isinstance(self.num_classes, int):
+                self.label_emb = nn.Embedding(num_classes, time_embed_dim)
+            elif self.num_classes == "continuous":
+                print("setting up linear c_adm embedding layer")
+                self.label_emb = nn.Linear(1, time_embed_dim)
+            else:
+                raise ValueError()
+
+        self.input_blocks = nn.ModuleList(
+            [
+                TimestepEmbedSequential(
+                    conv_nd(dims, in_channels, model_channels, 3, padding=1)
+                )
+            ]
+        )
+        self._feature_size = model_channels
+        input_block_chans = [model_channels]
+        ch = model_channels
+        ds = 1
+        for level, mult in enumerate(channel_mult):
+            for nr in range(self.num_res_blocks[level]):
+                layers = [
+                    ResBlock(
+                        ch,
+                        time_embed_dim,
+                        dropout,
+                        out_channels=mult * model_channels,
+                        dims=dims,
+                        use_checkpoint=use_checkpoint,
+                        use_scale_shift_norm=use_scale_shift_norm,
+                    )
+                ]
+                ch = mult * model_channels
+                if ds in attention_resolutions:
+                    if num_head_channels == -1:
+                        dim_head = ch // num_heads
+                    else:
+                        num_heads = ch // num_head_channels
+                        dim_head = num_head_channels
+                    if legacy:
+                        #num_heads = 1
+                        dim_head = ch // num_heads if use_spatial_transformer else num_head_channels
+                    if exists(disable_self_attentions):
+                        disabled_sa = disable_self_attentions[level]
+                    else:
+                        disabled_sa = False
+
+                    if not exists(num_attention_blocks) or nr < num_attention_blocks[level]:
+                        layers.append(
+                            AttentionBlock(
+                                ch,
+                                use_checkpoint=use_checkpoint,
+                                num_heads=num_heads,
+                                num_head_channels=dim_head,
+                                use_new_attention_order=use_new_attention_order,
+                            ) if not use_spatial_transformer else SpatialTransformer(
+                                ch, num_heads, dim_head, depth=transformer_depth, context_dim=context_dim,
+                                disable_self_attn=disabled_sa, use_linear=use_linear_in_transformer,
+                                use_checkpoint=use_checkpoint
+                            )
+                        )
+                self.input_blocks.append(TimestepEmbedSequential(*layers))
+                self._feature_size += ch
+                input_block_chans.append(ch)
+            if level != len(channel_mult) - 1:
+                out_ch = ch
+                self.input_blocks.append(
+                    TimestepEmbedSequential(
+                        ResBlock(
+                            ch,
+                            time_embed_dim,
+                            dropout,
+                            out_channels=out_ch,
+                            dims=dims,
+                            use_checkpoint=use_checkpoint,
+                            use_scale_shift_norm=use_scale_shift_norm,
+                            down=True,
+                        )
+                        if resblock_updown
+                        else Downsample(
+                            ch, conv_resample, dims=dims, out_channels=out_ch
+                        )
+                    )
+                )
+                ch = out_ch
+                input_block_chans.append(ch)
+                ds *= 2
+                self._feature_size += ch
+
+        if num_head_channels == -1:
+            dim_head = ch // num_heads
+        else:
+            num_heads = ch // num_head_channels
+            dim_head = num_head_channels
+        if legacy:
+            #num_heads = 1
+            dim_head = ch // num_heads if use_spatial_transformer else num_head_channels
+        self.middle_block = TimestepEmbedSequential(
+            ResBlock(
+                ch,
+                time_embed_dim,
+                dropout,
+                dims=dims,
+                use_checkpoint=use_checkpoint,
+                use_scale_shift_norm=use_scale_shift_norm,
+            ),
+            AttentionBlock(
+                ch,
+                use_checkpoint=use_checkpoint,
+                num_heads=num_heads,
+                num_head_channels=dim_head,
+                use_new_attention_order=use_new_attention_order,
+            ) if not use_spatial_transformer else SpatialTransformer(  # always uses a self-attn
+                            ch, num_heads, dim_head, depth=transformer_depth, context_dim=context_dim,
+                            disable_self_attn=disable_middle_self_attn, use_linear=use_linear_in_transformer,
+                            use_checkpoint=use_checkpoint
+                        ),
+            ResBlock(
+                ch,
+                time_embed_dim,
+                dropout,
+                dims=dims,
+                use_checkpoint=use_checkpoint,
+                use_scale_shift_norm=use_scale_shift_norm,
+            ),
+        )
+        self._feature_size += ch
+
+        self.output_blocks = nn.ModuleList([])
+        for level, mult in list(enumerate(channel_mult))[::-1]:
+            for i in range(self.num_res_blocks[level] + 1):
+                ich = input_block_chans.pop()
+                layers = [
+                    ResBlock(
+                        ch + ich,
+                        time_embed_dim,
+                        dropout,
+                        out_channels=model_channels * mult,
+                        dims=dims,
+                        use_checkpoint=use_checkpoint,
+                        use_scale_shift_norm=use_scale_shift_norm,
+                    )
+                ]
+                ch = model_channels * mult
+                if ds in attention_resolutions:
+                    if num_head_channels == -1:
+                        dim_head = ch // num_heads
+                    else:
+                        num_heads = ch // num_head_channels
+                        dim_head = num_head_channels
+                    if legacy:
+                        #num_heads = 1
+                        dim_head = ch // num_heads if use_spatial_transformer else num_head_channels
+                    if exists(disable_self_attentions):
+                        disabled_sa = disable_self_attentions[level]
+                    else:
+                        disabled_sa = False
+
+                    if not exists(num_attention_blocks) or i < num_attention_blocks[level]:
+                        layers.append(
+                            AttentionBlock(
+                                ch,
+                                use_checkpoint=use_checkpoint,
+                                num_heads=num_heads_upsample,
+                                num_head_channels=dim_head,
+                                use_new_attention_order=use_new_attention_order,
+                            ) if not use_spatial_transformer else SpatialTransformer(
+                                ch, num_heads, dim_head, depth=transformer_depth, context_dim=context_dim,
+                                disable_self_attn=disabled_sa, use_linear=use_linear_in_transformer,
+                                use_checkpoint=use_checkpoint
+                            )
+                        )
+                if level and i == self.num_res_blocks[level]:
+                    out_ch = ch
+                    layers.append(
+                        ResBlock(
+                            ch,
+                            time_embed_dim,
+                            dropout,
+                            out_channels=out_ch,
+                            dims=dims,
+                            use_checkpoint=use_checkpoint,
+                            use_scale_shift_norm=use_scale_shift_norm,
+                            up=True,
+                        )
+                        if resblock_updown
+                        else Upsample(ch, conv_resample, dims=dims, out_channels=out_ch)
+                    )
+                    ds //= 2
+                self.output_blocks.append(TimestepEmbedSequential(*layers))
+                self._feature_size += ch
+
+        self.out = nn.Sequential(
+            normalization(ch),
+            nn.SiLU(),
+            zero_module(conv_nd(dims, model_channels, out_channels, 3, padding=1)),
+        )
+        if self.predict_codebook_ids:
+            self.id_predictor = nn.Sequential(
+            normalization(ch),
+            conv_nd(dims, model_channels, n_embed, 1),
+            #nn.LogSoftmax(dim=1)  # change to cross_entropy and produce non-normalized logits
+        )
+    
+    def forward(self, x, timesteps, context=None, y=None,**kwargs):
+        """
+        Apply the model to an input batch.
+        :param x: an [N x C x ...] Tensor of inputs.
+        :param timesteps: a 1-D batch of timesteps.
+        :param context: conditioning plugged in via crossattn
+        :param y: an [N] Tensor of labels, if class-conditional.
+        :return: an [N x C x ...] Tensor of outputs.
+        """
+        assert (y is not None) == (
+            self.num_classes is not None
+        ), "must specify y if and only if the model is class-conditional"
+        hs = []
+        t_emb = timestep_embedding(timesteps, self.model_channels, repeat_only=False)
+        emb = self.time_embed(t_emb)
+
+        if self.num_classes is not None:
+            assert y.shape[0] == x.shape[0]
+            emb = emb + self.label_emb(y)
+
+        h = x.type(self.dtype)
+        for module in self.input_blocks:
+            h = module(h, emb, context)
+            hs.append(h)
+        h = self.middle_block(h, emb, context)
+        for module in self.output_blocks:
+            h = th.cat([h, hs.pop()], dim=1)
+            h = module(h, emb, context)
+        h = h.type(x.dtype)
+        if self.predict_codebook_ids:
+            return self.id_predictor(h)
+        else:
+            return self.out(h)

model/util.py

@@ -0,0 +1,225 @@
+# adopted from
+# https://github.com/openai/improved-diffusion/blob/main/improved_diffusion/gaussian_diffusion.py
+# and
+# https://github.com/lucidrains/denoising-diffusion-pytorch/blob/7706bdfc6f527f58d33f84b7b522e61e6e3164b3/denoising_diffusion_pytorch/denoising_diffusion_pytorch.py
+# and
+# https://github.com/openai/guided-diffusion/blob/0ba878e517b276c45d1195eb29f6f5f72659a05b/guided_diffusion/nn.py
+#
+# thanks!
+
+
+import os
+import math
+from inspect import isfunction
+import torch
+import torch.nn as nn
+import numpy as np
+from einops import repeat
+
+
+def exists(val):
+    return val is not None
+
+
+def default(val, d):
+    if exists(val):
+        return val
+    return d() if isfunction(d) else d
+
+
+def checkpoint(func, inputs, params, flag):
+    """
+    Evaluate a function without caching intermediate activations, allowing for
+    reduced memory at the expense of extra compute in the backward pass.
+    :param func: the function to evaluate.
+    :param inputs: the argument sequence to pass to `func`.
+    :param params: a sequence of parameters `func` depends on but does not
+                   explicitly take as arguments.
+    :param flag: if False, disable gradient checkpointing.
+    """
+    if flag:
+        args = tuple(inputs) + tuple(params)
+        return CheckpointFunction.apply(func, len(inputs), *args)
+    else:
+        return func(*inputs)
+
+
+# class CheckpointFunction(torch.autograd.Function):
+#     @staticmethod
+#     def forward(ctx, run_function, length, *args):
+#         ctx.run_function = run_function
+#         ctx.input_tensors = list(args[:length])
+#         ctx.input_params = list(args[length:])
+#         ctx.gpu_autocast_kwargs = {"enabled": torch.is_autocast_enabled(),
+#                                    "dtype": torch.get_autocast_gpu_dtype(),
+#                                    "cache_enabled": torch.is_autocast_cache_enabled()}
+#         with torch.no_grad():
+#             output_tensors = ctx.run_function(*ctx.input_tensors)
+#         return output_tensors
+
+#     @staticmethod
+#     def backward(ctx, *output_grads):
+#         ctx.input_tensors = [x.detach().requires_grad_(True) for x in ctx.input_tensors]
+#         with torch.enable_grad(), \
+#                 torch.cuda.amp.autocast(**ctx.gpu_autocast_kwargs):
+#             # Fixes a bug where the first op in run_function modifies the
+#             # Tensor storage in place, which is not allowed for detach()'d
+#             # Tensors.
+#             shallow_copies = [x.view_as(x) for x in ctx.input_tensors]
+#             output_tensors = ctx.run_function(*shallow_copies)
+#         input_grads = torch.autograd.grad(
+#             output_tensors,
+#             ctx.input_tensors + ctx.input_params,
+#             output_grads,
+#             allow_unused=True,
+#         )
+#         del ctx.input_tensors
+#         del ctx.input_params
+#         del output_tensors
+#         return (None, None) + input_grads
+
+
+# Fixes: When we set unet parameters with requires_grad=False, the original CheckpointFunction
+# still tries to compute gradient for unet parameters.
+# https://discuss.pytorch.org/t/get-runtimeerror-one-of-the-differentiated-tensors-does-not-require-grad-in-pytorch-lightning/179738/6
+class CheckpointFunction(torch.autograd.Function):
+    @staticmethod
+    def forward(ctx, run_function, length, *args):
+        ctx.run_function = run_function
+        ctx.input_tensors = list(args[:length])
+        ctx.input_params = list(args[length:])
+        ctx.gpu_autocast_kwargs = {"enabled": torch.is_autocast_enabled(),
+                                   "dtype": torch.get_autocast_gpu_dtype(),
+                                   "cache_enabled": torch.is_autocast_cache_enabled()}
+        with torch.no_grad():
+            output_tensors = ctx.run_function(*ctx.input_tensors)
+        return output_tensors
+
+    @staticmethod
+    def backward(ctx, *output_grads):
+        ctx.input_tensors = [x.detach().requires_grad_(True) for x in ctx.input_tensors]
+        with torch.enable_grad(), \
+                torch.cuda.amp.autocast(**ctx.gpu_autocast_kwargs):
+            # Fixes a bug where the first op in run_function modifies the
+            # Tensor storage in place, which is not allowed for detach()'d
+            # Tensors.
+            shallow_copies = [x.view_as(x) for x in ctx.input_tensors]
+            output_tensors = ctx.run_function(*shallow_copies)
+        grads = torch.autograd.grad(
+            output_tensors,
+            ctx.input_tensors + [x for x in ctx.input_params if x.requires_grad],
+            output_grads,
+            allow_unused=True,
+        )
+        grads = list(grads)
+        # Assign gradients to the correct positions, matching None for those that do not require gradients
+        input_grads = []
+        for tensor in ctx.input_tensors + ctx.input_params:
+            if tensor.requires_grad:
+                input_grads.append(grads.pop(0))  # Get the next computed gradient
+            else:
+                input_grads.append(None)  # No gradient required for this tensor
+        del ctx.input_tensors
+        del ctx.input_params
+        del output_tensors
+        return (None, None) + tuple(input_grads)
+
+
+def timestep_embedding(timesteps, dim, max_period=10000, repeat_only=False):
+    """
+    Create sinusoidal timestep embeddings.
+    :param timesteps: a 1-D Tensor of N indices, one per batch element.
+                      These may be fractional.
+    :param dim: the dimension of the output.
+    :param max_period: controls the minimum frequency of the embeddings.
+    :return: an [N x dim] Tensor of positional embeddings.
+    """
+    if not repeat_only:
+        half = dim // 2
+        freqs = torch.exp(
+            -math.log(max_period) * torch.arange(start=0, end=half, dtype=torch.float32) / half
+        ).to(device=timesteps.device)
+        args = timesteps[:, None].float() * freqs[None]
+        embedding = torch.cat([torch.cos(args), torch.sin(args)], dim=-1)
+        if dim % 2:
+            embedding = torch.cat([embedding, torch.zeros_like(embedding[:, :1])], dim=-1)
+    else:
+        embedding = repeat(timesteps, 'b -> b d', d=dim)
+    return embedding
+
+
+def zero_module(module):
+    """
+    Zero out the parameters of a module and return it.
+    """
+    for p in module.parameters():
+        p.detach().zero_()
+    return module
+
+
+def scale_module(module, scale):
+    """
+    Scale the parameters of a module and return it.
+    """
+    for p in module.parameters():
+        p.detach().mul_(scale)
+    return module
+
+
+def mean_flat(tensor):
+    """
+    Take the mean over all non-batch dimensions.
+    """
+    return tensor.mean(dim=list(range(1, len(tensor.shape))))
+
+
+def normalization(channels):
+    """
+    Make a standard normalization layer.
+    :param channels: number of input channels.
+    :return: an nn.Module for normalization.
+    """
+    return GroupNorm32(32, channels)
+
+
+# PyTorch 1.7 has SiLU, but we support PyTorch 1.5.
+class SiLU(nn.Module):
+    def forward(self, x):
+        return x * torch.sigmoid(x)
+
+
+class GroupNorm32(nn.GroupNorm):
+    def forward(self, x):
+        return super().forward(x.float()).type(x.dtype)
+
+def conv_nd(dims, *args, **kwargs):
+    """
+    Create a 1D, 2D, or 3D convolution module.
+    """
+    if dims == 1:
+        return nn.Conv1d(*args, **kwargs)
+    elif dims == 2:
+        return nn.Conv2d(*args, **kwargs)
+    elif dims == 3:
+        return nn.Conv3d(*args, **kwargs)
+    raise ValueError(f"unsupported dimensions: {dims}")
+
+
+def linear(*args, **kwargs):
+    """
+    Create a linear module.
+    """
+    return nn.Linear(*args, **kwargs)
+
+
+def avg_pool_nd(dims, *args, **kwargs):
+    """
+    Create a 1D, 2D, or 3D average pooling module.
+    """
+    if dims == 1:
+        return nn.AvgPool1d(*args, **kwargs)
+    elif dims == 2:
+        return nn.AvgPool2d(*args, **kwargs)
+    elif dims == 3:
+        return nn.AvgPool3d(*args, **kwargs)
+    raise ValueError(f"unsupported dimensions: {dims}")

model/vae.py

@@ -0,0 +1,567 @@
+import math
+import torch
+import torch.nn as nn
+from torch.nn import functional as F
+import numpy as np
+from einops import rearrange
+from typing import Optional, Any
+
+from model.distributions import DiagonalGaussianDistribution
+from model.config import Config, AttnMode
+
+
+def nonlinearity(x):
+    # swish
+    return x*torch.sigmoid(x)
+
+
+def Normalize(in_channels, num_groups=32):
+    return torch.nn.GroupNorm(num_groups=num_groups, num_channels=in_channels, eps=1e-6, affine=True)
+
+
+class Upsample(nn.Module):
+    def __init__(self, in_channels, with_conv):
+        super().__init__()
+        self.with_conv = with_conv
+        if self.with_conv:
+            self.conv = torch.nn.Conv2d(in_channels,
+                                        in_channels,
+                                        kernel_size=3,
+                                        stride=1,
+                                        padding=1)
+
+    def forward(self, x):
+        x = torch.nn.functional.interpolate(x, scale_factor=2.0, mode="nearest")
+        if self.with_conv:
+            x = self.conv(x)
+        return x
+
+
+class Downsample(nn.Module):
+    def __init__(self, in_channels, with_conv):
+        super().__init__()
+        self.with_conv = with_conv
+        if self.with_conv:
+            # no asymmetric padding in torch conv, must do it ourselves
+            self.conv = torch.nn.Conv2d(in_channels,
+                                        in_channels,
+                                        kernel_size=3,
+                                        stride=2,
+                                        padding=0)
+
+    def forward(self, x):
+        if self.with_conv:
+            pad = (0,1,0,1)
+            x = torch.nn.functional.pad(x, pad, mode="constant", value=0)
+            x = self.conv(x)
+        else:
+            x = torch.nn.functional.avg_pool2d(x, kernel_size=2, stride=2)
+        return x
+
+
+class ResnetBlock(nn.Module):
+    def __init__(self, *, in_channels, out_channels=None, conv_shortcut=False,
+                 dropout, temb_channels=512):
+        super().__init__()
+        self.in_channels = in_channels
+        out_channels = in_channels if out_channels is None else out_channels
+        self.out_channels = out_channels
+        self.use_conv_shortcut = conv_shortcut
+
+        self.norm1 = Normalize(in_channels)
+        self.conv1 = torch.nn.Conv2d(in_channels,
+                                     out_channels,
+                                     kernel_size=3,
+                                     stride=1,
+                                     padding=1)
+        if temb_channels > 0:
+            self.temb_proj = torch.nn.Linear(temb_channels,
+                                             out_channels)
+        self.norm2 = Normalize(out_channels)
+        self.dropout = torch.nn.Dropout(dropout)
+        self.conv2 = torch.nn.Conv2d(out_channels,
+                                     out_channels,
+                                     kernel_size=3,
+                                     stride=1,
+                                     padding=1)
+        if self.in_channels != self.out_channels:
+            if self.use_conv_shortcut:
+                self.conv_shortcut = torch.nn.Conv2d(in_channels,
+                                                     out_channels,
+                                                     kernel_size=3,
+                                                     stride=1,
+                                                     padding=1)
+            else:
+                self.nin_shortcut = torch.nn.Conv2d(in_channels,
+                                                    out_channels,
+                                                    kernel_size=1,
+                                                    stride=1,
+                                                    padding=0)
+
+    def forward(self, x, temb):
+        h = x
+        h = self.norm1(h)
+        h = nonlinearity(h)
+        h = self.conv1(h)
+
+        if temb is not None:
+            h = h + self.temb_proj(nonlinearity(temb))[:,:,None,None]
+
+        h = self.norm2(h)
+        h = nonlinearity(h)
+        h = self.dropout(h)
+        h = self.conv2(h)
+
+        if self.in_channels != self.out_channels:
+            if self.use_conv_shortcut:
+                x = self.conv_shortcut(x)
+            else:
+                x = self.nin_shortcut(x)
+
+        return x+h
+
+
+class AttnBlock(nn.Module):
+    def __init__(self, in_channels):
+        super().__init__()
+        print(f"building AttnBlock (vanilla) with {in_channels} in_channels")
+
+        self.in_channels = in_channels
+
+        self.norm = Normalize(in_channels)
+        self.q = torch.nn.Conv2d(in_channels,
+                                 in_channels,
+                                 kernel_size=1,
+                                 stride=1,
+                                 padding=0)
+        self.k = torch.nn.Conv2d(in_channels,
+                                 in_channels,
+                                 kernel_size=1,
+                                 stride=1,
+                                 padding=0)
+        self.v = torch.nn.Conv2d(in_channels,
+                                 in_channels,
+                                 kernel_size=1,
+                                 stride=1,
+                                 padding=0)
+        self.proj_out = torch.nn.Conv2d(in_channels,
+                                        in_channels,
+                                        kernel_size=1,
+                                        stride=1,
+                                        padding=0)
+
+    def forward(self, x):
+        h_ = x
+        h_ = self.norm(h_)
+        q = self.q(h_)
+        k = self.k(h_)
+        v = self.v(h_)
+
+        # compute attention
+        b,c,h,w = q.shape
+        q = q.reshape(b,c,h*w)
+        q = q.permute(0,2,1)   # b,hw,c
+        k = k.reshape(b,c,h*w) # b,c,hw
+        w_ = torch.bmm(q,k)     # b,hw,hw    w[b,i,j]=sum_c q[b,i,c]k[b,c,j]
+        w_ = w_ * (int(c)**(-0.5))
+        w_ = torch.nn.functional.softmax(w_, dim=2)
+
+        # attend to values
+        v = v.reshape(b,c,h*w)
+        w_ = w_.permute(0,2,1)   # b,hw,hw (first hw of k, second of q)
+        h_ = torch.bmm(v,w_)     # b, c,hw (hw of q) h_[b,c,j] = sum_i v[b,c,i] w_[b,i,j]
+        h_ = h_.reshape(b,c,h,w)
+
+        h_ = self.proj_out(h_)
+
+        return x+h_
+
+
+class MemoryEfficientAttnBlock(nn.Module):
+    """
+        Uses xformers efficient implementation,
+        see https://github.com/MatthieuTPHR/diffusers/blob/d80b531ff8060ec1ea982b65a1b8df70f73aa67c/src/diffusers/models/attention.py#L223
+        Note: this is a single-head self-attention operation
+    """
+    #
+    def __init__(self, in_channels):
+        super().__init__()
+        print(f"building MemoryEfficientAttnBlock (xformers) with {in_channels} in_channels")
+        self.in_channels = in_channels
+
+        self.norm = Normalize(in_channels)
+        self.q = torch.nn.Conv2d(in_channels,
+                                 in_channels,
+                                 kernel_size=1,
+                                 stride=1,
+                                 padding=0)
+        self.k = torch.nn.Conv2d(in_channels,
+                                 in_channels,
+                                 kernel_size=1,
+                                 stride=1,
+                                 padding=0)
+        self.v = torch.nn.Conv2d(in_channels,
+                                 in_channels,
+                                 kernel_size=1,
+                                 stride=1,
+                                 padding=0)
+        self.proj_out = torch.nn.Conv2d(in_channels,
+                                        in_channels,
+                                        kernel_size=1,
+                                        stride=1,
+                                        padding=0)
+        self.attention_op: Optional[Any] = None
+
+    def forward(self, x):
+        h_ = x
+        h_ = self.norm(h_)
+        q = self.q(h_)
+        k = self.k(h_)
+        v = self.v(h_)
+
+        # compute attention
+        B, C, H, W = q.shape
+        q, k, v = map(lambda x: rearrange(x, 'b c h w -> b (h w) c'), (q, k, v))
+
+        q, k, v = map(
+            lambda t: t.unsqueeze(3)
+            .reshape(B, t.shape[1], 1, C)
+            .permute(0, 2, 1, 3)
+            .reshape(B * 1, t.shape[1], C)
+            .contiguous(),
+            (q, k, v),
+        )
+        out = Config.xformers.ops.memory_efficient_attention(q, k, v, attn_bias=None, op=self.attention_op)
+        
+        out = (
+            out.unsqueeze(0)
+            .reshape(B, 1, out.shape[1], C)
+            .permute(0, 2, 1, 3)
+            .reshape(B, out.shape[1], C)
+        )
+        out = rearrange(out, 'b (h w) c -> b c h w', b=B, h=H, w=W, c=C)
+        out = self.proj_out(out)
+        return x+out
+
+
+class SDPAttnBlock(nn.Module):
+
+    def __init__(self, in_channels):
+        super().__init__()
+        print(f"building SDPAttnBlock (sdp) with {in_channels} in_channels")
+        self.in_channels = in_channels
+
+        self.norm = Normalize(in_channels)
+        self.q = torch.nn.Conv2d(in_channels,
+                                 in_channels,
+                                 kernel_size=1,
+                                 stride=1,
+                                 padding=0)
+        self.k = torch.nn.Conv2d(in_channels,
+                                 in_channels,
+                                 kernel_size=1,
+                                 stride=1,
+                                 padding=0)
+        self.v = torch.nn.Conv2d(in_channels,
+                                 in_channels,
+                                 kernel_size=1,
+                                 stride=1,
+                                 padding=0)
+        self.proj_out = torch.nn.Conv2d(in_channels,
+                                        in_channels,
+                                        kernel_size=1,
+                                        stride=1,
+                                        padding=0)
+
+    def forward(self, x):
+        h_ = x
+        h_ = self.norm(h_)
+        q = self.q(h_)
+        k = self.k(h_)
+        v = self.v(h_)
+
+        # compute attention
+        B, C, H, W = q.shape
+        q, k, v = map(lambda x: rearrange(x, 'b c h w -> b (h w) c'), (q, k, v))
+
+        q, k, v = map(
+            lambda t: t.unsqueeze(3)
+            .reshape(B, t.shape[1], 1, C)
+            .permute(0, 2, 1, 3)
+            .reshape(B * 1, t.shape[1], C)
+            .contiguous(),
+            (q, k, v),
+        )
+        out = F.scaled_dot_product_attention(q, k, v)
+
+        out = (
+            out.unsqueeze(0)
+            .reshape(B, 1, out.shape[1], C)
+            .permute(0, 2, 1, 3)
+            .reshape(B, out.shape[1], C)
+        )
+        out = rearrange(out, 'b (h w) c -> b c h w', b=B, h=H, w=W, c=C)
+        out = self.proj_out(out)
+        return x+out
+
+
+def make_attn(in_channels, attn_type="vanilla", attn_kwargs=None):
+    assert attn_type in ["vanilla", "sdp", "xformers", "linear", "none"], f'attn_type {attn_type} unknown'
+    if attn_type == "vanilla":
+        assert attn_kwargs is None
+        return AttnBlock(in_channels)
+    elif attn_type == "sdp":
+        return SDPAttnBlock(in_channels)
+    elif attn_type == "xformers":
+        return MemoryEfficientAttnBlock(in_channels)
+    elif attn_type == "none":
+        return nn.Identity(in_channels)
+    else:
+        raise NotImplementedError()
+
+
+class Encoder(nn.Module):
+    def __init__(self, *, ch, out_ch, ch_mult=(1,2,4,8), num_res_blocks,
+                 attn_resolutions, dropout=0.0, resamp_with_conv=True, in_channels,
+                 resolution, z_channels, double_z=True, use_linear_attn=False,
+                 **ignore_kwargs):
+        super().__init__()
+        ### setup attention type
+        if Config.attn_mode == AttnMode.SDP:
+            attn_type = "sdp"
+        elif Config.attn_mode == AttnMode.XFORMERS:
+            attn_type = "xformers"
+        else:
+            attn_type = "vanilla"
+        if use_linear_attn: attn_type = "linear"
+        self.ch = ch
+        self.temb_ch = 0
+        self.num_resolutions = len(ch_mult)
+        self.num_res_blocks = num_res_blocks
+        self.resolution = resolution
+        self.in_channels = in_channels
+
+        # downsampling
+        self.conv_in = torch.nn.Conv2d(in_channels,
+                                       self.ch,
+                                       kernel_size=3,
+                                       stride=1,
+                                       padding=1)
+
+        curr_res = resolution
+        in_ch_mult = (1,)+tuple(ch_mult)
+        self.in_ch_mult = in_ch_mult
+        self.down = nn.ModuleList()
+        for i_level in range(self.num_resolutions):
+            block = nn.ModuleList()
+            attn = nn.ModuleList()
+            block_in = ch*in_ch_mult[i_level]
+            block_out = ch*ch_mult[i_level]
+            for i_block in range(self.num_res_blocks):
+                block.append(ResnetBlock(in_channels=block_in,
+                                         out_channels=block_out,
+                                         temb_channels=self.temb_ch,
+                                         dropout=dropout))
+                block_in = block_out
+                if curr_res in attn_resolutions:
+                    attn.append(make_attn(block_in, attn_type=attn_type))
+            down = nn.Module()
+            down.block = block
+            down.attn = attn
+            if i_level != self.num_resolutions-1:
+                down.downsample = Downsample(block_in, resamp_with_conv)
+                curr_res = curr_res // 2
+            self.down.append(down)
+
+        # middle
+        self.mid = nn.Module()
+        self.mid.block_1 = ResnetBlock(in_channels=block_in,
+                                       out_channels=block_in,
+                                       temb_channels=self.temb_ch,
+                                       dropout=dropout)
+        self.mid.attn_1 = make_attn(block_in, attn_type=attn_type)
+        self.mid.block_2 = ResnetBlock(in_channels=block_in,
+                                       out_channels=block_in,
+                                       temb_channels=self.temb_ch,
+                                       dropout=dropout)
+
+        # end
+        self.norm_out = Normalize(block_in)
+        self.conv_out = torch.nn.Conv2d(block_in,
+                                        2*z_channels if double_z else z_channels,
+                                        kernel_size=3,
+                                        stride=1,
+                                        padding=1)
+
+    def forward(self, x):
+        # timestep embedding
+        temb = None
+
+        # downsampling
+        hs = [self.conv_in(x)]
+        for i_level in range(self.num_resolutions):
+            for i_block in range(self.num_res_blocks):
+                h = self.down[i_level].block[i_block](hs[-1], temb)
+                if len(self.down[i_level].attn) > 0:
+                    h = self.down[i_level].attn[i_block](h)
+                hs.append(h)
+            if i_level != self.num_resolutions-1:
+                hs.append(self.down[i_level].downsample(hs[-1]))
+
+        # middle
+        h = hs[-1]
+        h = self.mid.block_1(h, temb)
+        h = self.mid.attn_1(h)
+        h = self.mid.block_2(h, temb)
+
+        # end
+        h = self.norm_out(h)
+        h = nonlinearity(h)
+        h = self.conv_out(h)
+        return h
+
+
+class Decoder(nn.Module):
+    def __init__(self, *, ch, out_ch, ch_mult=(1,2,4,8), num_res_blocks,
+                 attn_resolutions, dropout=0.0, resamp_with_conv=True, in_channels,
+                 resolution, z_channels, give_pre_end=False, tanh_out=False, use_linear_attn=False,
+                 **ignorekwargs):
+        super().__init__()
+        ### setup attention type
+        if Config.attn_mode == AttnMode.SDP:
+            attn_type = "sdp"
+        elif Config.attn_mode == AttnMode.XFORMERS:
+            attn_type = "xformers"
+        else:
+            attn_type = "vanilla"
+        if use_linear_attn: attn_type = "linear"
+        self.ch = ch
+        self.temb_ch = 0
+        self.num_resolutions = len(ch_mult)
+        self.num_res_blocks = num_res_blocks
+        self.resolution = resolution
+        self.in_channels = in_channels
+        self.give_pre_end = give_pre_end
+        self.tanh_out = tanh_out
+
+        # compute in_ch_mult, block_in and curr_res at lowest res
+        in_ch_mult = (1,)+tuple(ch_mult)
+        block_in = ch*ch_mult[self.num_resolutions-1]
+        curr_res = resolution // 2**(self.num_resolutions-1)
+        self.z_shape = (1,z_channels,curr_res,curr_res)
+
+        # z to block_in
+        self.conv_in = torch.nn.Conv2d(z_channels,
+                                       block_in,
+                                       kernel_size=3,
+                                       stride=1,
+                                       padding=1)
+
+        # middle
+        self.mid = nn.Module()
+        self.mid.block_1 = ResnetBlock(in_channels=block_in,
+                                       out_channels=block_in,
+                                       temb_channels=self.temb_ch,
+                                       dropout=dropout)
+        self.mid.attn_1 = make_attn(block_in, attn_type=attn_type)
+        self.mid.block_2 = ResnetBlock(in_channels=block_in,
+                                       out_channels=block_in,
+                                       temb_channels=self.temb_ch,
+                                       dropout=dropout)
+
+        # upsampling
+        self.up = nn.ModuleList()
+        for i_level in reversed(range(self.num_resolutions)):
+            block = nn.ModuleList()
+            attn = nn.ModuleList()
+            block_out = ch*ch_mult[i_level]
+            for i_block in range(self.num_res_blocks+1):
+                block.append(ResnetBlock(in_channels=block_in,
+                                         out_channels=block_out,
+                                         temb_channels=self.temb_ch,
+                                         dropout=dropout))
+                block_in = block_out
+                if curr_res in attn_resolutions:
+                    attn.append(make_attn(block_in, attn_type=attn_type))
+            up = nn.Module()
+            up.block = block
+            up.attn = attn
+            if i_level != 0:
+                up.upsample = Upsample(block_in, resamp_with_conv)
+                curr_res = curr_res * 2
+            self.up.insert(0, up) # prepend to get consistent order
+
+        # end
+        self.norm_out = Normalize(block_in)
+        self.conv_out = torch.nn.Conv2d(block_in,
+                                        out_ch,
+                                        kernel_size=3,
+                                        stride=1,
+                                        padding=1)
+
+    def forward(self, z):
+        #assert z.shape[1:] == self.z_shape[1:]
+        self.last_z_shape = z.shape
+
+        # timestep embedding
+        temb = None
+
+        # z to block_in
+        h = self.conv_in(z)
+
+        # middle
+        h = self.mid.block_1(h, temb)
+        h = self.mid.attn_1(h)
+        h = self.mid.block_2(h, temb)
+
+        # upsampling
+        for i_level in reversed(range(self.num_resolutions)):
+            for i_block in range(self.num_res_blocks+1):
+                h = self.up[i_level].block[i_block](h, temb)
+                if len(self.up[i_level].attn) > 0:
+                    h = self.up[i_level].attn[i_block](h)
+            if i_level != 0:
+                h = self.up[i_level].upsample(h)
+
+        # end
+        if self.give_pre_end:
+            return h
+
+        h = self.norm_out(h)
+        h = nonlinearity(h)
+        h = self.conv_out(h)
+        if self.tanh_out:
+            h = torch.tanh(h)
+        return h
+
+
+class AutoencoderKL(nn.Module):
+    
+    def __init__(self, ddconfig, embed_dim):
+        super().__init__()
+        self.encoder = Encoder(**ddconfig)
+        self.decoder = Decoder(**ddconfig)
+        assert ddconfig["double_z"]
+        self.quant_conv = torch.nn.Conv2d(2*ddconfig["z_channels"], 2*embed_dim, 1)
+        self.post_quant_conv = torch.nn.Conv2d(embed_dim, ddconfig["z_channels"], 1)
+        self.embed_dim = embed_dim
+    
+    def encode(self, x):
+        h = self.encoder(x)
+        moments = self.quant_conv(h)
+        posterior = DiagonalGaussianDistribution(moments)
+        return posterior
+
+    def decode(self, z):
+        z = self.post_quant_conv(z)
+        dec = self.decoder(z)
+        return dec
+
+    def forward(self, input, sample_posterior=True):
+        posterior = self.encode(input)
+        if sample_posterior:
+            z = posterior.sample()
+        else:
+            z = posterior.mode()
+        dec = self.decode(z)
+        return dec, posterior