models/archs/vqgan_arch.py

# pytorch_diffusion + derived encoder decoder
import math
from urllib.request import proxy_bypass

import numpy as np
import torch
import torch.nn as nn
import torch.nn.functional as F
from einops import rearrange


class VectorQuantizer(nn.Module):
    """
    Improved version over VectorQuantizer, can be used as a drop-in replacement. Mostly
    avoids costly matrix multiplications and allows for post-hoc remapping of indices.
    """

    # NOTE: due to a bug the beta term was applied to the wrong term. for
    # backwards compatibility we use the buggy version by default, but you can
    # specify legacy=False to fix it.
    def __init__(self,
                 n_e,
                 e_dim,
                 beta,
                 remap=None,
                 unknown_index="random",
                 sane_index_shape=False,
                 legacy=True):
        super().__init__()
        self.n_e = n_e
        self.e_dim = e_dim
        self.beta = beta
        self.legacy = legacy

        self.embedding = nn.Embedding(self.n_e, self.e_dim)
        self.embedding.weight.data.uniform_(-1.0 / self.n_e, 1.0 / self.n_e)

        self.remap = remap
        if self.remap is not None:
            self.register_buffer("used", torch.tensor(np.load(self.remap)))
            self.re_embed = self.used.shape[0]
            self.unknown_index = unknown_index  # "random" or "extra" or integer
            if self.unknown_index == "extra":
                self.unknown_index = self.re_embed
                self.re_embed = self.re_embed + 1
            print(f"Remapping {self.n_e} indices to {self.re_embed} indices. "
                  f"Using {self.unknown_index} for unknown indices.")
        else:
            self.re_embed = n_e

        self.sane_index_shape = sane_index_shape

    def remap_to_used(self, inds):
        ishape = inds.shape
        assert len(ishape) > 1
        inds = inds.reshape(ishape[0], -1)
        used = self.used.to(inds)
        match = (inds[:, :, None] == used[None, None, ...]).long()
        new = match.argmax(-1)
        unknown = match.sum(2) < 1
        if self.unknown_index == "random":
            new[unknown] = torch.randint(
                0, self.re_embed,
                size=new[unknown].shape).to(device=new.device)
        else:
            new[unknown] = self.unknown_index
        return new.reshape(ishape)

    def unmap_to_all(self, inds):
        ishape = inds.shape
        assert len(ishape) > 1
        inds = inds.reshape(ishape[0], -1)
        used = self.used.to(inds)
        if self.re_embed > self.used.shape[0]:  # extra token
            inds[inds >= self.used.shape[0]] = 0  # simply set to zero
        back = torch.gather(used[None, :][inds.shape[0] * [0], :], 1, inds)
        return back.reshape(ishape)

    def forward(self, z, temp=None, rescale_logits=False, return_logits=False):
        assert temp is None or temp == 1.0, "Only for interface compatible with Gumbel"
        assert rescale_logits == False, "Only for interface compatible with Gumbel"
        assert return_logits == False, "Only for interface compatible with Gumbel"
        # reshape z -> (batch, height, width, channel) and flatten
        z = rearrange(z, 'b c h w -> b h w c').contiguous()
        z_flattened = z.view(-1, self.e_dim)
        # distances from z to embeddings e_j (z - e)^2 = z^2 + e^2 - 2 e * z

        d = torch.sum(z_flattened ** 2, dim=1, keepdim=True) + \
            torch.sum(self.embedding.weight**2, dim=1) - 2 * \
            torch.einsum('bd,dn->bn', z_flattened, rearrange(self.embedding.weight, 'n d -> d n'))

        min_encoding_indices = torch.argmin(d, dim=1)
        z_q = self.embedding(min_encoding_indices).view(z.shape)
        perplexity = None
        min_encodings = None

        # compute loss for embedding
        if not self.legacy:
            loss = self.beta * torch.mean((z_q.detach()-z)**2) + \
                   torch.mean((z_q - z.detach()) ** 2)
        else:
            loss = torch.mean((z_q.detach()-z)**2) + self.beta * \
                   torch.mean((z_q - z.detach()) ** 2)

        # preserve gradients
        z_q = z + (z_q - z).detach()

        # reshape back to match original input shape
        z_q = rearrange(z_q, 'b h w c -> b c h w').contiguous()

        if self.remap is not None:
            min_encoding_indices = min_encoding_indices.reshape(
                z.shape[0], -1)  # add batch axis
            min_encoding_indices = self.remap_to_used(min_encoding_indices)
            min_encoding_indices = min_encoding_indices.reshape(-1,
                                                                1)  # flatten

        if self.sane_index_shape:
            min_encoding_indices = min_encoding_indices.reshape(
                z_q.shape[0], z_q.shape[2], z_q.shape[3])

        return z_q, loss, (perplexity, min_encodings, min_encoding_indices)

    def get_codebook_entry(self, indices, shape):
        # shape specifying (batch, height, width, channel)
        if self.remap is not None:
            indices = indices.reshape(shape[0], -1)  # add batch axis
            indices = self.unmap_to_all(indices)
            indices = indices.reshape(-1)  # flatten again

        # get quantized latent vectors
        z_q = self.embedding(indices)

        if shape is not None:
            z_q = z_q.view(shape)
            # reshape back to match original input shape
            z_q = z_q.permute(0, 3, 1, 2).contiguous()

        return z_q


class VectorQuantizerTexture(nn.Module):
    """
    Improved version over VectorQuantizer, can be used as a drop-in replacement. Mostly
    avoids costly matrix multiplications and allows for post-hoc remapping of indices.
    """

    # NOTE: due to a bug the beta term was applied to the wrong term. for
    # backwards compatibility we use the buggy version by default, but you can
    # specify legacy=False to fix it.
    def __init__(self,
                 n_e,
                 e_dim,
                 beta,
                 remap=None,
                 unknown_index="random",
                 sane_index_shape=False,
                 legacy=True):
        super().__init__()
        self.n_e = n_e
        self.e_dim = e_dim
        self.beta = beta
        self.legacy = legacy

        # TODO: decide number of embeddings
        self.embedding_list = nn.ModuleList(
            [nn.Embedding(self.n_e, self.e_dim) for i in range(18)])
        for embedding in self.embedding_list:
            embedding.weight.data.uniform_(-1.0 / self.n_e, 1.0 / self.n_e)

        self.remap = remap
        if self.remap is not None:
            self.register_buffer("used", torch.tensor(np.load(self.remap)))
            self.re_embed = self.used.shape[0]
            self.unknown_index = unknown_index  # "random" or "extra" or integer
            if self.unknown_index == "extra":
                self.unknown_index = self.re_embed
                self.re_embed = self.re_embed + 1
            print(f"Remapping {self.n_e} indices to {self.re_embed} indices. "
                  f"Using {self.unknown_index} for unknown indices.")
        else:
            self.re_embed = n_e

        self.sane_index_shape = sane_index_shape

    def remap_to_used(self, inds):
        ishape = inds.shape
        assert len(ishape) > 1
        inds = inds.reshape(ishape[0], -1)
        used = self.used.to(inds)
        match = (inds[:, :, None] == used[None, None, ...]).long()
        new = match.argmax(-1)
        unknown = match.sum(2) < 1
        if self.unknown_index == "random":
            new[unknown] = torch.randint(
                0, self.re_embed,
                size=new[unknown].shape).to(device=new.device)
        else:
            new[unknown] = self.unknown_index
        return new.reshape(ishape)

    def unmap_to_all(self, inds):
        ishape = inds.shape
        assert len(ishape) > 1
        inds = inds.reshape(ishape[0], -1)
        used = self.used.to(inds)
        if self.re_embed > self.used.shape[0]:  # extra token
            inds[inds >= self.used.shape[0]] = 0  # simply set to zero
        back = torch.gather(used[None, :][inds.shape[0] * [0], :], 1, inds)
        return back.reshape(ishape)

    def forward(self,
                z,
                segm_map,
                temp=None,
                rescale_logits=False,
                return_logits=False):
        assert temp is None or temp == 1.0, "Only for interface compatible with Gumbel"
        assert rescale_logits == False, "Only for interface compatible with Gumbel"
        assert return_logits == False, "Only for interface compatible with Gumbel"

        segm_map = F.interpolate(segm_map, size=z.size()[2:], mode='nearest')
        # reshape z -> (batch, height, width, channel) and flatten
        z = rearrange(z, 'b c h w -> b h w c').contiguous()
        z_flattened = z.view(-1, self.e_dim)

        # flatten segm_map (b, h, w)
        segm_map_flatten = segm_map.view(-1)

        z_q = torch.zeros_like(z_flattened)
        min_encoding_indices_list = []
        min_encoding_indices_continual = torch.full(
            segm_map_flatten.size(),
            fill_value=-1,
            dtype=torch.long,
            device=segm_map_flatten.device)
        for codebook_idx in range(18):
            min_encoding_indices = torch.full(
                segm_map_flatten.size(),
                fill_value=-1,
                dtype=torch.long,
                device=segm_map_flatten.device)
            if torch.sum(segm_map_flatten == codebook_idx) > 0:
                z_selected = z_flattened[segm_map_flatten == codebook_idx]
                # distances from z to embeddings e_j (z - e)^2 = z^2 + e^2 - 2 e * z
                d_selected = torch.sum(
                    z_selected**2, dim=1, keepdim=True) + torch.sum(
                        self.embedding_list[codebook_idx].weight**2,
                        dim=1) - 2 * torch.einsum(
                            'bd,dn->bn', z_selected,
                            rearrange(self.embedding_list[codebook_idx].weight,
                                      'n d -> d n'))
                min_encoding_indices_selected = torch.argmin(d_selected, dim=1)
                z_q_selected = self.embedding_list[codebook_idx](
                    min_encoding_indices_selected)
                z_q[segm_map_flatten == codebook_idx] = z_q_selected
                min_encoding_indices[
                    segm_map_flatten ==
                    codebook_idx] = min_encoding_indices_selected
                min_encoding_indices_continual[
                    segm_map_flatten ==
                    codebook_idx] = min_encoding_indices_selected + 1024 * codebook_idx
            min_encoding_indices = min_encoding_indices.reshape(
                z.shape[0], z.shape[1], z.shape[2])
            min_encoding_indices_list.append(min_encoding_indices)

        min_encoding_indices_continual = min_encoding_indices_continual.reshape(
            z.shape[0], z.shape[1], z.shape[2])
        z_q = z_q.view(z.shape)
        perplexity = None

        # compute loss for embedding
        if not self.legacy:
            loss = self.beta * torch.mean((z_q.detach()-z)**2) + \
                   torch.mean((z_q - z.detach()) ** 2)
        else:
            loss = torch.mean((z_q.detach()-z)**2) + self.beta * \
                   torch.mean((z_q - z.detach()) ** 2)

        # preserve gradients
        z_q = z + (z_q - z).detach()

        # reshape back to match original input shape
        z_q = rearrange(z_q, 'b h w c -> b c h w').contiguous()

        return z_q, loss, (perplexity, min_encoding_indices_continual,
                           min_encoding_indices_list)

    def get_codebook_entry(self, indices_list, segm_map, shape):
        # flatten segm_map (b, h, w)
        segm_map = F.interpolate(
            segm_map, size=(shape[1], shape[2]), mode='nearest')
        segm_map_flatten = segm_map.view(-1)

        z_q = torch.zeros((shape[0] * shape[1] * shape[2]),
                          self.e_dim).to(segm_map.device)
        for codebook_idx in range(18):
            if torch.sum(segm_map_flatten == codebook_idx) > 0:
                min_encoding_indices_selected = indices_list[
                    codebook_idx].view(-1)[segm_map_flatten == codebook_idx]
                z_q_selected = self.embedding_list[codebook_idx](
                    min_encoding_indices_selected)
                z_q[segm_map_flatten == codebook_idx] = z_q_selected

        z_q = z_q.view(shape)
        # reshape back to match original input shape
        z_q = z_q.permute(0, 3, 1, 2).contiguous()

        return z_q


def sample_patches(inputs, patch_size=3, stride=1):
    """Extract sliding local patches from an input feature tensor.
    The sampled pathes are row-major.
    Args:
        inputs (Tensor): the input feature maps, shape: (n, c, h, w).
        patch_size (int): the spatial size of sampled patches. Default: 3.
        stride (int): the stride of sampling. Default: 1.
    Returns:
        patches (Tensor): extracted patches, shape: (n, c *  patch_size *
            patch_size, n_patches).
    """

    patches = F.unfold(inputs, (patch_size, patch_size), stride=stride)

    return patches


class VectorQuantizerSpatialTextureAware(nn.Module):
    """
    Improved version over VectorQuantizer, can be used as a drop-in replacement. Mostly
    avoids costly matrix multiplications and allows for post-hoc remapping of indices.
    """

    # NOTE: due to a bug the beta term was applied to the wrong term. for
    # backwards compatibility we use the buggy version by default, but you can
    # specify legacy=False to fix it.
    def __init__(self,
                 n_e,
                 e_dim,
                 beta,
                 spatial_size,
                 remap=None,
                 unknown_index="random",
                 sane_index_shape=False,
                 legacy=True):
        super().__init__()
        self.n_e = n_e
        self.e_dim = e_dim * spatial_size * spatial_size
        self.beta = beta
        self.legacy = legacy
        self.spatial_size = spatial_size

        # TODO: decide number of embeddings
        self.embedding_list = nn.ModuleList(
            [nn.Embedding(self.n_e, self.e_dim) for i in range(18)])
        for embedding in self.embedding_list:
            embedding.weight.data.uniform_(-1.0 / self.n_e, 1.0 / self.n_e)

        self.remap = remap
        if self.remap is not None:
            self.register_buffer("used", torch.tensor(np.load(self.remap)))
            self.re_embed = self.used.shape[0]
            self.unknown_index = unknown_index  # "random" or "extra" or integer
            if self.unknown_index == "extra":
                self.unknown_index = self.re_embed
                self.re_embed = self.re_embed + 1
            print(f"Remapping {self.n_e} indices to {self.re_embed} indices. "
                  f"Using {self.unknown_index} for unknown indices.")
        else:
            self.re_embed = n_e

        self.sane_index_shape = sane_index_shape

    def forward(self,
                z,
                segm_map,
                temp=None,
                rescale_logits=False,
                return_logits=False):
        assert temp is None or temp == 1.0, "Only for interface compatible with Gumbel"
        assert rescale_logits == False, "Only for interface compatible with Gumbel"
        assert return_logits == False, "Only for interface compatible with Gumbel"

        segm_map = F.interpolate(
            segm_map,
            size=(z.size(2) // self.spatial_size,
                  z.size(3) // self.spatial_size),
            mode='nearest')

        # reshape z -> (batch, height, width, channel) and flatten
        # z = rearrange(z, 'b c h w -> b h w c').contiguous() ?
        z_patches = sample_patches(
            z, patch_size=self.spatial_size,
            stride=self.spatial_size).permute(0, 2, 1)
        z_patches_flattened = z_patches.reshape(-1, self.e_dim)
        # distances from z to embeddings e_j (z - e)^2 = z^2 + e^2 - 2 e * z

        # flatten segm_map (b, h, w)
        segm_map_flatten = segm_map.view(-1)

        z_q = torch.zeros_like(z_patches_flattened)
        min_encoding_indices_list = []
        min_encoding_indices_continual = torch.full(
            segm_map_flatten.size(),
            fill_value=-1,
            dtype=torch.long,
            device=segm_map_flatten.device)

        for codebook_idx in range(18):
            min_encoding_indices = torch.full(
                segm_map_flatten.size(),
                fill_value=-1,
                dtype=torch.long,
                device=segm_map_flatten.device)
            if torch.sum(segm_map_flatten == codebook_idx) > 0:
                z_selected = z_patches_flattened[segm_map_flatten ==
                                                 codebook_idx]
                # distances from z to embeddings e_j (z - e)^2 = z^2 + e^2 - 2 e * z
                d_selected = torch.sum(
                    z_selected**2, dim=1, keepdim=True) + torch.sum(
                        self.embedding_list[codebook_idx].weight**2,
                        dim=1) - 2 * torch.einsum(
                            'bd,dn->bn', z_selected,
                            rearrange(self.embedding_list[codebook_idx].weight,
                                      'n d -> d n'))
                min_encoding_indices_selected = torch.argmin(d_selected, dim=1)
                z_q_selected = self.embedding_list[codebook_idx](
                    min_encoding_indices_selected)
                z_q[segm_map_flatten == codebook_idx] = z_q_selected
                min_encoding_indices[
                    segm_map_flatten ==
                    codebook_idx] = min_encoding_indices_selected
                min_encoding_indices_continual[
                    segm_map_flatten ==
                    codebook_idx] = min_encoding_indices_selected + self.n_e * codebook_idx
            min_encoding_indices = min_encoding_indices.reshape(
                z_patches.shape[0], segm_map.shape[2], segm_map.shape[3])
            min_encoding_indices_list.append(min_encoding_indices)

        z_q = F.fold(
            z_q.view(z_patches.shape).permute(0, 2, 1),
            z.size()[2:],
            kernel_size=(self.spatial_size, self.spatial_size),
            stride=self.spatial_size)

        perplexity = None

        # compute loss for embedding
        if not self.legacy:
            loss = self.beta * torch.mean((z_q.detach()-z)**2) + \
                   torch.mean((z_q - z.detach()) ** 2)
        else:
            loss = torch.mean((z_q.detach()-z)**2) + self.beta * \
                   torch.mean((z_q - z.detach()) ** 2)

        # preserve gradients
        z_q = z + (z_q - z).detach()

        return z_q, loss, (perplexity, min_encoding_indices_continual,
                           min_encoding_indices_list)

    def get_codebook_entry(self, indices_list, segm_map, shape):
        # flatten segm_map (b, h, w)
        segm_map = F.interpolate(
            segm_map, size=(shape[1], shape[2]), mode='nearest')
        segm_map_flatten = segm_map.view(-1)

        z_q = torch.zeros((shape[0] * shape[1] * shape[2]),
                          self.e_dim).to(segm_map.device)
        for codebook_idx in range(18):
            if torch.sum(segm_map_flatten == codebook_idx) > 0:
                min_encoding_indices_selected = indices_list[
                    codebook_idx].view(-1)[segm_map_flatten == codebook_idx]
                z_q_selected = self.embedding_list[codebook_idx](
                    min_encoding_indices_selected)
                z_q[segm_map_flatten == codebook_idx] = z_q_selected

        z_q = F.fold(
            z_q.view(((shape[0], shape[1] * shape[2],
                       self.e_dim))).permute(0, 2, 1),
            (shape[1] * self.spatial_size, shape[2] * self.spatial_size),
            kernel_size=(self.spatial_size, self.spatial_size),
            stride=self.spatial_size)

        return z_q


def get_timestep_embedding(timesteps, embedding_dim):
    """
    This matches the implementation in Denoising Diffusion Probabilistic Models:
    From Fairseq.
    Build sinusoidal embeddings.
    This matches the implementation in tensor2tensor, but differs slightly
    from the description in Section 3.5 of "Attention Is All You Need".
    """
    assert len(timesteps.shape) == 1

    half_dim = embedding_dim // 2
    emb = math.log(10000) / (half_dim - 1)
    emb = torch.exp(torch.arange(half_dim, dtype=torch.float32) * -emb)
    emb = emb.to(device=timesteps.device)
    emb = timesteps.float()[:, None] * emb[None, :]
    emb = torch.cat([torch.sin(emb), torch.cos(emb)], dim=1)
    if embedding_dim % 2 == 1:  # zero pad
        emb = torch.nn.functional.pad(emb, (0, 1, 0, 0))
    return emb


def nonlinearity(x):
    # swish
    return x * torch.sigmoid(x)


def Normalize(in_channels):
    return torch.nn.GroupNorm(
        num_groups=32, 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__()
        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 Model(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,
                 use_timestep=True):
        super().__init__()
        self.ch = ch
        self.temb_ch = self.ch * 4
        self.num_resolutions = len(ch_mult)
        self.num_res_blocks = num_res_blocks
        self.resolution = resolution
        self.in_channels = in_channels

        self.use_timestep = use_timestep
        if self.use_timestep:
            # timestep embedding
            self.temb = nn.Module()
            self.temb.dense = nn.ModuleList([
                torch.nn.Linear(self.ch, self.temb_ch),
                torch.nn.Linear(self.temb_ch, self.temb_ch),
            ])

        # 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.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(AttnBlock(block_in))
            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 = AttnBlock(block_in)
        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]
            skip_in = ch * ch_mult[i_level]
            for i_block in range(self.num_res_blocks + 1):
                if i_block == self.num_res_blocks:
                    skip_in = ch * in_ch_mult[i_level]
                block.append(
                    ResnetBlock(
                        in_channels=block_in + skip_in,
                        out_channels=block_out,
                        temb_channels=self.temb_ch,
                        dropout=dropout))
                block_in = block_out
                if curr_res in attn_resolutions:
                    attn.append(AttnBlock(block_in))
            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, x, t=None):
        #assert x.shape[2] == x.shape[3] == self.resolution

        if self.use_timestep:
            # timestep embedding
            assert t is not None
            temb = get_timestep_embedding(t, self.ch)
            temb = self.temb.dense[0](temb)
            temb = nonlinearity(temb)
            temb = self.temb.dense[1](temb)
        else:
            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)

        # 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](torch.cat([h, hs.pop()],
                                                              dim=1), 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
        h = self.norm_out(h)
        h = nonlinearity(h)
        h = self.conv_out(h)
        return h


class Encoder(nn.Module):

    def __init__(self,
                 ch,
                 num_res_blocks,
                 attn_resolutions,
                 in_channels,
                 resolution,
                 z_channels,
                 ch_mult=(1, 2, 4, 8),
                 dropout=0.0,
                 resamp_with_conv=True,
                 double_z=True):
        super().__init__()
        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.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(AttnBlock(block_in))
            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 = AttnBlock(block_in)
        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):
        #assert x.shape[2] == x.shape[3] == self.resolution, "{}, {}, {}".format(x.shape[2], x.shape[3], self.resolution)

        # 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,
                 in_channels,
                 resolution,
                 z_channels,
                 ch,
                 out_ch,
                 num_res_blocks,
                 attn_resolutions,
                 ch_mult=(1, 2, 4, 8),
                 dropout=0.0,
                 resamp_with_conv=True,
                 give_pre_end=False):
        super().__init__()
        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

        # 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 // 2)
        print("Working with z of shape {} = {} dimensions.".format(
            self.z_shape, np.prod(self.z_shape)))

        # 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 = AttnBlock(block_in)
        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(AttnBlock(block_in))
            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, bot_h=None):
        #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)
            if i_level == 4 and bot_h is not None:
                h += bot_h

        # end
        if self.give_pre_end:
            return h

        h = self.norm_out(h)
        h = nonlinearity(h)
        h = self.conv_out(h)
        return h

    def get_feature_top(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)
            if i_level == 4:
                return h

    def get_feature_middle(self, z, mid_h):
        #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)
            if i_level == 4:
                h += mid_h
            if i_level == 3:
                return h


class DecoderRes(nn.Module):

    def __init__(self,
                 in_channels,
                 resolution,
                 z_channels,
                 ch,
                 num_res_blocks,
                 ch_mult=(1, 2, 4, 8),
                 dropout=0.0,
                 give_pre_end=False):
        super().__init__()
        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

        # 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 // 2)
        print("Working with z of shape {} = {} dimensions.".format(
            self.z_shape, np.prod(self.z_shape)))

        # 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 = AttnBlock(block_in)
        self.mid.block_2 = ResnetBlock(
            in_channels=block_in,
            out_channels=block_in,
            temb_channels=self.temb_ch,
            dropout=dropout)

    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)

        return h


# patch based discriminator
class Discriminator(nn.Module):

    def __init__(self, nc, ndf, n_layers=3):
        super().__init__()

        layers = [
            nn.Conv2d(nc, ndf, kernel_size=4, stride=2, padding=1),
            nn.LeakyReLU(0.2, True)
        ]
        ndf_mult = 1
        ndf_mult_prev = 1
        for n in range(1,
                       n_layers):  # gradually increase the number of filters
            ndf_mult_prev = ndf_mult
            ndf_mult = min(2**n, 8)
            layers += [
                nn.Conv2d(
                    ndf * ndf_mult_prev,
                    ndf * ndf_mult,
                    kernel_size=4,
                    stride=2,
                    padding=1,
                    bias=False),
                nn.BatchNorm2d(ndf * ndf_mult),
                nn.LeakyReLU(0.2, True)
            ]

        ndf_mult_prev = ndf_mult
        ndf_mult = min(2**n_layers, 8)

        layers += [
            nn.Conv2d(
                ndf * ndf_mult_prev,
                ndf * ndf_mult,
                kernel_size=4,
                stride=1,
                padding=1,
                bias=False),
            nn.BatchNorm2d(ndf * ndf_mult),
            nn.LeakyReLU(0.2, True)
        ]

        layers += [
            nn.Conv2d(ndf * ndf_mult, 1, kernel_size=4, stride=1, padding=1)
        ]  # output 1 channel prediction map
        self.main = nn.Sequential(*layers)

    def forward(self, x):
        return self.main(x)