attention-refocusing's picture
Update gligen/ldm/modules/diffusionmodules/openaimodel.py
f412da9
from abc import abstractmethod
from functools import partial
import math
import numpy as np
import random
import torch as th
import torch.nn as nn
import torch.nn.functional as F
from ldm.modules.diffusionmodules.util import (
conv_nd,
linear,
avg_pool_nd,
zero_module,
normalization,
timestep_embedding,
)
from ldm.modules.attention import SpatialTransformer
# from .positionnet import PositionNet
from torch.utils import checkpoint
from ldm.util import instantiate_from_config
from copy import deepcopy
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, objs,t):
probs = []
self_prob_list = []
for layer in self:
if isinstance(layer, TimestepBlock):
x = layer(x, emb)
elif isinstance(layer, SpatialTransformer):
x, prob, self_prob = layer(x, context, objs,t)
probs.append(prob)
self_prob_list.append(self_prob)
else:
x = layer(x)
return x, probs, self_prob_list
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
# )
# if self.use_checkpoint and x.requires_grad:
# return checkpoint.checkpoint(self._forward, x, emb )
# else:
return self._forward(x, emb)
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 UNetModel(nn.Module):
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,
use_checkpoint=False,
num_heads=8,
use_scale_shift_norm=False,
transformer_depth=1,
positive_len = 768,
context_dim=None,
fuser_type = None,
is_inpaint = False,
is_style = False,
grounding_downsampler = None,
):
super().__init__()
self.image_size = image_size
self.in_channels = in_channels
self.model_channels = model_channels
self.out_channels = out_channels
self.num_res_blocks = num_res_blocks
self.attention_resolutions = attention_resolutions
self.dropout = dropout
self.channel_mult = channel_mult
self.conv_resample = conv_resample
self.use_checkpoint = use_checkpoint
self.num_heads = num_heads
self.context_dim = context_dim
self.fuser_type = fuser_type
self.is_inpaint = is_inpaint
self.positive_len = positive_len
assert fuser_type in ["gatedSA","gatedSA2","gatedCA"]
self.grounding_tokenizer_input = None # set externally
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.downsample_net = None
self.additional_channel_from_downsampler = 0
self.first_conv_type = "SD"
self.first_conv_restorable = True
if grounding_downsampler is not None:
self.downsample_net = instantiate_from_config(grounding_downsampler)
self.additional_channel_from_downsampler = self.downsample_net.out_dim
self.first_conv_type = "GLIGEN"
if is_inpaint:
# The new added channels are: masked image (encoded image) and mask, which is 4+1
in_c = in_channels+self.additional_channel_from_downsampler+in_channels+1
self.first_conv_restorable = False # in inpaint; You must use extra channels to take in masked real image
else:
in_c = in_channels+self.additional_channel_from_downsampler
self.input_blocks = nn.ModuleList([TimestepEmbedSequential(conv_nd(dims, in_c, model_channels, 3, padding=1))])
input_block_chans = [model_channels]
ch = model_channels
ds = 1
# = = = = = = = = = = = = = = = = = = = = Down Branch = = = = = = = = = = = = = = = = = = = = #
for level, mult in enumerate(channel_mult):
for _ in range(num_res_blocks):
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:
dim_head = ch // num_heads
layers.append(SpatialTransformer(ch, key_dim=context_dim, value_dim=context_dim, n_heads=num_heads, d_head=dim_head, depth=transformer_depth, fuser_type=fuser_type, use_checkpoint=use_checkpoint))
self.input_blocks.append(TimestepEmbedSequential(*layers))
input_block_chans.append(ch)
if level != len(channel_mult) - 1: # will not go to this downsample branch in the last feature
out_ch = ch
self.input_blocks.append( TimestepEmbedSequential( Downsample(ch, conv_resample, dims=dims, out_channels=out_ch ) ) )
ch = out_ch
input_block_chans.append(ch)
ds *= 2
dim_head = ch // num_heads
# self.input_blocks = [ C | RT RT D | RT RT D | RT RT D | R R ]
# = = = = = = = = = = = = = = = = = = = = BottleNeck = = = = = = = = = = = = = = = = = = = = #
self.middle_block = TimestepEmbedSequential(
ResBlock(ch,
time_embed_dim,
dropout,
dims=dims,
use_checkpoint=use_checkpoint,
use_scale_shift_norm=use_scale_shift_norm),
SpatialTransformer(ch, key_dim=context_dim, value_dim=context_dim, n_heads=num_heads, d_head=dim_head, depth=transformer_depth, fuser_type=fuser_type, use_checkpoint=use_checkpoint),
ResBlock(ch,
time_embed_dim,
dropout,
dims=dims,
use_checkpoint=use_checkpoint,
use_scale_shift_norm=use_scale_shift_norm))
# = = = = = = = = = = = = = = = = = = = = Up Branch = = = = = = = = = = = = = = = = = = = = #
self.output_blocks = nn.ModuleList([])
for level, mult in list(enumerate(channel_mult))[::-1]:
for i in range(num_res_blocks + 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:
dim_head = ch // num_heads
layers.append( SpatialTransformer(ch, key_dim=context_dim, value_dim=context_dim, n_heads=num_heads, d_head=dim_head, depth=transformer_depth, fuser_type=fuser_type, use_checkpoint=use_checkpoint) )
if level and i == num_res_blocks:
out_ch = ch
layers.append( Upsample(ch, conv_resample, dims=dims, out_channels=out_ch) )
ds //= 2
self.output_blocks.append(TimestepEmbedSequential(*layers))
# self.output_blocks = [ R R RU | RT RT RTU | RT RT RTU | RT RT RT ]
self.out = nn.Sequential(
normalization(ch),
nn.SiLU(),
zero_module(conv_nd(dims, model_channels, out_channels, 3, padding=1)),
)
# self.position_net = instantiate_from_config(grounding_tokenizer)
from .text_grounding_net import PositionNet
self.position_net = PositionNet(in_dim=positive_len, out_dim=context_dim)
def restore_first_conv_from_SD(self):
if self.first_conv_restorable:
device = self.input_blocks[0][0].weight.device
SD_weights = th.load("gligen/SD_input_conv_weight_bias.pth")
self.GLIGEN_first_conv_state_dict = deepcopy(self.input_blocks[0][0].state_dict())
self.input_blocks[0][0] = conv_nd(2, 4, 320, 3, padding=1)
self.input_blocks[0][0].load_state_dict(SD_weights)
self.input_blocks[0][0].to(device)
self.first_conv_type = "SD"
else:
print("First conv layer is not restorable and skipped this process, probably because this is an inpainting model?")
def restore_first_conv_from_GLIGEN(self):
breakpoint() # TODO
def forward_position_net(self,input):
# import pdb; pdb.set_trace()
if ("boxes" in input):
boxes, masks, text_embeddings = input["boxes"], input["masks"], input["text_embeddings"]
_ , self.max_box, _ = text_embeddings.shape
else:
dtype = input["x"].dtype
batch = input["x"].shape[0]
device = input["x"].device
boxes = th.zeros(batch, self.max_box, 4,).type(dtype).to(device)
masks = th.zeros(batch, self.max_box).type(dtype).to(device)
text_embeddings = th.zeros(batch, self.max_box, self.positive_len).type(dtype).to(device)
if self.training and random.random() < 0.1: # random drop for guidance
boxes, masks, text_embeddings = boxes*0, masks*0, text_embeddings*0
objs = self.position_net( boxes, masks, text_embeddings ) # B*N*C
return objs
def forward_position_net_with_image(self,input):
if ("boxes" in input):
boxes = input["boxes"]
masks = input["masks"]
text_masks = input["text_masks"]
image_masks = input["image_masks"]
text_embeddings = input["text_embeddings"]
image_embeddings = input["image_embeddings"]
_ , self.max_box, _ = text_embeddings.shape
else:
dtype = input["x"].dtype
batch = input["x"].shape[0]
device = input["x"].device
boxes = th.zeros(batch, self.max_box, 4,).type(dtype).to(device)
masks = th.zeros(batch, self.max_box).type(dtype).to(device)
text_masks = th.zeros(batch, self.max_box).type(dtype).to(device)
image_masks = th.zeros(batch, self.max_box).type(dtype).to(device)
text_embeddings = th.zeros(batch, self.max_box, self.positive_len).type(dtype).to(device)
image_embeddings = th.zeros(batch, self.max_box, self.positive_len).type(dtype).to(device)
if self.training and random.random() < 0.1: # random drop for guidance
boxes = boxes*0
masks = masks*0
text_masks = text_masks*0
image_masks = image_masks*0
text_embeddings = text_embeddings*0
image_embeddings = image_embeddings*0
objs = self.position_net( boxes, masks, text_masks, image_masks, text_embeddings, image_embeddings ) # B*N*C
return objs
def forward(self, input,unc=False):
if ("boxes" in input):
# grounding_input = input["grounding_input"]
boxes, masks, text_embeddings = input["boxes"], input["masks"], input["text_embeddings"]
_ , self.max_box, _ = text_embeddings.shape
else:
# Guidance null case
# grounding_input = self.grounding_tokenizer_input.get_null_input()
# boxes, masks, text_embeddings = input["boxes"]*0, input["masks"]*0, input["text_embeddings"]*0
dtype = input["x"].dtype
batch = input["x"].shape[0]
device = input["x"].device
boxes = th.zeros(batch, self.max_box, 4,).type(dtype).to(device)
masks = th.zeros(batch, self.max_box).type(dtype).to(device)
text_masks = th.zeros(batch, self.max_box).type(dtype).to(device)
image_masks = th.zeros(batch, self.max_box).type(dtype).to(device)
text_embeddings = th.zeros(batch, self.max_box, self.positive_len).type(dtype).to(device)
image_embeddings = th.zeros(batch, self.max_box, self.positive_len).type(dtype).to(device)
if self.training and random.random() < 0.1 : # random drop for guidance
boxes, masks, text_embeddings = boxes*0, masks*0, text_embeddings*0
objs = self.position_net( boxes, masks, text_embeddings )
# Time embedding
t_emb = timestep_embedding(input["timesteps"], self.model_channels, repeat_only=False)
emb = self.time_embed(t_emb)
# input tensor
h = input["x"]
t = input["timesteps"]
if self.downsample_net != None and self.first_conv_type=="GLIGEN":
temp = self.downsample_net(input["grounding_extra_input"])
h = th.cat( [h,temp], dim=1 )
if self.is_inpaint:#self.inpaint_mode:
if self.downsample_net != None:
breakpoint() # TODO: think about this case
h = th.cat( [h, input["inpainting_extra_input"]], dim=1 )
# Text input
context = input["context"]
# Start forwarding
hs = []
probs_first = []
self_prob_list_first = []
for module in self.input_blocks:
h,prob, self_prob = module(h, emb, context, objs,t)
hs.append(h)
probs_first.append(prob)
self_prob_list_first.append(self_prob)
h,mid_prob, self_prob_list_second = self.middle_block(h, emb, context, objs,t)
probs_third = []
self_prob_list_third = []
for module in self.output_blocks:
h = th.cat([h, hs.pop()], dim=1)
h, prob, self_prob = module(h, emb, context, objs,t)
probs_third.append(prob)
self_prob_list_third.append(self_prob)
return self.out(h),probs_third , mid_prob, probs_first, self_prob_list_first, [self_prob_list_second], self_prob_list_third