187 lines
6.9 KiB
Python
187 lines
6.9 KiB
Python
from typing import Dict, Union
|
|
|
|
import torch
|
|
import torch.nn as nn
|
|
|
|
from ..configuration_utils import ConfigMixin, register_to_config
|
|
from ..modeling_utils import ModelMixin
|
|
from .embeddings import GaussianFourierProjection, TimestepEmbedding, Timesteps
|
|
from .unet_blocks import UNetMidBlock2D, get_down_block, get_up_block
|
|
|
|
|
|
class UNet2DModel(ModelMixin, ConfigMixin):
|
|
@register_to_config
|
|
def __init__(
|
|
self,
|
|
sample_size=None,
|
|
in_channels=3,
|
|
out_channels=3,
|
|
center_input_sample=False,
|
|
time_embedding_type="positional",
|
|
freq_shift=0,
|
|
flip_sin_to_cos=True,
|
|
down_block_types=("DownBlock2D", "AttnDownBlock2D", "AttnDownBlock2D", "AttnDownBlock2D"),
|
|
up_block_types=("AttnUpBlock2D", "AttnUpBlock2D", "AttnUpBlock2D", "UpBlock2D"),
|
|
block_out_channels=(224, 448, 672, 896),
|
|
layers_per_block=2,
|
|
mid_block_scale_factor=1,
|
|
downsample_padding=1,
|
|
act_fn="silu",
|
|
attention_head_dim=8,
|
|
norm_num_groups=32,
|
|
norm_eps=1e-5,
|
|
):
|
|
super().__init__()
|
|
|
|
self.sample_size = sample_size
|
|
time_embed_dim = block_out_channels[0] * 4
|
|
|
|
# input
|
|
self.conv_in = nn.Conv2d(in_channels, block_out_channels[0], kernel_size=3, padding=(1, 1))
|
|
|
|
# time
|
|
if time_embedding_type == "fourier":
|
|
self.time_proj = GaussianFourierProjection(embedding_size=block_out_channels[0], scale=16)
|
|
timestep_input_dim = 2 * block_out_channels[0]
|
|
elif time_embedding_type == "positional":
|
|
self.time_proj = Timesteps(block_out_channels[0], flip_sin_to_cos, freq_shift)
|
|
timestep_input_dim = block_out_channels[0]
|
|
|
|
self.time_embedding = TimestepEmbedding(timestep_input_dim, time_embed_dim)
|
|
|
|
self.down_blocks = nn.ModuleList([])
|
|
self.mid_block = None
|
|
self.up_blocks = nn.ModuleList([])
|
|
|
|
# down
|
|
output_channel = block_out_channels[0]
|
|
for i, down_block_type in enumerate(down_block_types):
|
|
input_channel = output_channel
|
|
output_channel = block_out_channels[i]
|
|
is_final_block = i == len(block_out_channels) - 1
|
|
|
|
down_block = get_down_block(
|
|
down_block_type,
|
|
num_layers=layers_per_block,
|
|
in_channels=input_channel,
|
|
out_channels=output_channel,
|
|
temb_channels=time_embed_dim,
|
|
add_downsample=not is_final_block,
|
|
resnet_eps=norm_eps,
|
|
resnet_act_fn=act_fn,
|
|
attn_num_head_channels=attention_head_dim,
|
|
downsample_padding=downsample_padding,
|
|
)
|
|
self.down_blocks.append(down_block)
|
|
|
|
# mid
|
|
self.mid_block = UNetMidBlock2D(
|
|
in_channels=block_out_channels[-1],
|
|
temb_channels=time_embed_dim,
|
|
resnet_eps=norm_eps,
|
|
resnet_act_fn=act_fn,
|
|
output_scale_factor=mid_block_scale_factor,
|
|
resnet_time_scale_shift="default",
|
|
attn_num_head_channels=attention_head_dim,
|
|
resnet_groups=norm_num_groups,
|
|
)
|
|
|
|
# up
|
|
reversed_block_out_channels = list(reversed(block_out_channels))
|
|
output_channel = reversed_block_out_channels[0]
|
|
for i, up_block_type in enumerate(up_block_types):
|
|
prev_output_channel = output_channel
|
|
output_channel = reversed_block_out_channels[i]
|
|
input_channel = reversed_block_out_channels[min(i + 1, len(block_out_channels) - 1)]
|
|
|
|
is_final_block = i == len(block_out_channels) - 1
|
|
|
|
up_block = get_up_block(
|
|
up_block_type,
|
|
num_layers=layers_per_block + 1,
|
|
in_channels=input_channel,
|
|
out_channels=output_channel,
|
|
prev_output_channel=prev_output_channel,
|
|
temb_channels=time_embed_dim,
|
|
add_upsample=not is_final_block,
|
|
resnet_eps=norm_eps,
|
|
resnet_act_fn=act_fn,
|
|
attn_num_head_channels=attention_head_dim,
|
|
)
|
|
self.up_blocks.append(up_block)
|
|
prev_output_channel = output_channel
|
|
|
|
# out
|
|
num_groups_out = norm_num_groups if norm_num_groups is not None else min(block_out_channels[0] // 4, 32)
|
|
self.conv_norm_out = nn.GroupNorm(num_channels=block_out_channels[0], num_groups=num_groups_out, eps=norm_eps)
|
|
self.conv_act = nn.SiLU()
|
|
self.conv_out = nn.Conv2d(block_out_channels[0], out_channels, 3, padding=1)
|
|
|
|
def forward(
|
|
self, sample: torch.FloatTensor, timestep: Union[torch.Tensor, float, int]
|
|
) -> Dict[str, torch.FloatTensor]:
|
|
|
|
# 0. center input if necessary
|
|
if self.config.center_input_sample:
|
|
sample = 2 * sample - 1.0
|
|
|
|
# 1. time
|
|
timesteps = timestep
|
|
if not torch.is_tensor(timesteps):
|
|
timesteps = torch.tensor([timesteps], dtype=torch.long, device=sample.device)
|
|
elif torch.is_tensor(timesteps) and len(timesteps.shape) == 0:
|
|
timesteps = timesteps[None].to(sample.device)
|
|
|
|
# broadcast to batch dimension
|
|
timesteps = timesteps.broadcast_to(sample.shape[0])
|
|
|
|
t_emb = self.time_proj(timesteps)
|
|
emb = self.time_embedding(t_emb)
|
|
|
|
# 2. pre-process
|
|
skip_sample = sample
|
|
sample = self.conv_in(sample)
|
|
|
|
# 3. down
|
|
down_block_res_samples = (sample,)
|
|
for downsample_block in self.down_blocks:
|
|
if hasattr(downsample_block, "skip_conv"):
|
|
sample, res_samples, skip_sample = downsample_block(
|
|
hidden_states=sample, temb=emb, skip_sample=skip_sample
|
|
)
|
|
else:
|
|
sample, res_samples = downsample_block(hidden_states=sample, temb=emb)
|
|
|
|
down_block_res_samples += res_samples
|
|
|
|
# 4. mid
|
|
sample = self.mid_block(sample, emb)
|
|
|
|
# 5. up
|
|
skip_sample = None
|
|
for upsample_block in self.up_blocks:
|
|
res_samples = down_block_res_samples[-len(upsample_block.resnets) :]
|
|
down_block_res_samples = down_block_res_samples[: -len(upsample_block.resnets)]
|
|
|
|
if hasattr(upsample_block, "skip_conv"):
|
|
sample, skip_sample = upsample_block(sample, res_samples, emb, skip_sample)
|
|
else:
|
|
sample = upsample_block(sample, res_samples, emb)
|
|
|
|
# 6. post-process
|
|
# make sure hidden states is in float32
|
|
# when running in half-precision
|
|
sample = self.conv_norm_out(sample.float()).type(sample.dtype)
|
|
sample = self.conv_act(sample)
|
|
sample = self.conv_out(sample)
|
|
|
|
if skip_sample is not None:
|
|
sample += skip_sample
|
|
|
|
if self.config.time_embedding_type == "fourier":
|
|
timesteps = timesteps.reshape((sample.shape[0], *([1] * len(sample.shape[1:]))))
|
|
sample = sample / timesteps
|
|
|
|
output = {"sample": sample}
|
|
|
|
return output
|