Spaces:
Running on Zero
Running on Zero
File size: 5,833 Bytes
b701455 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 | """Residual block components for diffusion models."""
from abc import abstractmethod
from typing import Optional, Any, Dict
import torch
import torch.nn as nn
import torch.nn.functional as F
from src.NeuralNetwork import transformer
from src.Attention import Attention
from src.cond import cast
from src.sample import sampling_util
ops = cast.disable_weight_init
class TimestepBlock1(nn.Module):
"""Abstract timestep block."""
@abstractmethod
def forward(self, x: torch.Tensor, emb: torch.Tensor) -> torch.Tensor:
pass
def forward_timestep_embed1(ts, x, emb, context=None, transformer_options={}, output_shape=None,
time_context=None, num_video_frames=None, image_only_indicator=None):
"""Forward pass for timestep embedding."""
for layer in ts:
if isinstance(layer, TimestepBlock1):
x = layer(x, emb)
elif isinstance(layer, transformer.SpatialTransformer):
x = layer(x, context, transformer_options)
if "transformer_index" in transformer_options:
transformer_options["transformer_index"] += 1
elif isinstance(layer, Upsample1):
x = layer(x, output_shape=output_shape)
else:
x = layer(x)
return x
class Upsample1(nn.Module):
"""Upsample layer with optional conv."""
def __init__(self, channels, use_conv, dims=2, out_channels=None, padding=1,
dtype=None, device=None, operations=ops):
super().__init__()
self.channels = channels
self.out_channels = out_channels or channels
self.use_conv = use_conv
if use_conv:
self.conv = operations.conv_nd(dims, channels, self.out_channels, 3,
padding=padding, dtype=dtype, device=device)
def forward(self, x, output_shape=None):
shape = [x.shape[2] * 2, x.shape[3] * 2] if output_shape is None else [output_shape[2], output_shape[3]]
x = F.interpolate(x, size=shape, mode="nearest")
return self.conv(x) if self.use_conv else x
class Downsample1(nn.Module):
"""Downsample layer."""
def __init__(self, channels, use_conv, dims=2, out_channels=None, padding=1,
dtype=None, device=None, operations=ops):
super().__init__()
self.channels = channels
self.out_channels = out_channels or channels
stride = 2 if dims != 3 else (1, 2, 2)
self.op = operations.conv_nd(dims, channels, self.out_channels, 3, stride=stride,
padding=padding, dtype=dtype, device=device)
def forward(self, x):
return self.op(x)
class ResBlock1(TimestepBlock1):
"""Residual block with timestep embedding."""
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,
kernel_size=3, exchange_temb_dims=False, skip_t_emb=False,
dtype=None, device=None, operations=ops):
super().__init__()
self.out_channels = out_channels or channels
self.use_checkpoint = use_checkpoint
self.skip_t_emb = skip_t_emb
padding = kernel_size // 2
self.in_layers = nn.Sequential(
operations.GroupNorm(32, channels, dtype=dtype, device=device),
nn.SiLU(),
operations.conv_nd(dims, channels, self.out_channels, kernel_size, padding=padding, dtype=dtype, device=device))
self.emb_layers = nn.Sequential(
nn.SiLU(),
operations.Linear(emb_channels, 2 * self.out_channels if use_scale_shift_norm else self.out_channels,
dtype=dtype, device=device))
self.out_layers = nn.Sequential(
operations.GroupNorm(32, self.out_channels, dtype=dtype, device=device),
nn.SiLU(),
nn.Dropout(p=dropout),
operations.conv_nd(dims, self.out_channels, self.out_channels, kernel_size, padding=padding, dtype=dtype, device=device))
self.skip_connection = (nn.Identity() if self.out_channels == channels
else operations.conv_nd(dims, channels, self.out_channels, 1, dtype=dtype, device=device))
def forward(self, x, emb):
return sampling_util.checkpoint(self._forward, (x, emb), self.parameters(), self.use_checkpoint)
def _forward(self, x, emb):
h = self.in_layers(x)
if not self.skip_t_emb:
emb_out = self.emb_layers(emb).type(h.dtype)
while len(emb_out.shape) < len(h.shape):
emb_out = emb_out[..., None]
h = h + emb_out
return self.skip_connection(x) + self.out_layers(h)
class ResnetBlock(nn.Module):
"""VAE-style ResNet block."""
def __init__(self, *, in_channels, out_channels=None, conv_shortcut=False, dropout, temb_channels=512):
super().__init__()
out_channels = out_channels or in_channels
self.in_channels, self.out_channels = in_channels, out_channels
self.norm1 = Attention.Normalize(in_channels)
self.conv1 = ops.Conv2d(in_channels, out_channels, 3, 1, 1)
self.norm2 = Attention.Normalize(out_channels)
self.dropout = nn.Dropout(dropout, inplace=True)
self.conv2 = ops.Conv2d(out_channels, out_channels, 3, 1, 1)
self.nin_shortcut = ops.Conv2d(in_channels, out_channels, 1) if in_channels != out_channels else None
self.swish = nn.SiLU(inplace=True)
def forward(self, x, temb):
h = self.swish(self.norm1(x))
h = self.conv1(h)
h = self.dropout(self.swish(self.norm2(h)))
h = self.conv2(h)
return (self.nin_shortcut(x) if self.nin_shortcut else x) + h
|