File size: 5,230 Bytes
7667a87 |
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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 |
from typing import Optional
import numpy as np
import torch
import torch.nn as nn
from .unet_2d_blocks import UNetMidBlock2D, get_down_block, get_up_block
class Encoder(nn.Module):
def __init__(
self,
in_channels=3,
out_channels=3,
down_block_types=("DownEncoderBlock2D",),
block_out_channels=(64,),
layers_per_block=2,
norm_num_groups=32,
act_fn="silu",
double_z=True,
):
super().__init__()
self.layers_per_block = layers_per_block
self.conv_in = torch.nn.Conv2d(in_channels, block_out_channels[0], kernel_size=3, stride=1, padding=1)
self.mid_block = None
self.down_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=self.layers_per_block,
in_channels=input_channel,
out_channels=output_channel,
add_downsample=not is_final_block,
resnet_eps=1e-6,
downsample_padding=0,
resnet_act_fn=act_fn,
resnet_groups=norm_num_groups,
attn_num_head_channels=None,
temb_channels=None,
)
self.down_blocks.append(down_block)
# mid
self.mid_block = UNetMidBlock2D(
in_channels=block_out_channels[-1],
resnet_eps=1e-6,
resnet_act_fn=act_fn,
output_scale_factor=1,
resnet_time_scale_shift="default",
attn_num_head_channels=None,
resnet_groups=norm_num_groups,
temb_channels=None,
)
# out
self.conv_norm_out = nn.GroupNorm(num_channels=block_out_channels[-1], num_groups=norm_num_groups, eps=1e-6)
self.conv_act = nn.SiLU()
conv_out_channels = 2 * out_channels if double_z else out_channels
self.conv_out = nn.Conv2d(block_out_channels[-1], conv_out_channels, 3, padding=1)
def forward(self, x):
sample = x
sample = self.conv_in(sample)
# down
for down_block in self.down_blocks:
sample = down_block(sample)
# middle
sample = self.mid_block(sample)
# post-process
sample = self.conv_norm_out(sample)
sample = self.conv_act(sample)
sample = self.conv_out(sample)
return sample
class Decoder(nn.Module):
def __init__(
self,
in_channels=3,
out_channels=3,
up_block_types=("UpDecoderBlock2D",),
block_out_channels=(64,),
layers_per_block=2,
norm_num_groups=32,
act_fn="silu",
):
super().__init__()
self.layers_per_block = layers_per_block
self.conv_in = nn.Conv2d(in_channels, block_out_channels[-1], kernel_size=3, stride=1, padding=1)
self.mid_block = None
self.up_blocks = nn.ModuleList([])
# mid
self.mid_block = UNetMidBlock2D(
in_channels=block_out_channels[-1],
resnet_eps=1e-6,
resnet_act_fn=act_fn,
output_scale_factor=1,
resnet_time_scale_shift="default",
attn_num_head_channels=None,
resnet_groups=norm_num_groups,
temb_channels=None,
)
# 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]
is_final_block = i == len(block_out_channels) - 1
up_block = get_up_block(
up_block_type,
num_layers=self.layers_per_block + 1,
in_channels=prev_output_channel,
out_channels=output_channel,
prev_output_channel=None,
add_upsample=not is_final_block,
resnet_eps=1e-6,
resnet_act_fn=act_fn,
resnet_groups=norm_num_groups,
attn_num_head_channels=None,
temb_channels=None,
)
self.up_blocks.append(up_block)
prev_output_channel = output_channel
# out
self.conv_norm_out = nn.GroupNorm(num_channels=block_out_channels[0], num_groups=norm_num_groups, eps=1e-6)
self.conv_act = nn.SiLU()
self.conv_out = nn.Conv2d(block_out_channels[0], out_channels, 3, padding=1)
def forward(self, z):
sample = z
sample = self.conv_in(sample)
# middle
sample = self.mid_block(sample)
# up
for up_block in self.up_blocks:
sample = up_block(sample)
# post-process
sample = self.conv_norm_out(sample)
sample = self.conv_act(sample)
sample = self.conv_out(sample)
return sample
|