Spaces:
Running on Zero
Running on Zero
File size: 6,338 Bytes
e66463a | 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 | import torch
from torch import nn
from torch.nn import functional as F
import math
class SelfAttention(nn.Module):
def __init__(self, n_heads, d_embed, in_proj_bias=True, out_proj_bias=True):
super().__init__()
self.in_proj = nn.Linear(d_embed, 3 * d_embed, bias=in_proj_bias)
self.out_proj = nn.Linear(d_embed, d_embed, bias=out_proj_bias)
self.n_heads = n_heads
self.d_head = d_embed // n_heads
def forward(self, x, causal_mask=False):
input_shape = x.shape
batch_size, sequence_length, d_embed = input_shape
interim_shape = (batch_size, sequence_length, self.n_heads, self.d_head)
q, k, v = self.in_proj(x).chunk(3, dim=-1)
q = q.view(interim_shape).transpose(1, 2)
k = k.view(interim_shape).transpose(1, 2)
v = v.view(interim_shape).transpose(1, 2)
weight = q @ k.transpose(-1, -2)
if causal_mask:
mask = torch.ones_like(weight, dtype=torch.bool).triu(1)
weight.masked_fill_(mask, -torch.inf)
weight /= math.sqrt(self.d_head)
weight = F.softmax(weight, dim=-1)
output = weight @ v
output = output.transpose(1, 2)
output = output.reshape(input_shape)
output = self.out_proj(output)
return output
class VAE_AttentionBlock(nn.Module):
def __init__(self, channels):
super().__init__()
self.groupnorm = nn.GroupNorm(32, channels)
self.attention = SelfAttention(1, channels)
def forward(self, x):
residue = x
x = self.groupnorm(x)
n, c, h, w = x.shape
x = x.view((n, c, h * w)).transpose(1, 2)
x = self.attention(x)
x = x.transpose(1, 2).reshape((n, c, h, w))
x += residue
return x
class VAE_ResidualBlock(nn.Module):
def __init__(self, in_channels, out_channels):
super().__init__()
self.groupnorm_1 = nn.GroupNorm(32, in_channels)
self.conv_1 = nn.Conv2d(in_channels, out_channels, kernel_size=3, padding=1)
self.groupnorm_2 = nn.GroupNorm(32, out_channels)
self.conv_2 = nn.Conv2d(out_channels, out_channels, kernel_size=3, padding=1)
if in_channels == out_channels:
self.residual_layer = nn.Identity()
else:
self.residual_layer = nn.Conv2d(in_channels, out_channels, kernel_size=1, padding=0)
def forward(self, x):
residue = x
x = self.groupnorm_1(x)
x = F.silu(x)
x = self.conv_1(x)
x = self.groupnorm_2(x)
x = F.silu(x)
x = self.conv_2(x)
return x + self.residual_layer(residue)
class VAE_Encoder(nn.Module):
def __init__(self, in_channels=3, base_channels=128, channel_multipliers=[1, 2, 4, 4],
num_residual_blocks_per_level=[2, 2, 2, 2], z_channels=4):
super().__init__()
self.channel_multipliers = channel_multipliers
self.conv_in = nn.Conv2d(in_channels, base_channels, kernel_size=3, padding=1)
layers = []
current_channels = base_channels
for i, multiplier in enumerate(self.channel_multipliers):
out_channels = base_channels * multiplier
for _ in range(num_residual_blocks_per_level[i]):
layers.append(VAE_ResidualBlock(current_channels, out_channels))
current_channels = out_channels
if i < len(self.channel_multipliers) - 1:
layers.append(nn.Conv2d(current_channels, current_channels, kernel_size=3, stride=2, padding=1))
layers.extend([
VAE_ResidualBlock(current_channels, current_channels),
VAE_AttentionBlock(current_channels),
VAE_ResidualBlock(current_channels, current_channels),
nn.GroupNorm(32, current_channels),
nn.SiLU(),
nn.Conv2d(current_channels, 2 * z_channels, kernel_size=3, padding=1)
])
self.model = nn.ModuleList(layers)
def forward(self, x):
x = self.conv_in(x)
for module in self.model:
x = module(x)
mean, log_variance = torch.chunk(x, 2, dim=1)
return mean, log_variance
class VAE_Decoder(nn.Module):
def __init__(self, out_channels=3, base_channels=128, channel_multipliers=[1, 2, 4, 4],
num_residual_blocks_per_level=[2, 2, 2, 2], z_channels=4):
super().__init__()
current_channels = base_channels * channel_multipliers[-1]
self.conv_in = nn.Conv2d(z_channels, current_channels, kernel_size=3, padding=1)
layers = [
VAE_ResidualBlock(current_channels, current_channels),
VAE_AttentionBlock(current_channels),
VAE_ResidualBlock(current_channels, current_channels)
]
for i in reversed(range(len(channel_multipliers))):
out_channels_level = base_channels * channel_multipliers[i]
for _ in range(num_residual_blocks_per_level[i] + 1):
layers.append(VAE_ResidualBlock(current_channels, out_channels_level))
current_channels = out_channels_level
if i > 0:
layers.extend([
nn.Upsample(scale_factor=2, mode='bicubic'),
nn.Conv2d(current_channels, current_channels, kernel_size=3, padding=1)
])
layers.extend([
nn.GroupNorm(32, current_channels),
nn.SiLU(),
nn.Conv2d(current_channels, out_channels, kernel_size=3, padding=1),
nn.Tanh()
])
self.model = nn.ModuleList(layers)
def forward(self, x):
x = self.conv_in(x)
for module in self.model:
x = module(x)
return x
class VAE(nn.Module):
def __init__(self, encoder_config, decoder_config):
super().__init__()
self.encoder = VAE_Encoder(**encoder_config)
self.decoder = VAE_Decoder(**decoder_config)
def reparameterize(self, mu, logvar):
logvar = torch.clamp(logvar, min=-30.0, max=20.0)
std = torch.exp(0.5 * logvar)
eps = torch.randn_like(std)
return mu + eps * std
def forward(self, x):
mu, logvar = self.encoder(x)
z = self.reparameterize(mu, logvar)
reconstructed_x = self.decoder(z)
return reconstructed_x, mu, logvar, z
|