File size: 9,205 Bytes
b90fb2e | 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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 | """
Pixel Decoder: ViT-MAE style decoder following RAE architecture.
Takes 576×embed_dim ViT features and reconstructs 384×384×3 images.
Architecture: ViT-L decoder (24 layers, hidden=1024, heads=16, intermediate=4096).
"""
import math
import numpy as np
import torch
import torch.nn as nn
import torch.nn.functional as F
# ─── Sincos Positional Embeddings ───────────────────────────────────────────
def get_2d_sincos_pos_embed(embed_dim, grid_size, add_cls_token=False):
grid_h = np.arange(grid_size, dtype=np.float32)
grid_w = np.arange(grid_size, dtype=np.float32)
grid = np.meshgrid(grid_w, grid_h)
grid = np.stack(grid, axis=0).reshape([2, 1, grid_size, grid_size])
emb_h = get_1d_sincos_pos_embed(embed_dim // 2, grid[0].reshape(-1))
emb_w = get_1d_sincos_pos_embed(embed_dim // 2, grid[1].reshape(-1))
emb = np.concatenate([emb_h, emb_w], axis=1)
if add_cls_token:
emb = np.concatenate([np.zeros([1, embed_dim]), emb], axis=0)
return emb
def get_1d_sincos_pos_embed(embed_dim, pos):
omega = np.arange(embed_dim // 2, dtype=float)
omega /= embed_dim / 2.0
omega = 1.0 / 10000**omega
pos = pos.reshape(-1)
out = np.einsum("m,d->md", pos, omega)
return np.concatenate([np.sin(out), np.cos(out)], axis=1)
# ─── Transformer Components ────────────────────────────────────────────────
class MAESelfAttention(nn.Module):
def __init__(self, hidden_size, num_heads, qkv_bias=True, attn_drop=0.0, proj_drop=0.0):
super().__init__()
self.num_heads = num_heads
self.head_dim = hidden_size // num_heads
self.query = nn.Linear(hidden_size, hidden_size, bias=qkv_bias)
self.key = nn.Linear(hidden_size, hidden_size, bias=qkv_bias)
self.value = nn.Linear(hidden_size, hidden_size, bias=qkv_bias)
self.out_proj = nn.Linear(hidden_size, hidden_size)
self.attn_drop = attn_drop
def forward(self, x):
B, N, C = x.shape
q = self.query(x).reshape(B, N, self.num_heads, self.head_dim).permute(0, 2, 1, 3)
k = self.key(x).reshape(B, N, self.num_heads, self.head_dim).permute(0, 2, 1, 3)
v = self.value(x).reshape(B, N, self.num_heads, self.head_dim).permute(0, 2, 1, 3)
x = F.scaled_dot_product_attention(q, k, v, dropout_p=self.attn_drop if self.training else 0.0)
x = x.permute(0, 2, 1, 3).reshape(B, N, C)
return self.out_proj(x)
class MAEBlock(nn.Module):
"""Standard ViT block: pre-norm self-attention + pre-norm FFN."""
def __init__(self, hidden_size, num_heads, intermediate_size, hidden_act="gelu",
qkv_bias=True, layer_norm_eps=1e-6):
super().__init__()
self.layernorm_before = nn.LayerNorm(hidden_size, eps=layer_norm_eps)
self.attention = MAESelfAttention(hidden_size, num_heads, qkv_bias=qkv_bias)
self.layernorm_after = nn.LayerNorm(hidden_size, eps=layer_norm_eps)
self.intermediate = nn.Linear(hidden_size, intermediate_size)
self.output_proj = nn.Linear(intermediate_size, hidden_size)
self.act_fn = nn.GELU()
def forward(self, x):
# Self-attention with residual
x = x + self.attention(self.layernorm_before(x))
# FFN with residual
h = self.layernorm_after(x)
h = self.act_fn(self.intermediate(h))
x = x + self.output_proj(h)
return x
# ─── Main Pixel Decoder ────────────────────────────────────────────────────
class PixelDecoderMAE(nn.Module):
"""
ViT-MAE style pixel decoder following RAE.
Input: [B, 576, input_dim] ViT features (or FAE-reconstructed features)
Output: [B, 3, 384, 384] reconstructed images
Architecture (ViT-L):
- Linear projection: input_dim → decoder_hidden_size
- Trainable CLS token + sincos positional embeddings
- 24 Transformer blocks
- LayerNorm + linear head → patch_size² × 3 per token
- Unpatchify → full image
"""
def __init__(self, input_dim=1152, decoder_hidden_size=1024,
decoder_num_layers=24, decoder_num_heads=16,
decoder_intermediate_size=4096, patch_size=16,
img_size=384, num_channels=3, layer_norm_eps=1e-6):
super().__init__()
self.img_size = img_size
self.patch_size = patch_size
self.num_channels = num_channels
self.grid_size = img_size // patch_size # 24
self.num_patches = self.grid_size ** 2 # 576
# Project encoder features to decoder dimension + normalize
self.decoder_embed = nn.Linear(input_dim, decoder_hidden_size)
self.embed_norm = nn.LayerNorm(decoder_hidden_size, eps=layer_norm_eps)
# Trainable CLS token
self.cls_token = nn.Parameter(torch.zeros(1, 1, decoder_hidden_size))
# Fixed sincos positional embeddings (576 patches + 1 CLS)
pos_embed = get_2d_sincos_pos_embed(decoder_hidden_size, self.grid_size, add_cls_token=True)
self.decoder_pos_embed = nn.Parameter(
torch.from_numpy(pos_embed).float().unsqueeze(0),
requires_grad=False
)
# Transformer decoder blocks
self.decoder_layers = nn.ModuleList([
MAEBlock(
hidden_size=decoder_hidden_size,
num_heads=decoder_num_heads,
intermediate_size=decoder_intermediate_size,
layer_norm_eps=layer_norm_eps,
)
for _ in range(decoder_num_layers)
])
self.decoder_norm = nn.LayerNorm(decoder_hidden_size, eps=layer_norm_eps)
# Prediction head: project to pixel patches
self.decoder_pred = nn.Linear(
decoder_hidden_size, patch_size ** 2 * num_channels
)
self._init_weights()
def _init_weights(self):
nn.init.normal_(self.cls_token, std=0.02)
# Initialize decoder_embed like a linear layer
nn.init.xavier_uniform_(self.decoder_embed.weight)
if self.decoder_embed.bias is not None:
nn.init.zeros_(self.decoder_embed.bias)
# Initialize decoder_pred
nn.init.xavier_uniform_(self.decoder_pred.weight)
if self.decoder_pred.bias is not None:
nn.init.zeros_(self.decoder_pred.bias)
def unpatchify(self, x):
"""
x: [B, num_patches, patch_size²×3]
Returns: [B, 3, H, W]
"""
p = self.patch_size
h = w = self.grid_size
c = self.num_channels
x = x.reshape(-1, h, w, p, p, c)
x = torch.einsum("nhwpqc->nchpwq", x)
return x.reshape(-1, c, h * p, w * p)
def forward(self, features, noise_tau=0.0):
"""
Args:
features: [B, 576, input_dim] ViT features
noise_tau: max noise level applied AFTER normalization (where std≈1)
Returns:
images: [B, 3, 384, 384] reconstructed images in [-1, 1]
"""
# Project to decoder dimension and normalize
x = self.embed_norm(self.decoder_embed(features)) # [B, 576, decoder_hidden]
# Add noise after normalization (features now have std≈1, so tau=0.8 is meaningful)
if noise_tau > 0 and self.training:
noise_sigma = noise_tau * torch.rand(
(x.size(0),) + (1,) * (len(x.shape) - 1), device=x.device
)
x = x + noise_sigma * torch.randn_like(x)
# Prepend CLS token
cls_tokens = self.cls_token.expand(x.shape[0], -1, -1)
x = torch.cat([cls_tokens, x], dim=1) # [B, 577, decoder_hidden]
# Add positional embeddings
x = x + self.decoder_pos_embed
# Transformer blocks
for layer in self.decoder_layers:
x = layer(x)
x = self.decoder_norm(x)
# Predict pixel patches (remove CLS token)
x = self.decoder_pred(x[:, 1:, :]) # [B, 576, patch_size²×3]
# Unpatchify to full image
img = self.unpatchify(x) # [B, 3, 384, 384]
return img
class PatchGANDiscriminator(nn.Module):
"""PatchGAN discriminator for adversarial loss."""
def __init__(self, in_channels=3, ndf=64):
super().__init__()
self.model = nn.Sequential(
nn.Conv2d(in_channels, ndf, 4, stride=2, padding=1),
nn.LeakyReLU(0.2, inplace=True),
nn.Conv2d(ndf, ndf * 2, 4, stride=2, padding=1),
nn.InstanceNorm2d(ndf * 2),
nn.LeakyReLU(0.2, inplace=True),
nn.Conv2d(ndf * 2, ndf * 4, 4, stride=2, padding=1),
nn.InstanceNorm2d(ndf * 4),
nn.LeakyReLU(0.2, inplace=True),
nn.Conv2d(ndf * 4, ndf * 8, 4, stride=1, padding=1),
nn.InstanceNorm2d(ndf * 8),
nn.LeakyReLU(0.2, inplace=True),
nn.Conv2d(ndf * 8, 1, 4, stride=1, padding=1),
)
def forward(self, x):
return self.model(x)
|