File size: 8,706 Bytes
95f0e22 |
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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 |
import torch
import torch.nn as nn
import torch.nn.functional as F
from .uncertainty import BlockUncertaintyTracker
class ResidualAttentionBlock(nn.Module):
"""Residual attention block for capturing spatial dependencies"""
def __init__(self, in_channels):
super().__init__()
# Trunk branch
self.trunk = nn.Sequential(
nn.ReflectionPad2d(1),
nn.Conv2d(in_channels, in_channels, kernel_size=3, padding=0),
nn.BatchNorm2d(in_channels),
nn.SiLU(),
nn.ReflectionPad2d(1),
nn.Conv2d(in_channels, in_channels, kernel_size=3, padding=0),
nn.BatchNorm2d(in_channels)
)
# Mask branch for attention
self.mask = nn.Sequential(
nn.AdaptiveAvgPool2d(1),
nn.Conv2d(in_channels, in_channels, kernel_size=1),
nn.SiLU(),
nn.Conv2d(in_channels, in_channels, kernel_size=1),
nn.Sigmoid()
)
def forward(self, x):
# Trunk branch
trunk_output = self.trunk(x)
# Mask branch for attention weights
attention = self.mask(x)
# Apply attention and residual connection
out = x + attention * trunk_output
return F.silu(out)
class VectorQuantizer(nn.Module):
"""Vector quantizer for discrete latent representation"""
def __init__(self, n_embeddings=512, embedding_dim=256, beta=0.25):
super().__init__()
self.n_embeddings = n_embeddings
self.embedding_dim = embedding_dim
self.beta = beta
# Initialize embeddings
self.embeddings = nn.Parameter(torch.randn(n_embeddings, embedding_dim))
nn.init.uniform_(self.embeddings, -1.0 / n_embeddings, 1.0 / n_embeddings)
# Usage tracking
self.register_buffer('usage', torch.zeros(n_embeddings))
def forward(self, z):
# Reshape input for quantization
z_flattened = z.reshape(-1, self.embedding_dim)
# Calculate distances to embedding vectors
distances = torch.sum(z_flattened**2, dim=1, keepdim=True) + \
torch.sum(self.embeddings**2, dim=1) - \
2 * torch.matmul(z_flattened, self.embeddings.t())
# Find nearest embedding for each input vector
encoding_indices = torch.argmin(distances, dim=1)
# Update usage statistics
if self.training:
with torch.no_grad():
usage = torch.zeros_like(self.usage)
usage.scatter_add_(0, encoding_indices, torch.ones_like(encoding_indices, dtype=torch.float))
self.usage.mul_(0.99).add_(usage, alpha=0.01)
# Get quantized vectors
z_q = self.embeddings[encoding_indices].reshape(z.shape)
# Calculate loss terms
commitment_loss = F.mse_loss(z_q.detach(), z)
codebook_loss = F.mse_loss(z_q, z.detach())
# Combine losses
loss = codebook_loss + self.beta * commitment_loss
# Straight-through estimator
z_q = z + (z_q - z).detach()
if self.training:
return z_q, loss
else:
return z_q
class Encoder(nn.Module):
"""Encoder for VQ-VAE model"""
def __init__(self, in_channels=1, hidden_dims=[32, 64, 128, 256], embedding_dim=256):
super().__init__()
# Initial conv layer
layers = [
nn.Conv2d(in_channels, hidden_dims[0], kernel_size=3, stride=1, padding=1),
nn.BatchNorm2d(hidden_dims[0]),
nn.SiLU()
]
# Hidden layers with downsampling
for i in range(len(hidden_dims) - 1):
layers.extend([
nn.Conv2d(hidden_dims[i], hidden_dims[i+1], kernel_size=4, stride=2, padding=1),
nn.BatchNorm2d(hidden_dims[i+1]),
nn.SiLU()
])
# Residual attention blocks
for _ in range(2):
layers.append(ResidualAttentionBlock(hidden_dims[-1]))
# Final projection to embedding dimension
layers.extend([
nn.Conv2d(hidden_dims[-1], embedding_dim, kernel_size=1),
nn.BatchNorm2d(embedding_dim)
])
self.encoder = nn.Sequential(*layers)
def forward(self, x):
return self.encoder(x)
class Decoder(nn.Module):
"""Decoder for VQ-VAE model"""
def __init__(self, embedding_dim=256, hidden_dims=[256, 128, 64, 32], out_channels=1):
super().__init__()
# Reverse hidden dims for decoder
hidden_dims = hidden_dims[::-1]
# Initial processing
layers = [
nn.Conv2d(embedding_dim, hidden_dims[0], kernel_size=3, stride=1, padding=1),
nn.BatchNorm2d(hidden_dims[0]),
nn.SiLU()
]
# Residual attention blocks
for _ in range(2):
layers.append(ResidualAttentionBlock(hidden_dims[0]))
# Upsampling blocks
for i in range(len(hidden_dims) - 1):
layers.extend([
nn.ConvTranspose2d(hidden_dims[i], hidden_dims[i+1],
kernel_size=4, stride=2, padding=1),
nn.BatchNorm2d(hidden_dims[i+1]),
nn.SiLU()
])
# Final output layer
layers.append(
nn.Conv2d(hidden_dims[-1], out_channels, kernel_size=3, padding=1)
)
layers.append(nn.Sigmoid())
self.decoder = nn.Sequential(*layers)
def forward(self, x):
return self.decoder(x)
class VQVAE(nn.Module):
"""
Vector Quantized Variational Autoencoder with uncertainty awareness
for bathymetry super-resolution
"""
def __init__(self, in_channels=1, hidden_dims=[32, 64, 128, 256],
num_embeddings=512, embedding_dim=256, block_size=4, alpha=0.1):
super().__init__()
# Initialize block-wise uncertainty tracking
self.uncertainty_tracker = BlockUncertaintyTracker(
block_size=block_size,
alpha=alpha,
decay=0.99,
eps=1e-5
)
# Main model components
self.encoder = Encoder(
in_channels=in_channels,
hidden_dims=hidden_dims,
embedding_dim=embedding_dim
)
self.vq = VectorQuantizer(
n_embeddings=num_embeddings,
embedding_dim=embedding_dim,
beta=0.25
)
self.decoder = Decoder(
embedding_dim=embedding_dim,
hidden_dims=hidden_dims,
out_channels=in_channels
)
def forward(self, x):
"""Forward pass through the model"""
# Encode
z = self.encoder(x)
# Vector quantization
if self.training:
z_q, vq_loss = self.vq(z)
# Decode
reconstruction = self.decoder(z_q)
return reconstruction, vq_loss
else:
z_q = self.vq(z)
# Decode
reconstruction = self.decoder(z_q)
return reconstruction
def train_forward(self, x, y):
"""Training forward pass with uncertainty tracking"""
# Get reconstruction and VQ loss
reconstruction, vq_loss = self.forward(x)
# Calculate reconstruction error
error = torch.abs(reconstruction - y)
# Update uncertainty tracker
self.uncertainty_tracker.update(error)
# Get uncertainty map for loss weighting
uncertainty_map = self.uncertainty_tracker.get_uncertainty(error)
return reconstruction, vq_loss, uncertainty_map
def predict_with_uncertainty(self, x, confidence_level=0.95):
"""
Forward pass with calibrated uncertainty bounds
Args:
x: Input tensor
confidence_level: Confidence level for bounds (default: 0.95)
Returns:
tuple: (reconstruction, lower_bounds, upper_bounds)
"""
self.eval()
with torch.no_grad():
# Get reconstruction
reconstruction = self.forward(x)
# Get calibrated uncertainty bounds
lower_bounds, upper_bounds = self.uncertainty_tracker.get_bounds(
reconstruction, confidence_level
)
return reconstruction, lower_bounds, upper_bounds |