Spaces:
Runtime error
Runtime error
File size: 11,276 Bytes
c3ec853 | 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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 | """
MLP-based decoder for reconstructing point clouds from latent codes.
This module provides flexible MLP architectures for decoding latent representations
into point clouds, commonly used in generative models like VAE and cVAE.
"""
import torch
import torch.nn as nn
import torch.nn.functional as F
from typing import List, Optional
class PointCloudDecoder(nn.Module):
"""
MLP-based decoder for reconstructing point clouds from latent representations.
This decoder uses a series of fully connected layers with optional dropout
and normalization to transform a latent code into point cloud coordinates.
Args:
latent_dim: Dimensionality of input latent code
num_points: Number of output points in the point cloud
point_dim: Dimensionality of each point (default: 3 for XYZ coordinates)
hidden_dims: List of hidden layer dimensions (default: [1024, 2048])
dropout_rate: Dropout probability (default: 0.3)
use_batch_norm: Whether to use batch normalization (default: False)
activation: Activation function to use (default: 'relu')
Input:
Latent code of shape (B, latent_dim)
Output:
Point cloud of shape (B, num_points, point_dim)
"""
def __init__(
self,
latent_dim: int,
num_points: int,
point_dim: int = 3,
hidden_dims: Optional[List[int]] = None,
dropout_rate: float = 0.3,
use_batch_norm: bool = False,
activation: str = 'relu'
):
super(PointCloudDecoder, self).__init__()
self.latent_dim = latent_dim
self.num_points = num_points
self.point_dim = point_dim
self.dropout_rate = dropout_rate
self.use_batch_norm = use_batch_norm
# Default hidden dimensions
if hidden_dims is None:
hidden_dims = [1024, 2048]
self.hidden_dims = hidden_dims
# Select activation function
if activation == 'relu':
self.activation = nn.ReLU()
elif activation == 'leaky_relu':
self.activation = nn.LeakyReLU(0.2)
elif activation == 'elu':
self.activation = nn.ELU()
elif activation == 'gelu':
self.activation = nn.GELU()
else:
raise ValueError(f"Unsupported activation: {activation}")
# Build network layers
self.layers = nn.ModuleList()
self.batch_norms = nn.ModuleList() if use_batch_norm else None
self.dropouts = nn.ModuleList()
# Input layer
prev_dim = latent_dim
for hidden_dim in hidden_dims:
self.layers.append(nn.Linear(prev_dim, hidden_dim))
if use_batch_norm:
self.batch_norms.append(nn.BatchNorm1d(hidden_dim))
self.dropouts.append(nn.Dropout(dropout_rate))
prev_dim = hidden_dim
# Output layer
output_dim = num_points * point_dim
self.output_layer = nn.Linear(prev_dim, output_dim)
def forward(self, z: torch.Tensor) -> torch.Tensor:
"""
Decode latent code into point cloud.
Args:
z: Latent code of shape (B, latent_dim)
Returns:
Reconstructed point cloud of shape (B, num_points, point_dim)
"""
x = z
# Process through hidden layers
for i, layer in enumerate(self.layers):
x = layer(x)
if self.use_batch_norm:
x = self.batch_norms[i](x)
x = self.activation(x)
x = self.dropouts[i](x)
# Output layer
x = self.output_layer(x)
# Reshape to point cloud
x = x.view(-1, self.num_points, self.point_dim)
return x
def get_output_shape(self) -> tuple:
"""
Get the output point cloud shape (excluding batch dimension).
Returns:
Tuple of (num_points, point_dim)
"""
return (self.num_points, self.point_dim)
class ResidualMLPDecoder(nn.Module):
"""
MLP decoder with residual connections for improved gradient flow.
This decoder uses residual blocks to help with training deep networks.
Args:
latent_dim: Dimensionality of input latent code
num_points: Number of output points in the point cloud
point_dim: Dimensionality of each point (default: 3 for XYZ coordinates)
hidden_dim: Hidden layer dimension (default: 1024)
num_blocks: Number of residual blocks (default: 3)
dropout_rate: Dropout probability (default: 0.3)
use_batch_norm: Whether to use batch normalization (default: True)
Input:
Latent code of shape (B, latent_dim)
Output:
Point cloud of shape (B, num_points, point_dim)
"""
def __init__(
self,
latent_dim: int,
num_points: int,
point_dim: int = 3,
hidden_dim: int = 1024,
num_blocks: int = 3,
dropout_rate: float = 0.3,
use_batch_norm: bool = True
):
super(ResidualMLPDecoder, self).__init__()
self.latent_dim = latent_dim
self.num_points = num_points
self.point_dim = point_dim
self.hidden_dim = hidden_dim
# Initial projection
self.input_proj = nn.Linear(latent_dim, hidden_dim)
# Residual blocks
self.blocks = nn.ModuleList([
ResidualBlock(hidden_dim, dropout_rate, use_batch_norm)
for _ in range(num_blocks)
])
# Output projection
output_dim = num_points * point_dim
self.output_proj = nn.Linear(hidden_dim, output_dim)
def forward(self, z: torch.Tensor) -> torch.Tensor:
"""
Decode latent code into point cloud.
Args:
z: Latent code of shape (B, latent_dim)
Returns:
Reconstructed point cloud of shape (B, num_points, point_dim)
"""
# Initial projection
x = F.relu(self.input_proj(z))
# Residual blocks
for block in self.blocks:
x = block(x)
# Output projection
x = self.output_proj(x)
x = x.view(-1, self.num_points, self.point_dim)
return x
class ResidualBlock(nn.Module):
"""
Residual block with optional batch normalization and dropout.
Args:
hidden_dim: Dimension of the hidden layer
dropout_rate: Dropout probability
use_batch_norm: Whether to use batch normalization
"""
def __init__(
self,
hidden_dim: int,
dropout_rate: float = 0.3,
use_batch_norm: bool = True
):
super(ResidualBlock, self).__init__()
self.fc1 = nn.Linear(hidden_dim, hidden_dim)
self.fc2 = nn.Linear(hidden_dim, hidden_dim)
self.dropout = nn.Dropout(dropout_rate)
if use_batch_norm:
self.bn1 = nn.BatchNorm1d(hidden_dim)
self.bn2 = nn.BatchNorm1d(hidden_dim)
else:
self.bn1 = None
self.bn2 = None
def forward(self, x: torch.Tensor) -> torch.Tensor:
"""
Forward pass through residual block.
Args:
x: Input tensor of shape (B, hidden_dim)
Returns:
Output tensor of shape (B, hidden_dim)
"""
residual = x
out = self.fc1(x)
if self.bn1 is not None:
out = self.bn1(out)
out = F.relu(out)
out = self.dropout(out)
out = self.fc2(out)
if self.bn2 is not None:
out = self.bn2(out)
out = out + residual
out = F.relu(out)
return out
def create_pointcloud_decoder(
latent_dim: int,
num_points: int = 6890,
point_dim: int = 3,
architecture: str = 'mlp',
**kwargs
) -> nn.Module:
"""
Factory function to create a point cloud decoder with specified architecture.
Args:
latent_dim: Dimensionality of input latent code
num_points: Number of output points (default: 6890 for SMPL mesh)
point_dim: Dimensionality of each point (default: 3 for XYZ)
architecture: Decoder architecture type ('mlp' or 'residual')
**kwargs: Additional arguments passed to the decoder constructor
Returns:
Configured decoder instance
Examples:
>>> # Create a simple MLP decoder
>>> decoder = create_pointcloud_decoder(
... latent_dim=256,
... num_points=6890,
... architecture='mlp',
... hidden_dims=[1024, 2048],
... dropout_rate=0.3
... )
>>> # Create a residual decoder
>>> decoder = create_pointcloud_decoder(
... latent_dim=256,
... num_points=6890,
... architecture='residual',
... hidden_dim=1024,
... num_blocks=3
... )
"""
if architecture == 'mlp':
return PointCloudDecoder(
latent_dim=latent_dim,
num_points=num_points,
point_dim=point_dim,
**kwargs
)
elif architecture == 'residual':
return ResidualMLPDecoder(
latent_dim=latent_dim,
num_points=num_points,
point_dim=point_dim,
**kwargs
)
else:
raise ValueError(f"Unsupported architecture: {architecture}")
if __name__ == '__main__':
print("Testing Point Cloud Decoders...\n")
# Test parameters
batch_size = 32
latent_dim = 256
num_points = 6890 # SMPL mesh vertices
point_dim = 3
# Test standard MLP decoder
print("1. Standard MLP Decoder:")
mlp_decoder = create_pointcloud_decoder(
latent_dim=latent_dim,
num_points=num_points,
point_dim=point_dim,
architecture='mlp',
hidden_dims=[1024, 2048],
dropout_rate=0.3,
use_batch_norm=False
)
z = torch.randn(batch_size, latent_dim)
output = mlp_decoder(z)
print(f" Input shape: {z.shape}")
print(f" Output shape: {output.shape}")
print(f" Output expected shape: {mlp_decoder.get_output_shape()}")
print()
# Test MLP decoder with batch norm
print("2. MLP Decoder with Batch Normalization:")
mlp_bn_decoder = create_pointcloud_decoder(
latent_dim=latent_dim,
num_points=num_points,
architecture='mlp',
use_batch_norm=True,
activation='leaky_relu'
)
output = mlp_bn_decoder(z)
print(f" Output shape: {output.shape}")
print()
# Test residual MLP decoder
print("3. Residual MLP Decoder:")
residual_decoder = create_pointcloud_decoder(
latent_dim=latent_dim,
num_points=num_points,
architecture='residual',
hidden_dim=1024,
num_blocks=3,
dropout_rate=0.3
)
output = residual_decoder(z)
print(f" Input shape: {z.shape}")
print(f" Output shape: {output.shape}")
print()
# Test with different point dimensions
print("4. Decoder with 6D points (XYZ + RGB):")
decoder_6d = create_pointcloud_decoder(
latent_dim=latent_dim,
num_points=1000,
point_dim=6,
architecture='mlp',
hidden_dims=[512, 1024]
)
output = decoder_6d(z)
print(f" Output shape: {output.shape}")
print()
|