Spaces:
Running on Zero
Running on Zero
File size: 8,554 Bytes
632d805 | 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 | """
Prediction heads for the dynamics model: plain MLP, multi-scale, and
residual variants.
"""
import torch
import torch.nn as nn
import torch.nn.functional as F
from typing import List, Optional
class DynamicsPredictor(nn.Module):
"""
MLP predictor that takes state features and query points to predict dynamics.
"""
def __init__(
self,
state_dim: int = 512,
query_dim: int = 3, # 3D position
hidden_dims: List[int] = [256, 128, 64],
output_dim: int = 3, # 3D acceleration/displacement
activation: str = 'relu',
use_batch_norm: bool = False,
dropout: float = 0.1
):
super().__init__()
self.state_dim = state_dim
self.query_dim = query_dim
self.output_dim = output_dim
# Build MLP layers
layers = []
input_dim = state_dim + query_dim
for hidden_dim in hidden_dims:
layers.append(nn.Linear(input_dim, hidden_dim))
if use_batch_norm:
layers.append(nn.BatchNorm1d(hidden_dim))
if activation == 'relu':
layers.append(nn.SiLU())
elif activation == 'gelu':
layers.append(nn.SiLU())
elif activation == 'tanh':
layers.append(nn.Tanh())
else:
raise ValueError(f"Unknown activation: {activation}")
if dropout > 0:
layers.append(nn.Dropout(dropout))
input_dim = hidden_dim
# Output layer
layers.append(nn.Linear(hidden_dims[-1], output_dim))
layers.append(nn.Tanh()) # bounded output
self.output_scale = nn.Parameter(torch.ones(1) * 1.0)
self.mlp = nn.Sequential(*layers)
def forward(
self,
state_features: torch.Tensor,
query_points: torch.Tensor
) -> torch.Tensor:
"""
Predict dynamics for query points given state features.
Args:
state_features: Enhanced state features (B, N_state, D_state) or (B, D_state)
query_points: Query vertex positions (B, N_query, 3)
Returns:
Predicted accelerations/displacements (B, N_query, 3)
"""
B = query_points.shape[0]
N_query = query_points.shape[1]
# If state_features has sequence dimension, pool it
if state_features.dim() == 3:
# Global max pooling over state tokens
state_features = torch.max(state_features, dim=1)[0] # (B, D_state)
# Expand state features to match number of query points
state_features_expanded = state_features.unsqueeze(1).expand(B, N_query, -1)
# Concatenate state features with query points
combined = torch.cat([state_features_expanded, query_points], dim=-1)
# Reshape for batch processing through MLP
combined_flat = combined.view(B * N_query, -1)
# Apply MLP
output_flat = self.mlp(combined_flat) * self.output_scale
# Reshape back
output = output_flat.view(B, N_query, self.output_dim)
return output
class MultiScalePredictor(nn.Module):
"""
Multi-scale predictor that processes features at different resolutions.
"""
def __init__(
self,
state_dim: int = 512,
query_dim: int = 3,
scales: List[int] = [64, 128, 256],
output_dim: int = 3,
activation: str = 'relu',
dropout: float = 0.1
):
super().__init__()
self.scales = scales
# Create predictors for each scale
self.scale_predictors = nn.ModuleList([
DynamicsPredictor(
state_dim=state_dim,
query_dim=query_dim,
hidden_dims=[scale, scale // 2],
output_dim=output_dim,
activation=activation,
dropout=dropout
) for scale in scales
])
# Fusion layer
self.fusion = nn.Linear(output_dim * len(scales), output_dim)
def forward(
self,
state_features: torch.Tensor,
query_points: torch.Tensor
) -> torch.Tensor:
"""
Multi-scale prediction with fusion.
"""
predictions = []
for predictor in self.scale_predictors:
pred = predictor(state_features, query_points)
predictions.append(pred)
# Concatenate all scale predictions
combined = torch.cat(predictions, dim=-1)
# Fuse predictions
output = self.fusion(combined)
return output
class ResidualPredictor(nn.Module):
"""
Predictor with residual connections for better gradient flow.
"""
def __init__(
self,
state_dim: int = 512,
query_dim: int = 3,
hidden_dim: int = 256,
num_layers: int = 3,
output_dim: int = 3,
activation: str = 'relu',
dropout: float = 0.1
):
super().__init__()
self.input_projection = nn.Linear(state_dim + query_dim, hidden_dim)
# Residual blocks
self.res_blocks = nn.ModuleList([
ResidualBlock(hidden_dim, activation, dropout)
for _ in range(num_layers)
])
self.output_projection = nn.Linear(hidden_dim, output_dim)
def forward(
self,
state_features: torch.Tensor,
query_points: torch.Tensor
) -> torch.Tensor:
"""Forward pass with residual connections."""
B = query_points.shape[0]
N_query = query_points.shape[1]
# Pool state features if needed
if state_features.dim() == 3:
state_features = torch.max(state_features, dim=1)[0]
# Expand and concatenate
state_expanded = state_features.unsqueeze(1).expand(B, N_query, -1)
combined = torch.cat([state_expanded, query_points], dim=-1)
# Initial projection
x = self.input_projection(combined)
# Apply residual blocks
for block in self.res_blocks:
x = block(x)
# Output projection
output = self.output_projection(x)
return output
class ResidualBlock(nn.Module):
"""Single residual block."""
def __init__(self, hidden_dim: int, activation: str = 'relu', dropout: float = 0.1):
super().__init__()
self.fc1 = nn.Linear(hidden_dim, hidden_dim)
self.fc2 = nn.Linear(hidden_dim, hidden_dim)
if activation == 'relu':
self.activation = nn.ReLU()
elif activation == 'gelu':
self.activation = nn.GELU()
else:
self.activation = nn.ReLU()
self.dropout = nn.Dropout(dropout)
self.norm1 = nn.LayerNorm(hidden_dim)
self.norm2 = nn.LayerNorm(hidden_dim)
def forward(self, x):
residual = x
x = self.norm1(x)
x = self.fc1(x)
x = self.activation(x)
x = self.dropout(x)
x = self.norm2(x)
x = self.fc2(x)
x = self.dropout(x)
return x + residual
if __name__ == "__main__":
"""Test predictor modules."""
print("Testing DynamicsPredictor...")
predictor = DynamicsPredictor(
state_dim=512,
query_dim=3,
hidden_dims=[256, 128, 64],
output_dim=3
)
# Test inputs
batch_size = 4
state_features = torch.randn(batch_size, 10, 512) # With sequence dimension
query_points = torch.randn(batch_size, 8, 3) # 8 vertices
output = predictor(state_features, query_points)
print(f"Output shape: {output.shape}") # Should be (4, 8, 3)
print("\nTesting MultiScalePredictor...")
multi_predictor = MultiScalePredictor(state_dim=512, query_dim=3)
output_multi = multi_predictor(state_features, query_points)
print(f"Multi-scale output shape: {output_multi.shape}") # (4, 8, 3)
print("\nTesting ResidualPredictor...")
res_predictor = ResidualPredictor(state_dim=512, query_dim=3)
output_res = res_predictor(state_features, query_points)
print(f"Residual output shape: {output_res.shape}") # (4, 8, 3)
print("\nPredictor tests successful!")
|