Spaces:
Running
Running
File size: 11,012 Bytes
776877d |
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 |
import torch
import torch.nn as nn
import torch.nn.functional as F
import numpy as np
class GatingNetwork(nn.Module):
def __init__(self, N, M, Experts, dtype=torch.float32):
super().__init__()
self.conv = nn.Conv1d(N, N, kernel_size=2, padding=0, bias=True, dtype=dtype)
self.softmax_temp1 = nn.Parameter(torch.tensor([0.1], dtype=dtype))
self.D = nn.Parameter(torch.zeros(N, M, dtype=dtype))
self.D.data[:, :N] = torch.eye(N, dtype=dtype)
self.mlp_layer1 = nn.Linear(M + N, Experts, dtype=dtype)
self.mlp_layer2 = nn.Linear(Experts, Experts, dtype=dtype)
self.softmax_temp2 = nn.Parameter(torch.tensor([0.1], dtype=dtype))
self.sigma = nn.Parameter(torch.ones(N, dtype=dtype) * 0.05, requires_grad=True)
def forward(self, context, z, precomputed_cnn=None):
# context: (seq_length, batch_size, N)
# z: (M, batch_size)
# precomputed_cnn: Optional precomputed CNN features for inference (seq_length-1, batch_size, N)
seq_length, batch_size, N = context.shape
M = z.shape[0]
# Compute attention weights
z_obs = self.D @ z.detach()
z_current = z_obs + self.sigma.unsqueeze(1) * torch.randn(N, batch_size, dtype=z.dtype, device=z.device)
z_current_t = z_current.transpose(0, 1)
context_frames = context[:-1]
distances = torch.sum(torch.abs(context_frames - z_current_t.unsqueeze(0)), dim=2)
attention_weights = F.softmax(-distances / torch.abs(self.softmax_temp1[0]), dim=0)
# Process context with convolution
# Use precomputed CNN features if provided, otherwise compute them
if precomputed_cnn is not None:
encoded = precomputed_cnn
else:
context_for_conv = context.permute(1, 2, 0)
encoded = self.conv(context_for_conv)
encoded = encoded.permute(2, 0, 1)
# Build weighted embedding
weighted_encoded = encoded * attention_weights.unsqueeze(2)
embedding = torch.sum(weighted_encoded, dim=0)
embedding = embedding.transpose(0, 1)
# Predict expert weights
combined = torch.cat([embedding, z], dim=0)
combined_t = combined.transpose(0, 1)
mlp_output = self.mlp_layer2(F.relu(self.mlp_layer1(combined_t)))
w_exp = F.softmax(-mlp_output.transpose(0, 1) / torch.abs(self.softmax_temp2[0]), dim=0)
return w_exp
def gaussian_init(self, M, N, dtype=torch.float32):
return torch.randn(M, N, dtype=dtype) * 0.01
class ExpertNetwork(nn.Module):
"""Base class for different expert architectures."""
def __init__(self, M, P=0, probabilistic=False, dtype=torch.float32):
super().__init__()
self.M = M
self.P = P
self.probabilistic = probabilistic
self.dtype = dtype
# Parameter for probabilistic experts
if probabilistic:
self.sigma = nn.Parameter(torch.ones(1, dtype=dtype) * 0.05, requires_grad=True)
def forward(self, z):
raise NotImplementedError("Subclasses must implement forward method")
def add_noise(self, z):
"""Add stochasticity to the latent state if in probabilistic mode.
Args:
z: Input tensor
"""
if self.probabilistic:
batch_size = z.shape[1]
noise = torch.randn(self.M, batch_size, dtype=z.dtype, device=z.device)
return z + self.sigma * noise
return z
def gaussian_init(self, M, N):
return torch.randn(M, N, dtype=self.dtype) * 0.01
def normalized_positive_definite(self, M):
R = np.random.randn(M, M).astype(np.float32)
K = R.T @ R / M + np.eye(M)
lambd = np.max(np.abs(np.linalg.eigvals(K)))
return K / lambd
class AlmostLinearRNN(ExpertNetwork):
"""Almost linear RNN expert architecture."""
def __init__(self, M, P, probabilistic=False, dtype=torch.float32):
super().__init__(M, P, probabilistic, dtype=dtype)
self.A, self.W, self.h = self.initialize_A_W_h(M)
def forward(self, z):
# z: (M, batch_size)
# Split z into regular and ReLU parts
z1 = z[:-self.P, :]
z2 = F.relu(z[-self.P:, :])
zcat = torch.cat([z1, z2], dim=0)
output = self.A.unsqueeze(-1) * z + self.W @ zcat + self.h.unsqueeze(-1)
# Add stochasticity if probabilistic
if self.probabilistic:
output = self.add_noise(output)
return output
def initialize_A_W_h(self, M):
A = torch.nn.Parameter(torch.diag(torch.tensor(self.normalized_positive_definite(M), dtype=self.dtype)))
W = torch.nn.Parameter(self.gaussian_init(M, M))
h = torch.nn.Parameter(torch.zeros(M, dtype=self.dtype))
return A, W, h
class ClippedShallowPLRNN(ExpertNetwork):
"""Clipped shallow PLRNN expert architecture."""
def __init__(self, M, hidden_dim=50, probabilistic=False, dtype=torch.float32):
super().__init__(M, hidden_dim, probabilistic, dtype=dtype)
self.A = torch.nn.Parameter(torch.diag(torch.tensor(self.normalized_positive_definite(M), dtype=self.dtype)))
self.W1 = torch.nn.Parameter(self.gaussian_init(M, hidden_dim))
self.W2 = torch.nn.Parameter(self.gaussian_init(hidden_dim, M))
self.h1 = torch.nn.Parameter(torch.zeros(M, dtype=self.dtype))
self.h2 = torch.nn.Parameter(torch.zeros(hidden_dim, dtype=self.dtype))
def forward(self, z):
# z: (M, batch_size)
W2z = self.W2 @ z
output = (self.A.unsqueeze(-1) * z +
self.W1 @ (F.relu(W2z + self.h2.unsqueeze(-1)) - F.relu(W2z)) +
self.h1.unsqueeze(-1))
# Add stochasticity if probabilistic
if self.probabilistic:
output = self.add_noise(output)
return output
class DynaMix(nn.Module):
def __init__(self, M, N, Experts, P=2, hidden_dim=50, expert_type="almost_linear_rnn",
probabilistic_expert=False, dtype=torch.float32):
"""
Initialize a DynaMix model.
Args:
M: Dimension of latent state
N: Dimension of observation space
Experts: Number of experts
P: Number of ReLU dimensions
hidden_dim: Hidden dimension for clipped shallow PLRNN
expert_type: Type of expert to use ("almost_linear_rnn" or "clipped_shallow_plrnn")
probabilistic_expert: Whether to use probabilistic experts
dtype: Data type for model parameters (default: torch.float32)
"""
super().__init__()
self.expert_type = expert_type
self.probabilistic_expert = probabilistic_expert
self.experts = nn.ModuleList()
self.dtype = dtype
for _ in range(Experts):
if expert_type == "almost_linear_rnn":
self.experts.append(AlmostLinearRNN(M, P, probabilistic=probabilistic_expert, dtype=dtype))
elif expert_type == "clipped_shallow_plrnn":
self.experts.append(ClippedShallowPLRNN(M, hidden_dim, probabilistic=probabilistic_expert, dtype=dtype))
else:
raise ValueError(f"Unknown expert type: {expert_type}")
self.gating_network = GatingNetwork(N, M, Experts, dtype=dtype)
self.B = nn.Parameter(self.uniform_init((N, M), dtype=dtype))
self.N = N
self.Experts = Experts
self.P = P
self.hidden_dim = hidden_dim
self.M = M
def step(self, z, context, precomputed_cnn=None):
# z: (M, batch_size)
# context: (seq_length, batch_size, N)
# precomputed_cnn: Optional precomputed CNN features
# Compute expert weights
w_exp = self.gating_network(context, z, precomputed_cnn=precomputed_cnn) # (Experts, batch_size)
results = []
# Compute expert outputs
for i in range(self.Experts):
expert_output = self.experts[i](z)
results.append(expert_output * w_exp[i, :].unsqueeze(0))
# Combine expert outputs
return torch.sum(torch.stack(results, dim=0), dim=0)
def forward(self, z, context, precomputed_cnn=None):
"""
Forward pass through the DynaMix model.
Args:
z: Latent state of shape (M, batch_size)
context: Context data of shape (seq_length, batch_size, N)
precomputed_cnn: Optional precomputed CNN features to avoid redundant computation for inference
Shape should be (seq_length-1, batch_size, N)
Returns:
Updated latent state
"""
return self.step(z, context, precomputed_cnn=precomputed_cnn)
def precompute_cnn(self, context):
"""
Precompute CNN features for more efficient inference.
Args:
context: Context data of shape (seq_length, batch_size, N)
Returns:
Precomputed CNN features of shape (seq_length-1, batch_size, N)
"""
# Process context with convolution
context_for_conv = context.permute(1, 2, 0)
encoded = self.gating_network.conv(context_for_conv)
return encoded.permute(2, 0, 1)
def uniform_init(self, shape, dtype=torch.float32):
din = shape[-1]
r = 1 / np.sqrt(din)
return (torch.rand(shape, dtype=dtype) * 2 - 1) * r
def gaussian_init(self, M, N):
return torch.randn(M, N, dtype=self.dtype) * 0.01
def print_model_parameters(model):
"""Print simplified breakdown of model parameters by component."""
total_params = sum(p.numel() for p in model.parameters())
print("\n" + "-"*60)
print("Model Parameter Summary:")
print(f" Architecture: DynaMix with {model.expert_type} experts")
if model.expert_type == "almost_linear_rnn":
print(f" Dimensions: M={model.M}, N={model.N}, Experts={model.Experts}, P={model.P}")
else:
print(f" Dimensions: M={model.M}, N={model.N}, Experts={model.Experts}, Hidden dim={model.hidden_dim}")
print(f" Probabilistic experts: {model.probabilistic_expert}")
# Count parameters
gating_params = sum(p.numel() for p in model.gating_network.parameters())
expert_params = sum(p.numel() for expert in model.experts for p in expert.parameters())
b_params = model.B.numel()
# Print parameter counts
print(f"\nParameter counts:")
print(f" Gating Network: {gating_params:,} ({gating_params/total_params:.1%})")
print(f" Experts: {expert_params:,} ({expert_params/total_params:.1%})")
print(f" Observation matrix: {b_params:,} ({b_params/total_params:.1%})")
print(f" Total: {total_params:,} parameters")
print("-"*60) |