File size: 16,121 Bytes
97d8aaa |
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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 |
import types
from typing import Optional, List, Union, Callable
import torch
from torch import nn, Tensor
from torch.nn import functional as F
from torchvision.models.mobilenetv2 import MobileNetV2
from torchvision.models.resnet import ResNet
from torchvision.models.efficientnet import EfficientNet
from torchvision.models.vision_transformer import VisionTransformer
def compute_policy_loss(loss_sequence, mask_sequence, rewards):
losses = sum(mask * padded_loss for mask, padded_loss in zip(mask_sequence, loss_sequence))
returns = sum(padded_reward * mask for padded_reward, mask in zip(rewards, mask_sequence))
loss = torch.mean(losses * returns)
return loss
class TPBlock(nn.Module):
def __init__(self, depths: int, in_planes: int, out_planes: int = None, rank=1, shape_dims=3, channel_first=True, dtype=torch.float32) -> None:
super().__init__()
out_planes = in_planes if out_planes is None else out_planes
self.layers = torch.nn.ModuleList([self._make_layer(in_planes, out_planes, rank, shape_dims, channel_first, dtype) for _ in range(depths)])
def forward(self, x: Tensor) -> Tensor:
for layer in self.layers:
x = x + layer(x)
return x
def _make_layer(self, in_planes: int, out_planes: int = None, rank=1, shape_dims=3, channel_first=True, dtype=torch.float32) -> nn.Sequential:
class Permute(nn.Module):
def __init__(self, *dims):
super().__init__()
self.dims = dims
def forward(self, x):
return x.permute(*self.dims)
class RMSNorm(nn.Module):
__constants__ = ["eps"]
eps: float
def __init__(self, hidden_size, eps: float = 1e-6, device=None, dtype=None):
"""
LlamaRMSNorm is equivalent to T5LayerNorm.
"""
factory_kwargs = {"device": device, "dtype": dtype}
super().__init__()
self.eps = eps
self.weight = nn.Parameter(torch.ones(hidden_size, **factory_kwargs))
def forward(self, hidden_states):
input_dtype = hidden_states.dtype
hidden_states = hidden_states.to(torch.float32)
variance = hidden_states.pow(2).mean(dim=1, keepdim=True)
hidden_states = hidden_states * torch.rsqrt(variance + self.eps)
weight = self.weight.view(1, -1, *[1] * (hidden_states.ndim - 2))
return weight * hidden_states.to(input_dtype)
def extra_repr(self):
return f"{self.weight.shape[0]}, eps={self.eps}"
conv_map = {
2: (nn.Conv1d, (0, 2, 1), (0, 2, 1)),
3: (nn.Conv2d, (0, 3, 1, 2), (0, 2, 3, 1)),
4: (nn.Conv3d, (0, 4, 1, 2, 3), (0, 2, 3, 4, 1)),
}
Conv, pre_dims, post_dims = conv_map[shape_dims]
kernel_size, dilation, padding = self.generate_hyperparameters(rank)
pre_permute = nn.Identity() if channel_first else Permute(*pre_dims)
post_permute = nn.Identity() if channel_first else Permute(*post_dims)
conv1 = Conv(in_planes, out_planes, kernel_size, padding=padding, dilation=dilation, bias=False, dtype=dtype, device='cuda')
nn.init.zeros_(conv1.weight)
bn1 = RMSNorm(out_planes, dtype=dtype, device="cuda")
relu = nn.ReLU(inplace=True)
conv2 = Conv(out_planes, in_planes, kernel_size, padding=padding, dilation=dilation, bias=False, dtype=dtype, device='cuda')
nn.init.zeros_(conv2.weight)
bn2 = RMSNorm(in_planes, dtype=dtype, device="cuda")
return torch.nn.Sequential(pre_permute, conv1, bn1, relu, conv2, bn2, relu, post_permute)
@staticmethod
def generate_hyperparameters(rank: int):
"""
Generates kernel size and dilation rate pairs sorted by increasing padded kernel size.
Args:
rank: Number of (kernel_size, dilation) pairs to generate. Must be positive.
Returns:
Tuple[int, int]: A (kernel_size, dilation) tuple where:
- kernel_size: Always odd and >= 1
- dilation: Computed to maintain consistent padded kernel size growth
Note:
Padded kernel size is calculated as:
(kernel_size - 1) * dilation + 1
Pairs are generated first in order of increasing padded kernel size,
then by increasing kernel size for equal padded kernel sizes.
"""
pairs = [(1, 1, 0)] # Start with smallest possible
padded_kernel_size = 3
while len(pairs) < rank:
for kernel_size in range(3, padded_kernel_size + 1, 2):
if (padded_kernel_size - 1) % (kernel_size - 1) == 0:
dilation = (padded_kernel_size - 1) // (kernel_size - 1)
padding = dilation * (kernel_size - 1) // 2
pairs.append((kernel_size, dilation, padding))
if len(pairs) >= rank:
break
# Move to next odd padded kernel size
padded_kernel_size += 2
return pairs[-1]
class ResNetConfig:
@staticmethod
def gen_shared_head(self):
def func(hidden_states):
"""
Args:
hidden_states (Tensor): Hidden States tensor of shape [B, C, H, W].
Returns:
logits (Tensor): Logits tensor of shape [B, C].
"""
x = self.avgpool(hidden_states)
x = torch.flatten(x, 1)
logits = self.fc(x)
return logits
return func
@staticmethod
def gen_logits(self, shared_head):
def func(hidden_states):
"""
Args:
hidden_states (Tensor): Hidden States tensor of shape [B, L, hidden_units].
Returns:
logits_seqence (List[Tensor]): List of Logits tensors.
"""
logits_sequence = [shared_head(hidden_states)]
for layer in self.trp_blocks:
logits_sequence.append(shared_head(layer(hidden_states)))
return logits_sequence
return func
@staticmethod
def gen_mask(label_smoothing=0.0, top_k=1):
def func(logits_sequence, labels):
"""
Args:
logits_sequence (List[Tensor]): List of Logits tensors.
labels (Tensor): Target labels of shape [B] or [B, C].
Returns:
mask_sequence (List[Tensor]): List of Mask tensor.
returns (Tensor): Boolean mask tensor of shape [B*(L-1)].
"""
labels = torch.argmax(labels, dim=1) if label_smoothing > 0.0 else labels
mask_sequence = [torch.ones_like(labels, dtype=torch.float32, device=labels.device)]
for logits in logits_sequence:
with torch.no_grad():
topk_values, topk_indices = torch.topk(logits, top_k, dim=-1)
mask = torch.eq(topk_indices, labels[:, None]).any(dim=-1).to(torch.float32)
mask_sequence.append(mask_sequence[-1] * mask)
return mask_sequence
return func
@staticmethod
def gen_criterion(label_smoothing=0.0):
def func(logits_sequence, labels):
"""
Args:
logits_sequence (List[Tensor]): List of Logits tensor.
labels (Tensor): labels labels of shape [B] or [B, C].
Returns:
loss (Tensor): Scalar tensor representing the loss.
mask (Tensor): Boolean mask tensor of shape [B].
"""
labels = torch.argmax(labels, dim=1) if label_smoothing > 0.0 else labels
loss_sequence = []
for logits in logits_sequence:
loss_sequence.append(F.cross_entropy(logits, labels, reduction="none", label_smoothing=label_smoothing))
return loss_sequence
return func
@staticmethod
def gen_forward(rewards, label_smoothing=0.0, top_k=1):
def func(self, x: Tensor, targets=None) -> Tensor:
x = self.conv1(x)
x = self.bn1(x)
x = self.relu(x)
x = self.maxpool(x)
x = self.layer1(x)
x = self.layer2(x)
x = self.layer3(x)
hidden_states = self.layer4(x)
x = self.avgpool(hidden_states)
x = torch.flatten(x, 1)
logits = self.fc(x)
if self.training:
shared_head = ResNetConfig.gen_shared_head(self)
compute_logits = ResNetConfig.gen_logits(self, shared_head)
compute_mask = ResNetConfig.gen_mask(label_smoothing, top_k)
compute_loss = ResNetConfig.gen_criterion(label_smoothing)
logits_sequence = compute_logits(hidden_states)
mask_sequence = compute_mask(logits_sequence, targets)
loss_sequence = compute_loss(logits_sequence, targets)
loss = compute_policy_loss(loss_sequence, mask_sequence, rewards)
return logits, loss
return logits
return func
class MobileNetV2Config(ResNetConfig):
@staticmethod
def gen_shared_head(self):
def func(hidden_states):
"""
Args:
hidden_states (Tensor): Hidden States tensor of shape [B, C, H, W].
Returns:
logits (Tensor): Logits tensor of shape [B, C].
"""
x = nn.functional.adaptive_avg_pool2d(hidden_states, (1, 1))
x = torch.flatten(x, 1)
logits = self.classifier(x)
return logits
return func
@staticmethod
def gen_forward(rewards, label_smoothing=0.0, top_k=1):
def func(self, x: Tensor, targets=None) -> Tensor:
hidden_states = self.features(x)
# Cannot use "squeeze" as batch-size can be 1
x = nn.functional.adaptive_avg_pool2d(hidden_states, (1, 1))
x = torch.flatten(x, 1)
logits = self.classifier(x)
if self.training:
shared_head = MobileNetV2Config.gen_shared_head(self)
compute_logits = MobileNetV2Config.gen_logits(self, shared_head)
compute_mask = MobileNetV2Config.gen_mask(label_smoothing, top_k)
compute_loss = MobileNetV2Config.gen_criterion(label_smoothing)
logits_sequence = compute_logits(hidden_states)
mask_sequence = compute_mask(logits_sequence, targets)
loss_sequence = compute_loss(logits_sequence, targets)
loss = compute_policy_loss(loss_sequence, mask_sequence, rewards)
return logits, loss
return logits
return func
class EfficientNetConfig(ResNetConfig):
@staticmethod
def gen_shared_head(self):
def func(hidden_states):
"""
Args:
hidden_states (Tensor): Hidden States tensor of shape [B, C, H, W].
Returns:
logits (Tensor): Logits tensor of shape [B, C].
"""
x = self.avgpool(hidden_states)
x = torch.flatten(x, 1)
logits = self.classifier(x)
return logits
return func
@staticmethod
def gen_forward(rewards, label_smoothing=0.0, top_k=1):
def func(self, x: Tensor, targets=None) -> Tensor:
hidden_states = self.features(x)
x = self.avgpool(hidden_states)
x = torch.flatten(x, 1)
logits = self.classifier(x)
if self.training:
shared_head = EfficientNetConfig.gen_shared_head(self)
compute_logits = EfficientNetConfig.gen_logits(self, shared_head)
compute_mask = EfficientNetConfig.gen_mask(label_smoothing, top_k)
compute_loss = EfficientNetConfig.gen_criterion(label_smoothing)
logits_sequence = compute_logits(hidden_states)
mask_sequence = compute_mask(logits_sequence, targets)
loss_sequence = compute_loss(logits_sequence, targets)
loss = compute_policy_loss(loss_sequence, mask_sequence, rewards)
return logits, loss
return logits
return func
class VisionTransformerConfig(ResNetConfig):
@staticmethod
def gen_shared_head(self):
def func(hidden_states):
"""
Args:
hidden_states (Tensor): Hidden States tensor of shape [B, C, H, W].
Returns:
logits (Tensor): Logits tensor of shape [B, C].
"""
x = hidden_states[:, 0]
logits = self.heads(x)
return logits
return func
@staticmethod
def gen_forward(rewards, label_smoothing=0.0, top_k=1):
def func(self, images: Tensor, targets=None):
x = self._process_input(images)
n = x.shape[0]
batch_class_token = self.class_token.expand(n, -1, -1)
x = torch.cat([batch_class_token, x], dim=1)
hidden_states = self.encoder(x)
x = hidden_states[:, 0]
logits = self.heads(x)
if self.training:
shared_head = VisionTransformerConfig.gen_shared_head(self)
compute_logits = VisionTransformerConfig.gen_logits(self, shared_head)
compute_mask = VisionTransformerConfig.gen_mask(label_smoothing, top_k)
compute_loss = VisionTransformerConfig.gen_criterion(label_smoothing)
logits_sequence = compute_logits(hidden_states)
mask_sequence = compute_mask(logits_sequence, targets)
loss_sequence = compute_loss(logits_sequence, targets)
loss = compute_policy_loss(loss_sequence, mask_sequence, rewards)
return logits, loss
return logits
return func
def apply_trp(model, depths: List[int], in_planes: int, out_planes: int, rewards, **kwargs):
if isinstance(model, ResNet):
print("✅ Applying TRP to ResNet for Image Classification...")
model.trp_blocks = torch.nn.ModuleList([TPBlock(depths=d, in_planes=in_planes, out_planes=out_planes, rank=k) for k, d in enumerate(depths)])
model.forward = types.MethodType(ResNetConfig.gen_forward(rewards, label_smoothing=kwargs["label_smoothing"], top_k=1), model)
elif isinstance(model, MobileNetV2):
print("✅ Applying TRP to MobileNetV2 for Image Classification...")
model.trp_blocks = torch.nn.ModuleList([TPBlock(depths=d, in_planes=in_planes, out_planes=out_planes, rank=k) for k, d in enumerate(depths)])
model.forward = types.MethodType(MobileNetV2Config.gen_forward(rewards, label_smoothing=kwargs["label_smoothing"], top_k=1), model)
elif isinstance(model, EfficientNet):
print("✅ Applying TRP to EfficientNet for Image Classification...")
model.trp_blocks = torch.nn.ModuleList([TPBlock(depths=d, in_planes=in_planes, out_planes=out_planes, rank=k) for k, d in enumerate(depths)])
model.forward = types.MethodType(EfficientNetConfig.gen_forward(rewards, label_smoothing=kwargs["label_smoothing"], top_k=1), model)
elif isinstance(model, VisionTransformer):
print("✅ Applying TRP to VisionTransformer for Image Classification...")
model.trp_blocks = torch.nn.ModuleList([TPBlock(depths=d, in_planes=in_planes, out_planes=out_planes, rank=k, shape_dims=2, channel_first=False) for k, d in enumerate(depths)])
model.forward = types.MethodType(VisionTransformerConfig.gen_forward(rewards, label_smoothing=kwargs["label_smoothing"], top_k=1), model)
return model
|