Spaces:
Running
on
Zero
Running
on
Zero
File size: 18,337 Bytes
7c34c28 |
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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 |
import torch.nn as nn
import torch
from typing import Dict
import numpy as np
import torch.nn.functional as F
def build_patch_mlp_projector(
input_hidden_size: int, lm_hidden_size: int, num_layers: int
) -> nn.Module:
modules = [nn.Linear(input_hidden_size, lm_hidden_size)]
for _ in range(1, num_layers):
modules.append(nn.GELU())
modules.append(nn.Linear(lm_hidden_size, lm_hidden_size))
return nn.Sequential(*modules)
class _MLPVectorProjector(nn.Module):
def __init__(
self, input_hidden_size: int, lm_hidden_size: int, num_layers: int, width: int
):
super(_MLPVectorProjector, self).__init__()
self.mlps = nn.ModuleList()
for _ in range(width):
mlp = [nn.Linear(input_hidden_size, lm_hidden_size)]
for _ in range(1, num_layers):
mlp.append(nn.GELU())
mlp.append(nn.Linear(lm_hidden_size, lm_hidden_size))
self.mlps.append(nn.Sequential(*mlp))
def forward(self, x):
output = torch.cat([mlp(x) for mlp in self.mlps], dim=-2)
return output
def build_mlp_vector_projector(
input_hidden_size: int, lm_hidden_size: int, num_layers: int, num_tokens: int
):
return _MLPVectorProjector(
input_hidden_size, lm_hidden_size, num_layers, num_tokens
)
class MLPBackbone(nn.Module):
def __init__(self, input_size: int, output_size: int, num_layers: int, hidden_dim: int):
super(MLPBackbone, self).__init__()
self.output_size = output_size
mlp_layers = self._create_mlp_layers(input_size, output_size, num_layers, hidden_dim)
self.layers = nn.Sequential(*mlp_layers)
def _create_conv_layers(self, input_channels, num_conv_layers, hidden_dim):
layers = []
for _ in range(num_conv_layers):
layers += [
nn.Conv1d(input_channels, hidden_dim, kernel_size=3, padding=1),
nn.GELU(),
nn.MaxPool1d(kernel_size=2, stride=2)
]
input_channels = hidden_dim
return layers
def _create_mlp_layers(self, input_size, output_size, num_layers, hidden_dim):
if num_layers >=2:
layers = [nn.Linear(input_size, hidden_dim)]
layers.append(nn.GELU())
if num_layers > 2:
for _ in range(1, num_layers - 2):
layers += [
nn.Linear(hidden_dim, hidden_dim),
nn.GELU()
]
layers.append(nn.Linear(hidden_dim, output_size))
else:
layers = [nn.Linear(input_size, output_size)]
return layers
def forward(self, x):
return self.layers(x)
class MLPTaskHead(nn.Module):
def __init__(self, backbone: nn.Module, input_size: int, output_size: int, num_layers: int, hidden_dim: int, width: int = 1):
super(MLPTaskHead, self).__init__()
self.backbone = backbone
self.width = width
if width > 1:
self.layers = nn.ModuleList()
for i in range(width):
mlp_layers = [nn.GELU()]
mlp_layers += self._create_mlp_layers(input_size, output_size, num_layers, hidden_dim)
self.layers.append(nn.Sequential(*mlp_layers))
else:
mlp_layers = [nn.GELU()]
mlp_layers += self._create_mlp_layers(input_size, output_size, num_layers, hidden_dim)
self.layers = nn.Sequential(*mlp_layers)
def _create_mlp_layers(self, input_size, output_size, num_layers, hidden_dim):
if num_layers >=2:
layers = [nn.Linear(input_size, hidden_dim)]
layers.append(nn.GELU())
if num_layers > 2:
for _ in range(1, num_layers - 2):
layers += [
nn.Linear(hidden_dim, hidden_dim),
nn.GELU()
]
layers.append(nn.Linear(hidden_dim, output_size))
else:
layers = [nn.Linear(input_size, output_size)]
return layers
def _create_conv_layers(self, input_channels, num_conv_layers, hidden_dim):
layers = []
for _ in range(num_conv_layers):
layers += [
nn.Conv2d(in_channels = input_channels, out_channels = hidden_dim, kernel_size=(3,3), stride=1, padding=1),
nn.GELU(),
nn.MaxPool1d(kernel_size=2, stride=2)
]
input_channels = hidden_dim
return layers
def forward(self, x):
output = self.backbone.forward(x)
if self.width > 1:
return torch.cat([layer(output) for layer in self.layers], dim=-2)
else:
return self.layers(output)
class MLPTaskModule(nn.Module):
def __init__(self, input_size: int, output_size: int, num_layers: int, hidden_dim: int, width: int = 1):
super(MLPTaskModule, self).__init__()
self.width = width
if width > 1:
self.layers = nn.ModuleList()
for i in range(width):
mlp_layers = [nn.GELU()]
mlp_layers += self._create_mlp_layers(input_size, output_size, num_layers, hidden_dim)
self.layers.append(nn.Sequential(*mlp_layers))
else:
mlp_layers = [nn.GELU()]
mlp_layers += self._create_mlp_layers(input_size, output_size, num_layers, hidden_dim)
self.layers = nn.Sequential(*mlp_layers)
def _create_mlp_layers(self, input_size, output_size, num_layers, hidden_dim):
if num_layers >=2:
layers = [nn.Linear(input_size, hidden_dim)]
layers.append(nn.GELU())
if num_layers > 2:
for _ in range(1, num_layers - 2):
layers += [
nn.Linear(hidden_dim, hidden_dim),
nn.GELU()
]
layers.append(nn.Linear(hidden_dim, output_size))
else:
layers = [nn.Linear(input_size, output_size)]
return layers
def _create_conv_layers(self, input_channels, num_conv_layers, hidden_dim):
layers = []
for _ in range(num_conv_layers):
layers += [
nn.Conv2d(in_channels = input_channels, out_channels = hidden_dim, kernel_size=(3,3), stride=1, padding=1),
nn.GELU(),
nn.MaxPool1d(kernel_size=2, stride=2)
]
input_channels = hidden_dim
return layers
def forward(self, x):
if self.width > 1:
return torch.cat([layer(x) for layer in self.layers], dim=-2)
else:
return self.layers(x)
class MultiTaskModel(nn.Module):
def __init__(self, input_hidden_size: int, input_channels: int, time_average: bool, time_dimension: int, use_aggregator: bool, tasks: Dict):
super(MultiTaskModel, self).__init__()
self.tasks = tasks
self.time_average = time_average
self.time_dimension = time_dimension
self.use_aggregator = use_aggregator
if self.use_aggregator:
if (time_average):
self.aggregator = nn.Parameter(torch.randn((input_channels, 1), dtype = torch.float))
else:
self.aggregator = nn.Parameter(torch.randn((input_channels, 1, 1), dtype = torch.float))
self.backbone = MLPBackbone(input_hidden_size, self.tasks["backbone"]["output_size"], self.tasks["backbone"]["num_layers"], self.tasks["backbone"]["hidden_size"])
for task_name, task_head in self.tasks["task_heads"].items():
setattr(self, task_name, MLPTaskModule(self.tasks["backbone"]["output_size"], task_head["output_size"], task_head["num_layers"], task_head["hidden_size"], task_head["width"]))
if task_name in self.tasks["task_projectors"].keys():
task_projector = tasks["task_projectors"][task_name]
setattr(self, task_name + "_projector", MLPTaskModule(task_head["output_size"], task_projector["output_size"], task_projector["num_layers"], task_projector["hidden_size"], task_projector["width"]))
def forward(self, x):
task_head_outputs = {}
task_projector_outputs = []
if self.time_average:
x = x.mean(self.time_dimension)
if self.use_aggregator:
aggregator_weights = F.softmax(self.aggregator, dim=0)
aggregator_output = (x * aggregator_weights).sum(dim=0)
aggregator_output = aggregator_output.unsqueeze(0)
else:
aggregator_output = x
backbone_output = self.backbone(aggregator_output)
for task_name in self.tasks["task_heads"]:
if task_name != "lmm_projector":
task_head_output = getattr(self, task_name)(backbone_output)
min_val = torch.min(task_head_output)
max_val = torch.max(task_head_output)
normalized_task_head_output = (task_head_output - min_val) / (max_val - min_val)
task_head_outputs[task_name] = normalized_task_head_output
if task_name in self.tasks["task_projectors"].keys():
task_projector_outputs.append(getattr(self, task_name + "_projector")(task_head_outputs[task_name]))
else:
task_projector_outputs.append(getattr(self, task_name)(backbone_output))
task_projector_outputs_unsqueezed = [task_projector_output.unsqueeze(0) for task_projector_output in task_projector_outputs]
if len(task_projector_outputs_unsqueezed) > 0:
task_head_outputs["projectors"] = torch.cat(task_projector_outputs_unsqueezed, dim=-2)
return task_head_outputs
def build_mt_vector_projector(
input_hidden_size: int, lm_hidden_size: int, tasks: Dict
):
projector = nn.ModuleDict()
projector["backbone"] = MLPBackbone(input_hidden_size, tasks["backbone"]["output_size"], tasks["backbone"]["num_layers"], tasks["backbone"]["hidden_size"])
for task_name, task_head in tasks["task_heads"].items():
projector[task_name] = MLPTaskHead(projector["backbone"], task_head["hidden_size"], task_head["output_size"], task_head["num_layers"], task_head["hidden_size"], task_head["width"])
return projector
class Attention(nn.Module):
def __init__(self, input_dim, hidden_dim):
super(Attention, self).__init__()
self.linear_in = nn.Linear(input_dim, hidden_dim)
self.linear_out = nn.Linear(hidden_dim, 1)
def forward(self, x):
# Input shape: (batch_size, seq_len, input_dim)
energy = torch.tanh(self.linear_in(x))
attention_scores = torch.softmax(self.linear_out(energy), dim=1)
context_vector = torch.sum(attention_scores * x, dim=1)
return context_vector
class _CNNAttentionTokenizer(nn.Module):
def __init__(self, input_channels, output_size, width, hidden_dim, num_conv_layers):
super(_CNNAttentionTokenizer, self).__init__()
self.width = width
self.cnns = nn.ModuleList()
self.attentions = nn.ModuleList()
for _ in range(width):
cnn = self._create_conv_layers(input_channels, num_conv_layers)
self.cnns.append(cnn)
attention = [Attention(hidden_dim, 125)]
linear_input_size = hidden_dim
attention.append(nn.Linear(linear_input_size, output_size))
self.attentions.append(nn.Sequential(*attention))
def _create_conv_layers(self, input_channels, num_conv_layers):
layers = []
in_channels = input_channels
for _ in range(num_conv_layers):
layers += [
nn.Conv1d(in_channels, 64, kernel_size=3, padding=1),
nn.ReLU(),
nn.MaxPool1d(kernel_size=2, stride=2)
]
in_channels = 64
return nn.Sequential(*layers)
def forward(self, x):
outputs = []
for token in range(self.width):
# Input shape: (batch_size, input_channels, sequence_length)
token_output = self.cnns[token](x) # Apply convolutional layers
token_output = token_output.permute(0, 2, 1) # Reshape for attention mechanism (batch_size, sequence_length, input_dim
token_output = self.attentions[token](token_output) # Apply attention mechanism
outputs.append(token_output)
output = torch.cat(outputs, dim=-2)
output = torch.stack([output])
return output
def build_attentive_cnn_projector(
input_channels: int, lm_hidden_size: int, num_tokens: int, hidden_dim: int, num_layers: int
):
return _CNNAttentionTokenizer(input_channels, lm_hidden_size, num_tokens, hidden_dim, num_layers)
class _CNNMLPProjector(nn.Module):
def __init__(self, input_channels, input_size, output_size = 4096, width = 5, hidden_dim = 64, num_conv_layers = 1, num_mlp_layers = 2):
super(_CNNMLPProjector, self).__init__()
self.width = width
self.cnnmlps = nn.ModuleList()
for _ in range(self.width):
cnnmlp = self._create_conv_layers(input_channels, num_conv_layers, hidden_dim)
cnnmlp.append(nn.Flatten())
cnn_output_size = hidden_dim*((input_size + 2*1 - 3*num_conv_layers) // (2**num_conv_layers) + 1)
cnnmlp.append(nn.Linear(cnn_output_size, output_size))
cnnmlp.append(nn.GELU())
cnnmlp += self._create_mlp_layers(output_size, output_size, num_mlp_layers, output_size)
self.cnnmlps.append(nn.Sequential(*cnnmlp))
def _create_conv_layers(self, input_channels, num_conv_layers, hidden_dim):
layers = []
for _ in range(num_conv_layers):
layers += [
nn.Conv1d(input_channels, hidden_dim, kernel_size=3, padding=1),
nn.GELU(),
nn.MaxPool1d(kernel_size=2, stride=2)
]
input_channels = hidden_dim
return layers
def _create_mlp_layers(self, input_size, output_size, num_layers, hidden_dim):
if num_layers >=2:
layers = [nn.Linear(input_size, hidden_dim)]
layers.append(nn.GELU())
if num_layers > 2:
for _ in range(1, num_layers - 2):
layers += [
nn.Linear(hidden_dim, hidden_dim),
nn.GELU()
]
layers.append(nn.Linear(hidden_dim, output_size))
else:
layers = [nn.Linear(input_size, output_size)]
return layers
def forward(self, x):
return torch.stack([torch.cat([cnnmlp(x) for cnnmlp in self.cnnmlps], dim=-2)])
def build_cnn_mlp_projector(
input_channels: int, input_size: int, lm_hidden_size: int, num_tokens: int, hidden_dim: int, num_conv_layers: int, num_mlp_layers: int
):
return _CNNMLPProjector(input_channels, input_size, lm_hidden_size, num_tokens, hidden_dim, num_conv_layers, num_mlp_layers)
class _MultiLayeredCNNMLPProjector(nn.Module):
def __init__(self, input_channels, input_size, num_feature_layers, output_size = 4096, width = 5, hidden_dim = 64, num_conv_layers = 1, num_mlp_layers = 2):
super(_MultiLayeredCNNMLPProjector, self).__init__()
self.width = width
self.num_feature_layers = num_feature_layers
self.cnnmlps = nn.ModuleList()
for _ in range(self.width*self.num_feature_layers):
cnnmlp = self._create_conv_layers(input_channels, num_conv_layers, hidden_dim)
cnnmlp += [nn.GELU()]
cnnmlp += self._create_mlp_layers(input_size, output_size, num_mlp_layers, output_size)
self.cnnmlps.append(nn.Sequential(*cnnmlp))
def _create_conv_layers(self, input_channels, num_conv_layers, hidden_size):
layers = []
if input_channels >= hidden_size:
hidden_dim = int(input_channels/2)
else:
hidden_dim = hidden_size
layers += [nn.Conv1d(in_channels=input_channels, out_channels=hidden_dim, kernel_size=3, stride=1, padding=1), nn.GELU()]
if num_conv_layers > 2:
for _ in range(num_conv_layers - 2):
if hidden_dim/2 >= hidden_size:
output_dim = int(hidden_dim/2)
else:
output_dim = hidden_size
layers += [
nn.Conv1d(in_channels=hidden_dim, out_channels=output_dim, kernel_size=3, stride=1, padding=1),
nn.GELU(),
]
hidden_dim = output_dim
layers += [nn.Conv1d(in_channels=hidden_dim, out_channels=1, kernel_size=3, stride=1, padding=1)]
return layers
def _create_mlp_layers(self, input_size, output_size, num_layers, hidden_dim):
if num_layers >=2:
layers = [nn.Linear(input_size, hidden_dim)]
layers.append(nn.GELU())
if num_layers > 2:
for _ in range(1, num_layers - 2):
layers += [
nn.Linear(hidden_dim, hidden_dim),
nn.GELU()
]
layers.append(nn.Linear(hidden_dim, output_size))
else:
layers = [nn.Linear(input_size, output_size)]
return layers
def forward(self, x):
print("X SHAPE ", x.shape)
inp_feature_layers = []
for feature_id in range(self.num_feature_layers):
in_feat_layer = x[feature_id].unsqueeze(0).permute(0,2,1)
inp_feature_layers.append(in_feat_layer)
outputs = []
for layer_count in range(self.width*self.num_feature_layers):
feature_id = int(layer_count/self.width)
outputs+=[self.cnnmlps[layer_count](inp_feature_layers[feature_id])]
return torch.cat(outputs, dim=-2)
def build_multi_layer_cnn_mlp_projector(
input_channels: int, input_size: int, num_feature_layers: int, lm_hidden_size: int, num_tokens: int, hidden_dim: int, num_conv_layers: int, num_mlp_layers: int
):
assert(num_tokens % num_feature_layers == 0)
width = int(num_tokens/num_feature_layers)
return _MultiLayeredCNNMLPProjector(input_channels, input_size, num_feature_layers, lm_hidden_size, width, hidden_dim, num_conv_layers, num_mlp_layers)
|