File size: 14,531 Bytes
e2cffd9 |
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 |
"""
ISL Sign Language Translation - TechMatrix Solvers Initiative
Model definitions for body pose and hand pose estimation
Developed by: TechMatrix Solvers Team
"""
import torch
from collections import OrderedDict
import torch.nn as nn
def construct_layers(layer_config, no_relu_layers, prelu_layers=[]):
"""
Constructs neural network layers based on configuration
Args:
layer_config: Dictionary defining layer parameters
no_relu_layers: List of layers that shouldn't have ReLU activation
prelu_layers: List of layers that should use PReLU instead of ReLU
"""
layers = []
for layer_name, params in layer_config.items():
if 'pool' in layer_name:
layer = nn.MaxPool2d(kernel_size=params[0], stride=params[1], padding=params[2])
layers.append((layer_name, layer))
else:
conv2d = nn.Conv2d(
in_channels=params[0],
out_channels=params[1],
kernel_size=params[2],
stride=params[3],
padding=params[4]
)
layers.append((layer_name, conv2d))
if layer_name not in no_relu_layers:
if layer_name not in prelu_layers:
layers.append(('relu_' + layer_name, nn.ReLU(inplace=True)))
else:
layers.append(('prelu' + layer_name[4:], nn.PReLU(params[1])))
return nn.Sequential(OrderedDict(layers))
def construct_multi_conv_layers(layer_config, no_relu_layers):
"""
Constructs multiple convolution layers for complex architectures
"""
modules = []
for layer_name, params in layer_config.items():
layers = []
if 'pool' in layer_name:
layer = nn.MaxPool2d(kernel_size=params[0], stride=params[1], padding=params[2])
layers.append((layer_name, layer))
else:
conv2d = nn.Conv2d(
in_channels=params[0],
out_channels=params[1],
kernel_size=params[2],
stride=params[3],
padding=params[4]
)
layers.append((layer_name, conv2d))
if layer_name not in no_relu_layers:
layers.append(('Mprelu' + layer_name[5:], nn.PReLU(params[1])))
modules.append(nn.Sequential(OrderedDict(layers)))
return nn.ModuleList(modules)
class BodyPose25Model(nn.Module):
"""
Body pose estimation model using 25-point skeleton
Developed by TechMatrix Solvers for ISL translation
"""
def __init__(self):
super(BodyPose25Model, self).__init__()
# Define layers without ReLU activation
no_relu_layers = [
'Mconv7_stage0_L1', 'Mconv7_stage0_L2',
'Mconv7_stage1_L1', 'Mconv7_stage1_L2',
'Mconv7_stage2_L2', 'Mconv7_stage3_L2'
]
prelu_layers = ['conv4_2', 'conv4_3_CPM', 'conv4_4_CPM']
# Initial feature extraction layers
base_layers = OrderedDict([
('conv1_1', [3, 64, 3, 1, 1]),
('conv1_2', [64, 64, 3, 1, 1]),
('pool1_stage1', [2, 2, 0]),
('conv2_1', [64, 128, 3, 1, 1]),
('conv2_2', [128, 128, 3, 1, 1]),
('pool2_stage1', [2, 2, 0]),
('conv3_1', [128, 256, 3, 1, 1]),
('conv3_2', [256, 256, 3, 1, 1]),
('conv3_3', [256, 256, 3, 1, 1]),
('conv3_4', [256, 256, 3, 1, 1]),
('pool3_stage1', [2, 2, 0]),
('conv4_1', [256, 512, 3, 1, 1]),
('conv4_2', [512, 512, 3, 1, 1]),
('conv4_3_CPM', [512, 256, 3, 1, 1]),
('conv4_4_CPM', [256, 128, 3, 1, 1])
])
self.base_model = construct_layers(base_layers, no_relu_layers, prelu_layers)
# Multi-stage refinement blocks
stage_blocks = {}
# L2 branch - Stage 0
stage_blocks['Mconv1_stage0_L2'] = OrderedDict([
('Mconv1_stage0_L2_0', [128, 96, 3, 1, 1]),
('Mconv1_stage0_L2_1', [96, 96, 3, 1, 1]),
('Mconv1_stage0_L2_2', [96, 96, 3, 1, 1])
])
for i in range(2, 6):
stage_blocks[f'Mconv{i}_stage0_L2'] = OrderedDict([
(f'Mconv{i}_stage0_L2_0', [288, 96, 3, 1, 1]),
(f'Mconv{i}_stage0_L2_1', [96, 96, 3, 1, 1]),
(f'Mconv{i}_stage0_L2_2', [96, 96, 3, 1, 1])
])
stage_blocks['Mconv6_7_stage0_L2'] = OrderedDict([
('Mconv6_stage0_L2', [288, 256, 1, 1, 0]),
('Mconv7_stage0_L2', [256, 52, 1, 1, 0])
])
# L2 branch - Stages 1-3
for stage in range(1, 4):
stage_blocks[f'Mconv1_stage{stage}_L2'] = OrderedDict([
(f'Mconv1_stage{stage}_L2_0', [180, 128, 3, 1, 1]),
(f'Mconv1_stage{stage}_L2_1', [128, 128, 3, 1, 1]),
(f'Mconv1_stage{stage}_L2_2', [128, 128, 3, 1, 1])
])
for i in range(2, 6):
stage_blocks[f'Mconv{i}_stage{stage}_L2'] = OrderedDict([
(f'Mconv{i}_stage{stage}_L2_0', [384, 128, 3, 1, 1]),
(f'Mconv{i}_stage{stage}_L2_1', [128, 128, 3, 1, 1]),
(f'Mconv{i}_stage{stage}_L2_2', [128, 128, 3, 1, 1])
])
stage_blocks[f'Mconv6_7_stage{stage}_L2'] = OrderedDict([
(f'Mconv6_stage{stage}_L2', [384, 512, 1, 1, 0]),
(f'Mconv7_stage{stage}_L2', [512, 52, 1, 1, 0])
])
# L1 branch configurations
stage_blocks['Mconv1_stage0_L1'] = OrderedDict([
('Mconv1_stage0_L1_0', [180, 96, 3, 1, 1]),
('Mconv1_stage0_L1_1', [96, 96, 3, 1, 1]),
('Mconv1_stage0_L1_2', [96, 96, 3, 1, 1])
])
for i in range(2, 6):
stage_blocks[f'Mconv{i}_stage0_L1'] = OrderedDict([
(f'Mconv{i}_stage0_L1_0', [288, 96, 3, 1, 1]),
(f'Mconv{i}_stage0_L1_1', [96, 96, 3, 1, 1]),
(f'Mconv{i}_stage0_L1_2', [96, 96, 3, 1, 1])
])
stage_blocks['Mconv6_7_stage0_L1'] = OrderedDict([
('Mconv6_stage0_L1', [288, 256, 1, 1, 0]),
('Mconv7_stage0_L1', [256, 26, 1, 1, 0])
])
stage_blocks['Mconv1_stage1_L1'] = OrderedDict([
('Mconv1_stage1_L1_0', [206, 128, 3, 1, 1]),
('Mconv1_stage1_L1_1', [128, 128, 3, 1, 1]),
('Mconv1_stage1_L1_2', [128, 128, 3, 1, 1])
])
for i in range(2, 6):
stage_blocks[f'Mconv{i}_stage1_L1'] = OrderedDict([
(f'Mconv{i}_stage1_L1_0', [384, 128, 3, 1, 1]),
(f'Mconv{i}_stage1_L1_1', [128, 128, 3, 1, 1]),
(f'Mconv{i}_stage1_L1_2', [128, 128, 3, 1, 1])
])
stage_blocks['Mconv6_7_stage1_L1'] = OrderedDict([
('Mconv6_stage1_L1', [384, 512, 1, 1, 0]),
('Mconv7_stage1_L1', [512, 26, 1, 1, 0])
])
# Build multi-conv modules
for block_name in stage_blocks.keys():
stage_blocks[block_name] = construct_multi_conv_layers(stage_blocks[block_name], no_relu_layers)
self.stage_models = nn.ModuleDict(stage_blocks)
# Freeze parameters for efficiency
for param in self.parameters():
param.requires_grad = False
def _multi_conv_forward(self, x, models):
"""Forward pass through multi-convolution blocks"""
outputs = []
current_output = x
for model in models:
current_output = model(current_output)
outputs.append(current_output)
return torch.cat(outputs, 1)
def forward(self, x):
"""Forward pass through the body pose model"""
base_features = self.base_model(x)
# L2 branch processing
current_features = base_features
for stage in range(4):
current_features = self._multi_conv_forward(
current_features, self.stage_models[f'Mconv1_stage{stage}_L2']
)
for layer in range(2, 6):
current_features = self._multi_conv_forward(
current_features, self.stage_models[f'Mconv{layer}_stage{stage}_L2']
)
current_features = self.stage_models[f'Mconv6_7_stage{stage}_L2'][0](current_features)
current_features = self.stage_models[f'Mconv6_7_stage{stage}_L2'][1](current_features)
l2_output = current_features
current_features = torch.cat([base_features, current_features], 1)
# L1 branch - Stage 0
current_features = self._multi_conv_forward(
current_features, self.stage_models['Mconv1_stage0_L1']
)
for layer in range(2, 6):
current_features = self._multi_conv_forward(
current_features, self.stage_models[f'Mconv{layer}_stage0_L1']
)
current_features = self.stage_models['Mconv6_7_stage0_L1'][0](current_features)
current_features = self.stage_models['Mconv6_7_stage0_L1'][1](current_features)
stage0_l1_output = current_features
current_features = torch.cat([base_features, stage0_l1_output, l2_output], 1)
# L1 branch - Stage 1
current_features = self._multi_conv_forward(
current_features, self.stage_models['Mconv1_stage1_L1']
)
for layer in range(2, 6):
current_features = self._multi_conv_forward(
current_features, self.stage_models[f'Mconv{layer}_stage1_L1']
)
current_features = self.stage_models['Mconv6_7_stage1_L1'][0](current_features)
stage1_l1_output = self.stage_models['Mconv6_7_stage1_L1'][1](current_features)
return l2_output, stage1_l1_output
class HandPoseModel(nn.Module):
"""
Hand pose estimation model using 21-point hand landmarks
Developed by TechMatrix Solvers for ISL translation
"""
def __init__(self):
super(HandPoseModel, self).__init__()
# Layers without ReLU activation
no_relu_layers = [
'conv6_2_CPM', 'Mconv7_stage2', 'Mconv7_stage3',
'Mconv7_stage4', 'Mconv7_stage5', 'Mconv7_stage6'
]
# Stage 1 - Feature extraction
stage1_base = OrderedDict([
('conv1_1', [3, 64, 3, 1, 1]),
('conv1_2', [64, 64, 3, 1, 1]),
('pool1_stage1', [2, 2, 0]),
('conv2_1', [64, 128, 3, 1, 1]),
('conv2_2', [128, 128, 3, 1, 1]),
('pool2_stage1', [2, 2, 0]),
('conv3_1', [128, 256, 3, 1, 1]),
('conv3_2', [256, 256, 3, 1, 1]),
('conv3_3', [256, 256, 3, 1, 1]),
('conv3_4', [256, 256, 3, 1, 1]),
('pool3_stage1', [2, 2, 0]),
('conv4_1', [256, 512, 3, 1, 1]),
('conv4_2', [512, 512, 3, 1, 1]),
('conv4_3', [512, 512, 3, 1, 1]),
('conv4_4', [512, 512, 3, 1, 1]),
('conv5_1', [512, 512, 3, 1, 1]),
('conv5_2', [512, 512, 3, 1, 1]),
('conv5_3_CPM', [512, 128, 3, 1, 1])
])
stage1_prediction = OrderedDict([
('conv6_1_CPM', [128, 512, 1, 1, 0]),
('conv6_2_CPM', [512, 22, 1, 1, 0])
])
stage_blocks = {}
stage_blocks['stage1_base'] = stage1_base
stage_blocks['stage1_prediction'] = stage1_prediction
# Stages 2-6 refinement
for i in range(2, 7):
stage_blocks[f'stage{i}'] = OrderedDict([
(f'Mconv1_stage{i}', [150, 128, 7, 1, 3]),
(f'Mconv2_stage{i}', [128, 128, 7, 1, 3]),
(f'Mconv3_stage{i}', [128, 128, 7, 1, 3]),
(f'Mconv4_stage{i}', [128, 128, 7, 1, 3]),
(f'Mconv5_stage{i}', [128, 128, 7, 1, 3]),
(f'Mconv6_stage{i}', [128, 128, 1, 1, 0]),
(f'Mconv7_stage{i}', [128, 22, 1, 1, 0])
])
# Build all stage models
for block_name in stage_blocks.keys():
stage_blocks[block_name] = construct_layers(stage_blocks[block_name], no_relu_layers)
self.stage1_base_model = stage_blocks['stage1_base']
self.stage1_prediction_model = stage_blocks['stage1_prediction']
self.stage2_model = stage_blocks['stage2']
self.stage3_model = stage_blocks['stage3']
self.stage4_model = stage_blocks['stage4']
self.stage5_model = stage_blocks['stage5']
self.stage6_model = stage_blocks['stage6']
# Freeze parameters for efficiency
for param in self.parameters():
param.requires_grad = False
def forward(self, x):
"""Forward pass through the hand pose model"""
base_features = self.stage1_base_model(x)
stage1_output = self.stage1_prediction_model(base_features)
# Stage 2
stage2_input = torch.cat([stage1_output, base_features], 1)
stage2_output = self.stage2_model(stage2_input)
# Stage 3
stage3_input = torch.cat([stage2_output, base_features], 1)
stage3_output = self.stage3_model(stage3_input)
# Stage 4
stage4_input = torch.cat([stage3_output, base_features], 1)
stage4_output = self.stage4_model(stage4_input)
# Stage 5
stage5_input = torch.cat([stage4_output, base_features], 1)
stage5_output = self.stage5_model(stage5_input)
# Stage 6
stage6_input = torch.cat([stage5_output, base_features], 1)
stage6_output = self.stage6_model(stage6_input)
return stage6_output
# Factory functions for easy model instantiation
def create_bodypose_model():
"""Create and return body pose detection model"""
return BodyPose25Model()
def create_handpose_model():
"""Create and return hand pose detection model"""
return HandPoseModel() |