File size: 11,681 Bytes
f21948d |
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 |
import os
import sys
sys.path.append(os.path.dirname(os.path.dirname(os.path.abspath(__file__))))
from utils import suppress_warnings
import numpy as np
import tensorflow as tf
import torch
import torch.nn as nn
import logging
from pathlib import Path
logger = logging.getLogger(__name__)
class ModelLoader:
def __init__(self, weights_dir="models/weights"):
"""Initialize model loader with paths to weight files"""
self.weights_dir = Path(weights_dir)
# Model paths (adjust these based on your setup)
self.STGCN_WEIGHTS = self.weights_dir / "best_stgcn.weights.h5"
self.TRANSFORMER_MODEL_PATH = self.weights_dir / "Transformer_12rel_4_bs16_sl32.keras"
self.ANGLE_MODEL_PATH = self.weights_dir / "Transformer_12rel_4_angle3_branch_bs16_sl32.keras"
self.SWIN3D_WEIGHTS = self.weights_dir / "best_swin3d_b_22k.pth"
# Constants
self.SEQ_LEN = 32
self.NUM_CLASSES = 22
self.ACTIONS = [
"barbell biceps curl","lateral raise","push-up","bench press",
"chest fly machine","deadlift","decline bench press","hammer curl",
"hip thrust","incline bench press","lat pulldown","leg extension",
"leg raises","plank","pull Up","romanian deadlift","russian twist",
"shoulder press","squat","t bar row","tricep Pushdown","tricep dips"
]
# Device for PyTorch
self.device = "cuda" if torch.cuda.is_available() else "cpu"
# Initialize models
self.models = {}
self._load_all_models()
def _build_stgcn_model(self):
"""Build ST-GCN model architecture"""
from tensorflow.keras import layers, Model, regularizers
# ST-GCN Block
class STGCNBlock(layers.Layer):
def __init__(self, in_ch, out_ch, A_norm, stride=1, dropout=0.3, **kwargs):
super().__init__(**kwargs)
self.A_const = tf.constant(A_norm.astype(np.float32))
self.B = self.add_weight(shape=A_norm.shape, initializer="zeros", trainable=True)
self.conv1x1 = layers.Conv2D(out_ch, 1, use_bias=False,
kernel_regularizer=regularizers.l2(1e-5))
self.branches = [
layers.Conv2D(out_ch, (k,1), strides=(stride,1), padding="same",
use_bias=False, kernel_regularizer=regularizers.l2(1e-5))
for k in (3,5,9)
]
self.bn = layers.BatchNormalization()
if in_ch == out_ch and stride == 1:
self.res = lambda x, training: x
else:
self.res = tf.keras.Sequential([
layers.Conv2D(out_ch,1,strides=(stride,1),padding="same",
use_bias=False, kernel_regularizer=regularizers.l2(1e-5)),
layers.BatchNormalization()
])
self.act = layers.Activation("relu")
self.drop = layers.SpatialDropout2D(dropout)
def call(self, x, training=False):
A_mat = self.A_const + tf.nn.softmax(self.B, axis=1)
x_sp = tf.einsum('ij,btjk->btik', A_mat, x)
x_sp = self.conv1x1(x_sp)
out = sum(branch(x_sp) for branch in self.branches) / len(self.branches)
out = self.bn(out, training=training)
r = self.res(x, training=training) if callable(self.res) else self.res(x)
y = self.act(out + r)
return self.drop(y, training=training)
# Build adjacency matrix
V = 31 # Number of joints in MediaPipe
connections = [
(0,1),(1,2),(2,3),(3,7),(0,4),(4,5),(5,6),(6,8),(9,10),
(11,12),(11,13),(13,15),(15,17),(17,19),(19,21),
(12,14),(14,16),(16,18),(18,20),(20,22),
(11,23),(12,24),(23,24),(23,25),(25,27),(27,29),(27,31),
(24,26),(26,28),(28,30)
]
A = np.zeros((V, V), dtype=np.float32)
for u, v in connections:
if u < V and v < V:
A[u, v] = A[v, u] = 1.0
# Normalize adjacency matrix
D = A.sum(axis=1)
D_inv = np.diag(1.0 / np.sqrt(D + 1e-6))
A_norm = D_inv @ A @ D_inv
# Build model
C = 4 # x, y, z, visibility
inp = layers.Input((self.SEQ_LEN, V, C))
x = STGCNBlock(C, 64, A_norm, stride=1)(inp)
x = STGCNBlock(64, 64, A_norm, stride=2)(x)
x = STGCNBlock(64, 128, A_norm, stride=2)(x)
x = STGCNBlock(128,256, A_norm, stride=2)(x)
x = layers.GlobalAveragePooling2D()(x)
out = layers.Dense(self.NUM_CLASSES, activation="softmax",
kernel_regularizer=regularizers.l2(1e-5))(x)
return Model(inp, out)
def _build_swin3d_model(self):
"""Build Swin3D model"""
from torchvision.models.video import swin3d_b, Swin3D_B_Weights
model = swin3d_b(weights=Swin3D_B_Weights.KINETICS400_IMAGENET22K_V1)
model.head = nn.Linear(model.head.in_features, self.NUM_CLASSES)
model = model.to(self.device)
return model
@tf.keras.utils.register_keras_serializable()
class PositionalEncoding(tf.keras.layers.Layer):
"""Positional encoding for Transformer models"""
def __init__(self, maxlen, dm, **kwargs):
super().__init__(**kwargs)
pos = np.arange(maxlen)[:, None]
i = np.arange(dm)[None, :]
angle = pos / np.power(10000, (2*(i//2))/dm)
pe = np.zeros((maxlen, dm), dtype=np.float32)
pe[:,0::2] = np.sin(angle[:,0::2])
pe[:,1::2] = np.cos(angle[:,1::2])
self.pe = tf.constant(pe[None,...])
def call(self, x):
return x + self.pe[:, :tf.shape(x)[1],:]
def get_config(self):
cfg = super().get_config()
cfg.update({"maxlen": int(self.pe.shape[1]), "dm": int(self.pe.shape[2])})
return cfg
def _load_all_models(self):
"""Load all four models"""
logger.info("Loading all models...")
try:
# 1. Load ST-GCN
logger.info("Loading ST-GCN...")
if self.STGCN_WEIGHTS.exists():
model_stgcn = self._build_stgcn_model()
try:
model_stgcn.load_weights(str(self.STGCN_WEIGHTS))
logger.info("ST-GCN weights loaded successfully.")
except Exception as e:
logger.warning(f"ST-GCN weight load error: {e}. Using skip_mismatch.")
model_stgcn.load_weights(str(self.STGCN_WEIGHTS), skip_mismatch=True)
self.models['stgcn'] = model_stgcn
else:
logger.warning(f"ST-GCN weights not found: {self.STGCN_WEIGHTS}")
self.models['stgcn'] = None
# 2. Load Transformer 12rel
logger.info("Loading Transformer 12rel...")
if self.TRANSFORMER_MODEL_PATH.exists():
model_transformer = tf.keras.models.load_model(
str(self.TRANSFORMER_MODEL_PATH),
custom_objects={'PositionalEncoding': self.PositionalEncoding}
)
self.models['transformer_12rel'] = model_transformer
logger.info("Transformer 12rel loaded successfully.")
else:
logger.warning(f"Transformer 12rel not found: {self.TRANSFORMER_MODEL_PATH}")
self.models['transformer_12rel'] = None
# 3. Load Transformer angle branch
logger.info("Loading Transformer angle branch...")
if self.ANGLE_MODEL_PATH.exists():
model_angle = tf.keras.models.load_model(str(self.ANGLE_MODEL_PATH))
self.models['transformer_angle'] = model_angle
logger.info("Transformer angle branch loaded successfully.")
else:
logger.warning(f"Transformer angle not found: {self.ANGLE_MODEL_PATH}")
self.models['transformer_angle'] = None
# 4. Load Swin3D
logger.info("Loading Swin3D...")
if self.SWIN3D_WEIGHTS.exists():
model_swin3d = self._build_swin3d_model()
state = torch.load(str(self.SWIN3D_WEIGHTS), map_location=self.device)
model_swin3d.load_state_dict(state)
model_swin3d.eval()
self.models['swin3d'] = model_swin3d
logger.info("Swin3D loaded successfully.")
else:
logger.warning(f"Swin3D weights not found: {self.SWIN3D_WEIGHTS}")
self.models['swin3d'] = None
# Check if any models loaded
loaded_models = [name for name, model in self.models.items() if model is not None]
if not loaded_models:
logger.warning("No models could be loaded. App will run in demo mode with mock predictions.")
# Don't raise error, just warn - app can still work for testing UI
else:
logger.info(f"Successfully loaded models: {loaded_models}")
except Exception as e:
logger.error(f"Error loading models: {str(e)}")
raise
def get_model(self, model_name):
"""Get a specific model by name"""
return self.models.get(model_name)
def get_available_models(self):
"""Get list of available (loaded) models"""
return [name for name, model in self.models.items() if model is not None]
def predict_stgcn(self, X):
"""Predict using ST-GCN model"""
if self.models['stgcn'] is None:
return None
return self.models['stgcn'].predict(X, batch_size=32, verbose=0)
def predict_transformer_12rel(self, X):
"""Predict using Transformer 12rel model"""
if self.models['transformer_12rel'] is None:
return None
return self.models['transformer_12rel'].predict(X, batch_size=32, verbose=0)
def predict_transformer_angle(self, X_rel, X_ang):
"""Predict using Transformer angle branch model"""
if self.models['transformer_angle'] is None:
return None
return self.models['transformer_angle'].predict([X_rel, X_ang], batch_size=32, verbose=0)
def predict_swin3d(self, X):
"""Predict using Swin3D model"""
if self.models['swin3d'] is None:
return None
with torch.no_grad():
probas = []
for x in X:
x_batch = x.unsqueeze(0).to(self.device)
logits = self.models['swin3d'](x_batch)
proba = torch.softmax(logits, dim=1).cpu().numpy()
probas.append(proba)
return np.vstack(probas)
def cleanup(self):
"""Clean up model resources"""
logger.info("Cleaning up model resources...")
for name, model in self.models.items():
if model is not None:
if name == 'swin3d':
# PyTorch model cleanup
del model
torch.cuda.empty_cache()
else:
# TensorFlow model cleanup
del model
# Clear TensorFlow session
tf.keras.backend.clear_session() |