File size: 15,696 Bytes
f2688f7 |
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 |
import torch
import torch.nn as nn
import torch.optim as optim
import torch.distributed as dist
import torch.multiprocessing as mp
from torch.nn.parallel import DistributedDataParallel as DDP
from torch.utils.data import DataLoader, Dataset, DistributedSampler
import os
import matplotlib.pyplot as plt
from Deep_ANC_model_trim import CRN
import logging
from Pre_processing import Preprocessing
import random
from torch.optim.lr_scheduler import CosineAnnealingLR
from ranger import Ranger
from torch.optim.lr_scheduler import CosineAnnealingWarmRestarts
from torch.optim.lr_scheduler import OneCycleLR
# from deap import base, creator, tools, algorithms # For GA
# import pickle
# import json
# import optuna
# Set up logging
logging.basicConfig(level=logging.INFO)
logger = logging.getLogger(__name__)
# class NoisySpeechDataset(Dataset):
# def __init__(self, noisy_dir, clean_dir):
# self.noisy_files = sorted([os.path.join(noisy_dir, f) for f in os.listdir(noisy_dir) if f.endswith('.pt')])
# self.clean_files = sorted([os.path.join(clean_dir, f) for f in os.listdir(clean_dir) if f.endswith('.pt')])
# def __len__(self):
# return len(self.noisy_files)
# def __getitem__(self, idx):
# noisy_spectrogram = torch.load(self.noisy_files[idx], weights_only=True)
# clean_spectrogram = torch.load(self.clean_files[idx], weights_only=True)
# assert noisy_spectrogram.shape == clean_spectrogram.shape, "Mismatched tensor shapes"
# return noisy_spectrogram, clean_spectrogram
# class NoisySpeechDataset(Dataset):
# def __init__(self, noisy_dir, clean_dir, subset_size=50000):
# self.noisy_files = sorted([os.path.join(noisy_dir, f) for f in os.listdir(noisy_dir) if f.endswith('.pt')])
# self.clean_files = sorted([os.path.join(clean_dir, f) for f in os.listdir(clean_dir) if f.endswith('.pt')])
# assert len(self.noisy_files) == len(self.clean_files), "Mismatched noisy and clean datasets"
# subset_size = min(subset_size, len(self.noisy_files))
# self.noisy_files = self.noisy_files[:subset_size]
# self.clean_files = self.clean_files[:subset_size]
# def __len__(self):
# return len(self.noisy_files)
# def __getitem__(self, idx):
# noisy_spectrogram = torch.load(self.noisy_files[idx], weights_only=True)
# clean_spectrogram = torch.load(self.clean_files[idx], weights_only=True)
# return noisy_spectrogram, clean_spectrogram
def custom_loss_function(output, target):
if output.size() != target.size():
min_size = min(output.size(2), target.size(2))
output = output[:, :, :min_size, :]
target = target[:, :, :min_size, :]
return torch.mean((output - target) ** 2)
class NoisySpeechDataset(Dataset):
def __init__(self, noisy_dir, clean_dir, subset_size=50000, shuffle=True):
self.noisy_files = sorted([os.path.join(noisy_dir, f) for f in os.listdir(noisy_dir) if f.endswith('.pt')])
self.clean_files = sorted([os.path.join(clean_dir, f) for f in os.listdir(clean_dir) if f.endswith('.pt')])
assert len(self.noisy_files) == len(self.clean_files), "Mismatched noisy and clean datasets"
# If shuffle is True, shuffle the dataset
if shuffle:
combined = list(zip(self.noisy_files, self.clean_files))
random.shuffle(combined)
self.noisy_files, self.clean_files = zip(*combined)
# Limit the subset size if provided
subset_size = min(subset_size, len(self.noisy_files))
self.noisy_files = self.noisy_files[:subset_size]
self.clean_files = self.clean_files[:subset_size]
def __len__(self):
return len(self.noisy_files)
def __getitem__(self, idx):
noisy_spectrogram = torch.load(self.noisy_files[idx], weights_only=True)
clean_spectrogram = torch.load(self.clean_files[idx], weights_only=True)
return noisy_spectrogram, clean_spectrogram
# def snr_improvement(noisy, clean, enhanced):
# min_size = min(noisy.size(2), clean.size(2), enhanced.size(2))
# noisy = noisy[:, :, :min_size, :]
# clean = clean[:, :, :min_size, :]
# enhanced = enhanced[:, :, :min_size, :]
# noise = noisy - clean
# noise_est = enhanced - clean
# snr_before = torch.mean(clean ** 2) / torch.mean(noise ** 2)
# snr_after = torch.mean(clean ** 2) / torch.mean(noise_est ** 2)
# return 10 * torch.log10(snr_after / snr_before)
def snr_improvement(noisy, clean, enhanced):
min_size = min(noisy.size(2), clean.size(2), enhanced.size(2))
noisy = noisy[:, :, :min_size, :]
clean = clean[:, :, :min_size, :]
enhanced = enhanced[:, :, :min_size, :]
noise = noisy - clean
noise_est = enhanced - clean
# Ensure the denominator isn't zero to avoid NaN values
noise_power = torch.mean(noise ** 2)
noise_est_power = torch.mean(noise_est ** 2)
if noise_power == 0 or noise_est_power == 0:
return torch.tensor(0.0) # Avoid division by zero and return 0 SNR improvement
snr_before = torch.mean(clean ** 2) / noise_power
snr_after = torch.mean(clean ** 2) / noise_est_power
return 10 * torch.log10(snr_after / snr_before)
def plot_metrics(train_metrics, val_metrics, metric_name):
epochs = range(1, len(train_metrics) + 1)
plt.plot(epochs, train_metrics, 'bo', label=f'Training {metric_name}')
plt.plot(epochs, val_metrics, 'b', label=f'Validation {metric_name}')
plt.title(f'Training and Validation {metric_name}')
plt.xlabel('Epochs')
plt.ylabel(metric_name)
plt.legend()
plt.show()
def train_model(rank, world_size, model, train_loader, val_loader, num_epochs, learning_rate, save_path, best_save_path, checkpoint_path=None):
try:
# Enable anomaly detection
torch.autograd.set_detect_anomaly(True)
# Set the device for the current process
torch.cuda.set_device(rank)
model = model.to(rank)
model = DDP(model, device_ids=[rank])
# # Apply weight initialization
# def init_weights(m):
# if isinstance(m, nn.Conv2d) or isinstance(m, nn.ConvTranspose2d):
# nn.init.kaiming_uniform_(m.weight, nonlinearity='relu') # He initialization
# if m.bias is not None:
# nn.init.constant_(m.bias, 0)
# elif isinstance(m, nn.LSTM):
# for name, param in m.named_parameters():
# if 'weight' in name:
# nn.init.xavier_uniform_(param) # Xavier initialization
# elif 'bias' in name:
# nn.init.constant_(param, 0)
# # Initialize the weights of the model
# model.apply(init_weights)
# Initialize the Adam optimizer
optimizer = optim.Adam(model.parameters(), lr=0.001, amsgrad=True)
# Set up the learning rate scheduler
# scheduler = OneCycleLR(optimizer, max_lr=0.01, steps_per_epoch=1250, epochs=1500)
# Define the Cosine Annealing Warm Restarts scheduler
scheduler = optim.lr_scheduler.ReduceLROnPlateau(optimizer, mode='min',
factor=0.1, patience=20,
verbose=True)
# Initialize variables for tracking progress
start_epoch = 0
best_val_loss = float('inf')
best_val_snr_improvement = float('-inf')
# Load checkpoint if provided and if it exists
if checkpoint_path and os.path.exists(checkpoint_path):
try:
checkpoint = torch.load(checkpoint_path, map_location=torch.device(f'cuda:{rank}'))
print(f"Checkpoint keys: {checkpoint.keys()}") # Inspect the checkpoint keys
# Directly load the state dictionary into the model
model.load_state_dict(checkpoint)
logger.info(f"Model state loaded directly from checkpoint.")
# If available, load optimizer and scheduler states
if 'optimizer_state_dict' in checkpoint:
optimizer.load_state_dict(checkpoint['optimizer_state_dict'])
if 'scheduler_state_dict' in checkpoint:
scheduler.load_state_dict(checkpoint['scheduler_state_dict'])
# Load additional tracking variables if available
start_epoch = checkpoint.get('epoch', 0) + 1
best_val_loss = checkpoint.get('best_val_loss', float('inf'))
best_val_snr_improvement = checkpoint.get('best_val_snr_improvement', float('-inf'))
logger.info(f"Resuming training from epoch {start_epoch}")
except Exception as e:
logger.error(f"Error loading checkpoint: {e}")
raise e # Re-raise the exception to prevent further issues
# Training loop
model.train()
training_snr_improvements = []
validation_snr_improvements = []
for epoch in range(start_epoch, start_epoch + num_epochs):
running_loss = 0.0
train_snr_improvement = 0.0
total_samples = 0
batch_snr_improvements = []
for i, (noisy_spectrogram, clean_spectrogram) in enumerate(train_loader):
noisy_spectrogram = noisy_spectrogram.cuda(rank, non_blocking=True)
clean_spectrogram = clean_spectrogram.cuda(rank, non_blocking=True)
optimizer.zero_grad()
# Perform forward and backward pass with anomaly detection
with torch.amp.autocast(device_type='cuda'):
output = model(noisy_spectrogram)
loss = custom_loss_function(output, clean_spectrogram)
# Check for NaNs or Infs in loss
if torch.isnan(loss).any() or torch.isinf(loss).any():
print(f"NaN or Inf detected in loss at iteration {i}, epoch {epoch}")
continue
loss.backward()
# Apply gradient clipping
torch.nn.utils.clip_grad_norm_(model.parameters(), max_norm=1.0)
optimizer.step()
running_loss += loss.item()
# Calculate SNR Improvement
batch_snr_improvement = 0.0
for j in range(noisy_spectrogram.size(0)):
single_snr_improvement = snr_improvement(
noisy_spectrogram[j:j+1], clean_spectrogram[j:j+1], output[j:j+1]
).item()
batch_snr_improvement += single_snr_improvement
batch_snr_improvement /= noisy_spectrogram.size(0)
batch_snr_improvements.append(batch_snr_improvement)
total_samples += noisy_spectrogram.size(0)
# Aggregate batch SNR improvements
training_snr_improvement_avg = sum(batch_snr_improvements) / len(batch_snr_improvements)
training_snr_improvements.append(training_snr_improvement_avg)
print(f"Epoch {epoch+1}, Training SNR Improvement: {training_snr_improvement_avg}")
print(f"Epoch {epoch+1}, Total Samples Processed: {total_samples}")
# Validation phase
model.eval()
val_loss = 0.0
val_snr_improvement = 0.0
with torch.no_grad():
for noisy_spectrogram, clean_spectrogram in val_loader:
noisy_spectrogram = noisy_spectrogram.cuda(rank, non_blocking=True)
clean_spectrogram = clean_spectrogram.cuda(rank, non_blocking=True)
with torch.amp.autocast(device_type='cuda'):
output = model(noisy_spectrogram)
loss = custom_loss_function(output, clean_spectrogram)
val_loss += loss.item()
val_snr_improvement += snr_improvement(noisy_spectrogram, clean_spectrogram, output).item()
val_loss /= len(val_loader)
val_snr_improvement /= len(val_loader)
validation_snr_improvements.append(val_snr_improvement)
print(f"Epoch {epoch+1}, Validation Loss: {val_loss}, Validation SNR Improvement: {val_snr_improvement}")
model.train()
# Save the model every 50 epochs
if rank == 0:
if (epoch + 1) % 50 == 0:
torch.save(model.state_dict(), save_path)
print(f"Model saved at epoch {epoch+1}")
if val_loss < best_val_loss:
best_val_loss = val_loss
torch.save(model.state_dict(), best_save_path)
print(f"Best model saved at epoch {epoch+1} with validation loss {best_val_loss}")
if val_snr_improvement > best_val_snr_improvement:
best_val_snr_improvement = val_snr_improvement
# Step the learning rate scheduler
scheduler.step(val_loss)
if rank == 0:
print(f"Training complete for batch size {train_loader.batch_size}, learning rate {learning_rate}, epochs {num_epochs}")
print(f"Best Validation Loss: {best_val_loss}, Best Validation SNR Improvement: {best_val_snr_improvement}")
plot_metrics(training_snr_improvements, validation_snr_improvements, 'SNR Improvement')
except Exception as e:
print(f"Rank {rank} encountered an error: {e}")
finally:
torch.cuda.synchronize() # Ensure all operations are complete before cleanup
cleanup()
def setup(rank, world_size):
logger.info(f"Setting up distributed training on rank {rank}")
dist.init_process_group("nccl", rank=rank, world_size=world_size)
torch.cuda.set_device(rank)
# torch.distributed.barrier()
def cleanup():
try:
dist.destroy_process_group()
except Exception as e:
print(f"Error during cleanup: {e}")
def main_worker(rank, world_size, noisy_dir, clean_dir, save_dir, num_epochs, learning_rate, batch_size, checkpoint_path):
try:
setup(rank, world_size)
# Preprocessing the data
# preprocessor = Preprocessing(sample_rate=16000, n_fft=1024, hop_length=512, win_length=1024)
# preprocessor.create_dataset(noisy_dir, clean_dir, save_dir)
dataset = NoisySpeechDataset(os.path.join(save_dir, 'noisy'), os.path.join(save_dir, 'clean'), subset_size=50000)
train_size = int(0.8 * len(dataset))
val_size = len(dataset) - train_size
train_dataset, val_dataset = torch.utils.data.random_split(dataset, [train_size, val_size])
train_sampler = DistributedSampler(train_dataset, num_replicas=world_size, rank=rank)
val_sampler = DistributedSampler(val_dataset, num_replicas=world_size, rank=rank, shuffle=False)
train_loader = DataLoader(train_dataset, batch_size=batch_size, sampler=train_sampler, num_workers=2)
val_loader = DataLoader(val_dataset, batch_size=batch_size, sampler=val_sampler)
model = CRN()
save_path = f"/home/siddharth/Sid/ASR/ANC/DEEP_ANC_MODEL_trim_bs{batch_size}_lr{learning_rate}_ep{num_epochs}_og_trial.pth"
best_save_path = f"/home/siddharth/Sid/ASR/ANC/DEEP_ANC_MODEL_best_bs{batch_size}_lr{learning_rate}_ep{num_epochs}_og_trial.pth"
train_model(rank, world_size, model, train_loader, val_loader, num_epochs, learning_rate, save_path, best_save_path, checkpoint_path)
except Exception as e:
logger.error(f"An error occurred on rank {rank}: {e}")
finally:
cleanup()
|