File size: 5,467 Bytes
a445ec4 | 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 | import torch
from utils.dataset import Speech2Text, speech_collate_fn
from models.model import TransformerTransducer
from tqdm import tqdm
from models.loss import RNNTLoss
import argparse
import yaml
import os
def train_one_epoch(model, dataloader, optimizer, criterion, device):
model.train()
total_loss = 0.0
progress_bar = tqdm(dataloader, desc="🔁 Training", leave=False)
for batch_idx, batch in enumerate(progress_bar):
speech = batch["fbank"].to(device)
text = batch["text"].to(device)
speech_mask = batch["fbank_mask"].to(device)
text_mask = batch["text_mask"].to(device)
fbank_len = batch["fbank_len"].to(device)
text_len = batch["text_len"].to(device)
optimizer.zero_grad()
output, _, _ = model(
speech=speech,
speech_mask=speech_mask,
text=text,
text_mask=text_mask,
)
# Bỏ <s> ở đầu nếu có
loss = criterion(output, text, fbank_len, text_len)
loss.backward()
optimizer.step()
total_loss += loss.item()
# === In loss từng batch ===
progress_bar.set_postfix(batch_loss=loss.item())
avg_loss = total_loss / len(dataloader)
print(f"✅ Average training loss: {avg_loss:.4f}")
return avg_loss
from torchaudio.functional import rnnt_loss
def evaluate(model, dataloader, criterion, device):
model.eval()
total_loss = 0.0
progress_bar = tqdm(dataloader, desc="🧪 Evaluating", leave=False)
with torch.no_grad():
for batch in progress_bar:
speech = batch["fbank"].to(device)
text = batch["text"].to(device)
speech_mask = batch["fbank_mask"].to(device)
text_mask = batch["text_mask"].to(device)
fbank_len = batch["fbank_len"].to(device)
text_len = batch["text_len"].to(device)
output, _, _ = model(
speech=speech,
speech_mask=speech_mask,
text=text,
text_mask=text_mask,
)
loss = criterion(output, text, fbank_len, text_len)
total_loss += loss.item()
progress_bar.set_postfix(batch_loss=loss.item())
avg_loss = total_loss / len(dataloader)
print(f"✅ Average validation loss: {avg_loss:.4f}")
return avg_loss
def load_config(config_path):
with open(config_path, 'r') as f:
return yaml.safe_load(f)
def main():
from torch.optim import Adam
parser = argparse.ArgumentParser()
parser.add_argument("--config", type=str, required=True, help="Path to YAML config file")
args = parser.parse_args()
config = load_config(args.config)
training_cfg = config['training']
optimizer_cfg = config['optimizer']
# ==== Load Dataset ====
train_dataset = Speech2Text(
json_path=training_cfg['train_path'],
vocab_path=training_cfg['vocab_path'],
)
train_loader = torch.utils.data.DataLoader(
train_dataset,
batch_size= training_cfg['batch_size'],
shuffle=True,
collate_fn = speech_collate_fn
)
dev_dataset = Speech2Text(
json_path=training_cfg['dev_path'],
vocab_path=training_cfg['vocab_path']
)
dev_loader = torch.utils.data.DataLoader(
dev_dataset,
batch_size= training_cfg['batch_size'],
shuffle=True,
collate_fn = speech_collate_fn
)
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
model = TransformerTransducer(
in_features=config['model']['in_features'],
n_classes=len(train_dataset.vocab),
n_layers=config['model']['n_layers'],
n_dec_layers=config['model']['n_dec_layers'],
d_model=config['model']['d_model'],
ff_size=config['model']['ff_size'],
h=config['model']['h'],
joint_size=config['model']['joint_size'],
enc_left_size=config['model']['enc_left_size'],
enc_right_size=config['model']['enc_right_size'],
dec_left_size=config['model']['dec_left_size'],
dec_right_size=config['model']['dec_right_size'],
p_dropout=config['model']['p_dropout']
).to(device)
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
model.to(device)
# === Khởi tạo loss ===
# Giả sử <blank> = 0, và bạn chưa dùng reduction 'mean' toàn bộ batch
criterion = RNNTLoss(config["rnnt_loss"]["blank"] , config["rnnt_loss"]["reduction"]) # hoặc "sum" nếu bạn custom average
# === Optimizer ===
optimizer = Adam(model.parameters(), lr=optimizer_cfg['lr'])
# === Huấn luyện ===
num_epochs = config["training"]["epochs"]
for epoch in range(1, num_epochs + 1):
train_loss = train_one_epoch(model, train_loader, optimizer, criterion, device)
val_loss = evaluate(model, dev_loader, criterion, device)
print(f"📘 Epoch {epoch}: Train Loss = {train_loss:.4f}, Val Loss = {val_loss:.4f}")
# Save model checkpoint
model_filename = os.path.join(
config['training']['save_path'],
f"transformer_transducer_epoch_{epoch}"
)
torch.save({
'epoch': epoch,
'model_state_dict': model.state_dict(),
'optimizer_state_dict': optimizer.state_dict(),
}, model_filename)
if __name__ == "__main__":
main()
|