File size: 6,839 Bytes
fedcb95 |
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 |
import torch
import config
import math
import sys
import os
from tqdm import tqdm
from torch.optim import AdamW
from transformers import AutoTokenizer
from diffusion import WrapESM, Diffusion
from data_loader import get_dataloaders
def save_hyperparams(ckpt_dir):
hyperparms_txt_file = os.path.join(ckpt_dir, "hyperparameters.txt")
with open(hyperparms_txt_file, 'w') as f:
for k, v in vars(config).items():
if k.isupper():
f.write(f"{k}: {v}\n")
def train_and_validate(model, optimizer, device, train_loader, val_loader, num_epochs, ckpt_dir):
best_val_loss = float('inf')
for epoch in range(num_epochs):
model.train()
print(f"EPOCH {epoch+1}/{num_epochs}")
sys.stderr.flush()
total_loss = 0.0
train_tokens = 0
weighted_total_train_loss = 0.0
train_update_interval = len(train_loader) // 4
with tqdm(enumerate(train_loader), desc="Training batch", total=len(train_loader), leave=True, position=0, ncols=100) as trainbar:
for step, inputs in trainbar:
inputs = {k: v.to(device) for k, v in inputs.items()}
optimizer.zero_grad()
outputs = model(**inputs)
train_loss = diffusion_model.compute_loss(inputs["input_ids"], inputs['attention_mask'],
val=False).loss
train_loss.backward()
optimizer.step()
total_loss += train_loss.item()
weighted_total_train_loss += train_loss.item() * inputs['input_ids'].shape[1] # Loss * sequence length
train_tokens += inputs['input_ids'].shape[1]
if (step+1) % train_update_interval == 0:
trainbar.update(train_update_interval)
avg_train_loss = total_loss / len(train_loader)
avg_train_neg_log_likelihood = weighted_total_train_loss / train_tokens
train_perplexity = math.exp(avg_train_neg_log_likelihood)
# Save model every epoch
train_ckpt_path = os.path.join(config.Eval.CHECKPOINT_PATH, f'epoch{epoch+1}')
model.save_model(train_ckpt_path)
save_hyperparams(train_ckpt_path)
# Validate model
if val_loader:
model.eval()
total_val_loss = 0.0
weighted_total_val_loss = 0.0
val_tokens = 0
with torch.no_grad():
val_update_interval = len(val_loader) // 4
with tqdm(enumerate(val_loader), desc='Validiation batch', total=len(val_loader), leave=True, position=0) as valbar:
for step, inputs in valbar:
inputs = {k: v.to(device) for k, v in inputs.items()}
outputs = model(**inputs)
val_loss = diffusion_model.compute_loss(inputs['input_ids'], inputs['attention_mask'],
val=True).loss.item()
total_val_loss += val_loss
weighted_total_val_loss += val_loss * inputs['input_ids'].shape[1] # Loss * sequence length
val_tokens += inputs['input_ids'].shape[1]
if (step+1) % val_update_interval == 0:
valbar.update(val_update_interval)
avg_val_loss = total_val_loss / len(val_loader)
avg_val_log_likelihood = weighted_total_val_loss / val_tokens
val_perplexity = math.exp(avg_val_log_likelihood)
# Save the best model based on validation loss
if avg_val_loss < best_val_loss:
best_val_loss = avg_val_loss
val_ckpt_path = os.path.join(config.Eval.CHECKPOINT_PATH, "best_model_epoch")
model.save_model(val_ckpt_path)
save_hyperparams(val_ckpt_path)
print(f"Average train loss: {avg_train_loss}")
print(f"Average train perplexity: {train_perplexity}\n")
sys.stdout.flush()
print(f"Average validation loss: {avg_val_loss}")
print(f"Average validation perplexity: {val_perplexity}\n")
sys.stdout.flush()
return avg_train_loss, train_perplexity, avg_val_loss, val_perplexity
def test(model, test_loader, device):
model.to(device).eval()
total_test_loss = 0.0
weighted_total_test_loss = 0.0
test_tokens = 0
with torch.no_grad():
for step, inputs in enumerate(test_loader):
inputs = {k: v.to(device) for k, v in inputs.items()}
outputs = model(**inputs)
test_loss = diffusion_model.compute_loss(inputs['input_ids'], inputs['attention_mask'],
val=True).loss.item()
total_test_loss += test_loss
weighted_total_test_loss += test_loss * inputs['input_ids'].shape[1] # loss * sequence length
test_tokens += inputs['input_ids'].shape[1]
avg_test_loss = total_test_loss / len(test_loader)
avg_test_log_likelihood = weighted_total_test_loss / test_tokens
test_perplexity = math.exp(avg_test_log_likelihood)
return avg_test_loss, test_perplexity
if __name__ == "__main__":
device = torch.device('cuda' if torch.cuda.is_available() else "cpu")
tokenizer = AutoTokenizer.from_pretrained(config.MODEL_NAME)
esm_model = WrapESM()
diffusion_model = Diffusion(config, tokenizer=tokenizer)
print(f'Trainable params before unfreezing: {sum(p.numel() for p in esm_model.parameters() if p.requires_grad)}')
esm_model.to(device)
diffusion_model.to(device)
esm_model.freeze_model()
esm_model.unfreeze_n_layers()
print(f'Trainable params after unfreezing: {sum(p.numel() for p in esm_model.parameters() if p.requires_grad)}')
train_loader, val_loader, test_loader = get_dataloaders(config)
optimizer = AdamW(filter(lambda p: p.requires_grad, esm_model.parameters()), lr=config.Optim.LR)
# Train and test the model
avg_train_loss, train_ppl, avg_val_loss, val_ppl = train_and_validate(esm_model, optimizer, device, train_loader, val_loader, config.Training.NUM_EPOCHS, config.Eval.CHECKPOINT_PATH)
avg_test_loss, test_ppl = test(esm_model, test_loader, device)
results_dict = {"Average train loss": avg_train_loss,
"Average train perplexity": train_ppl,
"Average val loss": avg_val_loss,
"Average val perplexity": val_ppl,
"Average test loss": avg_test_loss,
"Average test perplexity": test_ppl,
}
print("TRAIN AND TEST RESULTS")
for k, v in results_dict.items():
print(f"{k}: {v}\n") |