144 lines
4.3 KiB
Python
144 lines
4.3 KiB
Python
#! /usr/bin/env python3
|
|
from transformers import BartForConditionalGeneration, BartTokenizer, AdamW
|
|
import torch
|
|
from dl import load_dataset
|
|
from tqdm import tqdm
|
|
import os
|
|
|
|
|
|
class FT:
|
|
def __init__(self):
|
|
# Enable cudnn optimizations
|
|
torch.backends.cudnn.benchmark = True
|
|
|
|
self.device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
|
|
|
|
# load tokenizer and model
|
|
self.tokenizer = BartTokenizer.from_pretrained("facebook/bart-base")
|
|
self.model = BartForConditionalGeneration.from_pretrained("facebook/bart-base")
|
|
self.model.to(self.device)
|
|
|
|
# set up optimizer
|
|
self.optimizer = torch.optim.AdamW(self.model.parameters(), lr=1e-5)
|
|
|
|
try:
|
|
from torch.cuda.amp import GradScaler, autocast
|
|
|
|
self.scaler = GradScaler()
|
|
except ImportError:
|
|
|
|
class autocast:
|
|
def __enter__(self):
|
|
pass
|
|
|
|
def __exit__(self, *args):
|
|
pass
|
|
|
|
self.scaler = None # We won't use a scaler if we don't have Amp
|
|
|
|
def train_model(self, dataloader):
|
|
self.model.train()
|
|
total_loss = 0
|
|
print("Training model...")
|
|
for batch in tqdm(dataloader):
|
|
self.optimizer.zero_grad()
|
|
|
|
inputs = self.tokenizer(
|
|
batch[1],
|
|
return_tensors="pt",
|
|
padding=True,
|
|
truncation=True,
|
|
max_length=512,
|
|
)
|
|
inputs.to(self.device)
|
|
labels = self.tokenizer(
|
|
batch[0],
|
|
return_tensors="pt",
|
|
padding=True,
|
|
truncation=True,
|
|
max_length=512,
|
|
)
|
|
labels.to(self.device)
|
|
|
|
outputs = self.model(
|
|
input_ids=inputs["input_ids"],
|
|
attention_mask=inputs["attention_mask"],
|
|
labels=labels["input_ids"],
|
|
)
|
|
|
|
loss = outputs.loss
|
|
loss.backward()
|
|
|
|
self.optimizer.step()
|
|
total_loss += loss.item()
|
|
|
|
avg_train_loss = total_loss / len(dataloader)
|
|
return avg_train_loss
|
|
|
|
def test_model(self, dataloader):
|
|
self.model.eval()
|
|
total_loss = 0
|
|
print("Testing model...")
|
|
for batch in tqdm(dataloader):
|
|
with torch.no_grad():
|
|
inputs = self.tokenizer(
|
|
batch[1],
|
|
return_tensors="pt",
|
|
padding=True,
|
|
truncation=True,
|
|
max_length=512,
|
|
)
|
|
inputs.to(self.device)
|
|
labels = self.tokenizer(
|
|
batch[0],
|
|
return_tensors="pt",
|
|
padding=True,
|
|
truncation=True,
|
|
max_length=512,
|
|
)
|
|
labels.to(self.device)
|
|
outputs = self.model(
|
|
input_ids=inputs["input_ids"],
|
|
attention_mask=inputs["attention_mask"],
|
|
labels=labels["input_ids"],
|
|
)
|
|
loss = outputs.loss
|
|
total_loss += loss.item()
|
|
|
|
avg_test_loss = total_loss / len(dataloader)
|
|
return avg_test_loss
|
|
|
|
def save_checkpoint(self, epoch):
|
|
os.makedirs("./checkpoints/ft", exist_ok=True)
|
|
torch.save(
|
|
{
|
|
"epoch": epoch,
|
|
"model_state_dict": self.model.state_dict(),
|
|
"optimizer_state_dict": self.optimizer.state_dict(),
|
|
},
|
|
f"./checkpoints/ft/ft_{epoch}.pt",
|
|
)
|
|
|
|
def train(self):
|
|
train_dataloader, test_dataloader = load_dataset(
|
|
"../datasets/deu_mixed-typical_2011_1M/deu_mixed-typical_2011_1M-sentences.txt",
|
|
100,
|
|
100,
|
|
32,
|
|
test_ratio=0.2,
|
|
)
|
|
num_epochs = 3
|
|
for epoch in range(num_epochs):
|
|
avg_train_loss = self.train_model(train_dataloader)
|
|
print(f"Train loss for epoch {epoch+1}: {avg_train_loss}")
|
|
self.save_checkpoint(epoch)
|
|
print("Checkpoint saved!")
|
|
|
|
avg_test_loss = self.test_model(test_dataloader)
|
|
print(f"Test loss for epoch {epoch+1}: {avg_test_loss}")
|
|
|
|
|
|
if __name__ == "__main__":
|
|
trainer = FT()
|
|
trainer.train()
|