Create train.py
Browse files
train.py
ADDED
@@ -0,0 +1,78 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import torch
|
2 |
+
import pickle
|
3 |
+
from transformers import MarianMTModel, MarianTokenizer, Seq2SeqTrainer, Seq2SeqTrainingArguments, DataCollatorForSeq2Seq
|
4 |
+
from datasets import load_dataset
|
5 |
+
from torch.utils.data import Dataset
|
6 |
+
|
7 |
+
# Load dataset (limit to 100 samples)
|
8 |
+
dataset = load_dataset("Helsinki-NLP/tatoeba_mt", "ara-eng")
|
9 |
+
train_data = dataset["test"].select(range(100)) # Use only first 100 samples
|
10 |
+
val_data = dataset["validation"].select(range(100))
|
11 |
+
|
12 |
+
# Load tokenizer and model
|
13 |
+
model_name = "Helsinki-NLP/opus-mt-ar-en"
|
14 |
+
tokenizer = MarianTokenizer.from_pretrained(model_name)
|
15 |
+
model = MarianMTModel.from_pretrained(model_name)
|
16 |
+
|
17 |
+
# Custom Dataset class
|
18 |
+
class TranslationDataset(Dataset):
|
19 |
+
def __init__(self, data, tokenizer, max_length=128):
|
20 |
+
self.data = data
|
21 |
+
self.tokenizer = tokenizer
|
22 |
+
self.max_length = max_length
|
23 |
+
|
24 |
+
def __len__(self):
|
25 |
+
return len(self.data)
|
26 |
+
|
27 |
+
def __getitem__(self, idx):
|
28 |
+
src_text = self.data[idx]["sourceString"]
|
29 |
+
tgt_text = self.data[idx]["targetString"]
|
30 |
+
src_encoded = self.tokenizer(src_text, truncation=True, max_length=self.max_length, padding="max_length", return_tensors="pt")
|
31 |
+
tgt_encoded = self.tokenizer(tgt_text, truncation=True, max_length=self.max_length, padding="max_length", return_tensors="pt")
|
32 |
+
return {
|
33 |
+
"input_ids": src_encoded["input_ids"].squeeze(0),
|
34 |
+
"attention_mask": src_encoded["attention_mask"].squeeze(0),
|
35 |
+
"labels": tgt_encoded["input_ids"].squeeze(0),
|
36 |
+
}
|
37 |
+
|
38 |
+
# Create dataset instances
|
39 |
+
train_dataset = TranslationDataset(train_data, tokenizer)
|
40 |
+
val_dataset = TranslationDataset(val_data, tokenizer)
|
41 |
+
|
42 |
+
# Data collator
|
43 |
+
data_collator = DataCollatorForSeq2Seq(tokenizer, model=model, padding=True)
|
44 |
+
|
45 |
+
# Training arguments (reduce epochs & batch size)
|
46 |
+
training_args = Seq2SeqTrainingArguments(
|
47 |
+
output_dir="./results",
|
48 |
+
evaluation_strategy="epoch",
|
49 |
+
save_strategy="epoch",
|
50 |
+
per_device_train_batch_size=8, # Reduce batch size
|
51 |
+
per_device_eval_batch_size=8,
|
52 |
+
learning_rate=5e-5,
|
53 |
+
weight_decay=0.01,
|
54 |
+
num_train_epochs=2, # Reduce epochs
|
55 |
+
logging_dir="./logs",
|
56 |
+
logging_steps=5, # Log more frequently
|
57 |
+
save_total_limit=1,
|
58 |
+
predict_with_generate=True,
|
59 |
+
)
|
60 |
+
|
61 |
+
# Trainer setup
|
62 |
+
trainer = Seq2SeqTrainer(
|
63 |
+
model=model,
|
64 |
+
args=training_args,
|
65 |
+
train_dataset=train_dataset,
|
66 |
+
eval_dataset=val_dataset,
|
67 |
+
tokenizer=tokenizer,
|
68 |
+
data_collator=data_collator,
|
69 |
+
)
|
70 |
+
|
71 |
+
# Train model
|
72 |
+
trainer.train()
|
73 |
+
|
74 |
+
# Save model
|
75 |
+
with open("nmt_model.pkl", "wb") as f:
|
76 |
+
pickle.dump(model, f)
|
77 |
+
|
78 |
+
print("Model training complete and saved as nmt_model.pkl")
|