First commit finetune - untested, even locally

codegen_model_comparison/src/finetune.py

@@ -0,0 +1,88 @@
+import argparse
+import logging
+import pickle
+
+import datasets
+import evaluate
+import torch
+from transformers import (AutoModelForCausalLM, AutoTokenizer,
+                          DataCollatorWithPadding, Trainer, TrainingArguments)
+
+
+def tokenize_function(example):
+    return tokenizer(example['text'],
+                     padding="max_length",
+                     truncation=True,
+                     max_length=500)
+
+
+def add_labels(example):
+    example['label'] = example['input_ids']
+    return example
+
+
+def main(args):
+    checkpoint = args.checkpoint
+    data_path = args.data_path
+    batch_size = args.batch_size
+    seq_length = args.seq_length
+
+    handler = logging.StreamHandler()
+    logger = logging.getLogger(__name__)
+    logger.addHandler(handler)
+    logger.setLevel(logging.INFO)
+
+    logger.info('Load data')
+    with open(data_path, "rb") as f:
+        data = pickle.load(f)
+
+    device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
+
+    logger.info('Load tokenizer and model from HF')
+    tokenizer = AutoTokenizer.from_pretrained(checkpoint)
+    tokenizer.pad_token = tokenizer.eos_token
+    model = AutoModelForCausalLM.from_pretrained(
+        checkpoint, trust_remote_code=True).to(device)
+
+    logger.info('Tokenize data')
+    tokenized_dataset = data.map(tokenize_function,
+                                 batched=True,
+                                 batch_size=batch_size).map(
+                                     add_labels,
+                                     batched=True,
+                                     batch_size=batch_size)
+
+    # Data collator - Assembles data into batches for training
+    data_collator = DataCollatorWithPadding(tokenizer=tokenizer)
+
+    training_args = TrainingArguments(output_dir='trainer_' + checkpoint,
+                                      evaluation_strategy="epoch")
+
+    trainer = Trainer(
+        model,
+        training_args,
+        train_dataset=tokenized_dataset['train'],
+        eval_dataset=tokenized_dataset['test'],
+        data_collator=data_collator,
+        #compute_metrics=compute_bleu_score,
+        tokenizer=tokenizer)
+
+    trainer.train()
+
+
+def parse_args():
+    parser = argparse.ArgumentParser()
+    parser.add_argument("--checkpoint", type=str)
+    parser.add_argument("--data_path", type=str)
+    parser.add_argument("--batch_size", type=int)
+    parser.add_argument("--seq_length", type=int)
+    args = parser.parse_args()
+
+    return args
+
+
+if __name__ == "__main__":
+
+    args = parse_args()
+
+    main(args)