init

training_scripts/finetune_t5.py

@@ -148,7 +148,7 @@ def train(
                 args=Seq2SeqTrainingArguments(
                     num_train_epochs=epoch,
                     learning_rate=lr_tmp,
-                    output_dir=f'{output_dir}/runs',
+                    output_dir=f"{output_dir}/model_{n}",
                     evaluation_strategy='steps',
                     eval_steps=eval_steps,
                     per_device_eval_batch_size=eval_batch_size,
@@ -160,17 +160,21 @@ def train(
                 eval_dataset=tokenized_dataset['validation_ds'],
                 compute_metrics=compute_metric,
             )
-            os.makedirs(f'{output_dir}/model_{n}', exist_ok=True)
-            trainer.train()
-            trainer.save_model(f'{output_dir}/model_{n}')
-            tokenizer.save_pretrained(f'{output_dir}/model_{n}')
-            # grid search
-            with open(f'{output_dir}/model_{n}/hyperparameters.json', 'w') as f:
-                json.dump({"learning_rate": lr_tmp, "batch_size": batch_tmp}, f)
+            result = trainer.train()
+            trainer.save_model()  # Saves the tokenizer too for easy upload
+            metrics = result.metrics
+            trainer.log_metrics("train", metrics)
+            trainer.save_metrics("train", metrics)
+            trainer.save_state()
+
+            # trainer.save_model(f'{output_dir}/model_{n}')
+            # tokenizer.save_pretrained(f'{output_dir}/model_{n}')
+            # # grid search
+            # with open(f'{output_dir}/model_{n}/hyperparameters.json', 'w') as f:
+            #     json.dump({"learning_rate": lr_tmp, "batch_size": batch_tmp}, f)
             del trainer
             gc.collect()
             torch.cuda.empty_cache()
-            logging.info(f'model saved at {output_dir}/model_{n}')
     else:
         logging.info('skip hyperparameter search & model training (already done)')