Load a model saved using trainer API

Beginners
1

Hi, I’m using this code to train (DAN is a model class, DataCollatorWithPadding is a Datacollector):

  co = DataCollatorWithPadding()
  training_args = TrainingArguments("DAN",
                                    num_train_epochs= 10, #must be at least 10.
                                    per_device_train_batch_size=32,
                                    per_device_eval_batch_size=4,
                                    learning_rate= 0.001,
                                    save_total_limit=1,
                                    log_level="error",
                                    evaluation_strategy="epoch")
  model = DAN()  
  trainer = Trainer(
      model=model,
      data_collator=co,
      args=training_args,
      callbacks = [],
      train_dataset=small_train_dataset,
      eval_dataset=small_eval_dataset,
      compute_metrics=compute_metrics,
  )
  trainer.train()

  preds = trainer.predict(small_eval_dataset)
  print(compute_metrics(preds))

The best model gets saved to DAN/checkpoint-5500.

I tried adding the resume_from_checkpoint="DAN/checkpoint-5500" flag to the trainer args, but it performed worse than it should (0.13 accuracy vs ~0.3)

Thanks