Simpletransformers: Save And load a model

[zbaida-achraf]

For MultiLabelClassificationModel
How can i save the model after the trainnig ?
And how i can load it?
I don't know why it takes a lot of time in the training (4H) for a 222K ?

[ThilinaRajapakse]

Please read the docs.

The models are saved automatically depending on the parameters you set in args.

You can load the model by providing the path to a saved model directory as model_name.

Do you mean for 222K training samples? What is your GPU?

[zbaida-achraf]

I know that it save automatically in outputs but i just want to know if i want to use it in other notebook how i can do it (for example torch.save to save a pytorch model) or i need to the full repository ( vocab, model.bin, config.json...)
Yes 222k samples for training.
I run it in kaggle (NVidia K80 GPUs)

[Shane-Neeley]

What is the minimum we need from the outputs to load a saved model for inference only? @ThilinaRajapakse .. attached screenshot is everything that was generated in outputs.

[ThilinaRajapakse]

@Zbaida I'm sorry, I seem to have missed your reply.

You don't need to manually torch.save(). You can copy the outputs folder to use it with another notebook. You can also copy a specific checkpoint or "best_model" directory. Essentially, any output directory with a pytorch_model.bin file is like a saved model.

You can safely delete/ignore the optimizer.pt and scheduler.pt files if you don't need to continue training from a checkpoint. Those are only needed for resuming training. It's best to keep all the other files though.

[aaronbriel]

What is the difference between 'best_model' and the model saved directly under 'outputs'? Is best_model the model that achieved best metrics during one of the epochs (perhaps not the last) while outputs stores the final model? If this is the case, what is the best way to ensure that only the best model is stored?

[ThilinaRajapakse]

best_model is the checkpoint with the best evaluation metric. You can prevent checkpoints from being saved by setting save_eval_checkpoints: False. However, you can't explicitly prevent the final model from being saved. I don't want to add more saving related flags to control this as I feel like there are already too many such flags.

You could set up your script to delete the final model once the training completes though.

[aaronbriel]

Thank you for the quick and detailed response. It is much appreciated!