The "do-eval" option will rewrite trained model

[zjjj]

The "do-eval" option will rewrite trained model

I see, yea the saving of checkpoints occurs with evaluation. You can comment this out, add a --do-save flag, or change the name of the checkpoint file. I would recommend the third option, which was in the original code before I cleaned things up. For example, something like:

def save_pretrained(args, filepath=None):
    if filepath is None:
        filepath = os.path.join(args.output_dir, f'{args.task}_model_{args.seed}.pt')
    torch.save(self.state_dict(), filepath)
    print(f"Model weights saved in {filepath}")

In terms of loading a checkpoint, you can use any typical loading function. An example:

def load_checkpoint(model, load_dir, checkpoint_name):  
    ckpt_path = os.path.join(load_dir, checkpoint_name)
    checkpoint = torch.load(ckpt_path, map_location='cpu')
    model.load_state_dict(checkpoint)
    model.eval()
    model.to(device)
    return model

Hope that helps!