self-training-text-classification/finetuning.py [280:331]:
- - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - -
                    if checkpoints is None:
                        checkpoints = np.array([new_checkpoint])
                        eval_results = np.array([new_eval_result])
                        best_checkpoint = new_checkpoint
                        best_eval_result = new_eval_result
                    else:
                        if new_eval_result - best_eval_result > args.early_stopping_threshold:
                            best_checkpoint = new_checkpoint
                            best_eval_result = new_eval_result
                            early_stopping_patience_counter = 0
                        else:
                            if new_eval_result == best_eval_result:
                                best_checkpoint = new_checkpoint
                                best_eval_result = new_eval_result
                            early_stopping_patience_counter += 1

                        if early_stopping_patience_counter >= args.early_stopping_patience:
                            should_training_stop = True

                        checkpoints = np.append(checkpoints, [new_checkpoint], axis=0)
                        eval_results = np.append(eval_results, [new_eval_result], axis=0)
                        sorted_ids = np.argsort(eval_results)
                        eval_results = eval_results[sorted_ids]
                        checkpoints = checkpoints[sorted_ids]

                    if len(checkpoints) > args.keep_checkpoint_max:
                        # Delete the current worst checkpoint
                        checkpoint_to_remove, *checkpoints = checkpoints
                        eval_results = eval_results[1:]
                        if checkpoint_to_remove != new_checkpoint:
                            if accelerator.is_main_process:
                                shutil.rmtree(os.path.join(args.output_dir, checkpoint_to_remove), ignore_errors=True)
                            accelerator.wait_for_everyone()

                    if new_checkpoint in checkpoints:
                        # Save model checkpoint
                        checkpoint_output_dir = os.path.join(args.output_dir, new_checkpoint)
                        if accelerator.is_main_process:
                            if not os.path.exists(checkpoint_output_dir):
                                os.makedirs(checkpoint_output_dir)
                        accelerator.wait_for_everyone()
                        unwrapped_model = accelerator.unwrap_model(model)
                        unwrapped_model.save_pretrained(checkpoint_output_dir, save_function=accelerator.save)
                        if accelerator.is_main_process:
                            tokenizer.save_pretrained(checkpoint_output_dir)
                            logger.info("Saving model checkpoint to %s", checkpoint_output_dir)

            if completed_steps >= args.max_steps:
                break

            if should_training_stop:
                break
- - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - -



self-training-text-classification/finetuning.py [342:393]:
- - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - -
            if checkpoints is None:
                checkpoints = np.array([new_checkpoint])
                eval_results = np.array([new_eval_result])
                best_checkpoint = new_checkpoint
                best_eval_result = new_eval_result
            else:
                if new_eval_result - best_eval_result > args.early_stopping_threshold:
                    best_checkpoint = new_checkpoint
                    best_eval_result = new_eval_result
                    early_stopping_patience_counter = 0
                else:
                    if new_eval_result == best_eval_result:
                        best_checkpoint = new_checkpoint
                        best_eval_result = new_eval_result
                    early_stopping_patience_counter += 1

                if early_stopping_patience_counter >= args.early_stopping_patience:
                    should_training_stop = True

                checkpoints = np.append(checkpoints, [new_checkpoint], axis=0)
                eval_results = np.append(eval_results, [new_eval_result], axis=0)
                sorted_ids = np.argsort(eval_results)
                eval_results = eval_results[sorted_ids]
                checkpoints = checkpoints[sorted_ids]

            if len(checkpoints) > args.keep_checkpoint_max:
                # Delete the current worst checkpoint
                checkpoint_to_remove, *checkpoints = checkpoints
                eval_results = eval_results[1:]
                if checkpoint_to_remove != new_checkpoint:
                    if accelerator.is_main_process:
                        shutil.rmtree(os.path.join(args.output_dir, checkpoint_to_remove), ignore_errors=True)
                    accelerator.wait_for_everyone()

            if new_checkpoint in checkpoints:
                # Save model checkpoint
                checkpoint_output_dir = os.path.join(args.output_dir, new_checkpoint)
                if accelerator.is_main_process:
                    if not os.path.exists(checkpoint_output_dir):
                        os.makedirs(checkpoint_output_dir)
                accelerator.wait_for_everyone()
                unwrapped_model = accelerator.unwrap_model(model)
                unwrapped_model.save_pretrained(checkpoint_output_dir, save_function=accelerator.save)
                if accelerator.is_main_process:
                    tokenizer.save_pretrained(checkpoint_output_dir)
                    logger.info("Saving model checkpoint to %s", checkpoint_output_dir)

        if completed_steps >= args.max_steps:
            break

        if should_training_stop:
            break
- - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - -