break data, x_train, y_train, x_test, y_test, norm_constant = naf.one_file_training_data( actual_class_dir, file, forecast_length, backcast_length, batch_size, device) while i < limit: #difference > threshold and i += 1 global_step = naf.train_full_grad_steps( data, device, net, optimiser, test_losses, training_models + training_checkpoint, x_train.shape[0]) train_eval = naf.evaluate_training(backcast_length, forecast_length, net, norm_constant, test_losses, x_train, y_train, the_lowest_error, device) experiment.log_metric('train_loss', train_eval) new_eval = naf.evaluate_training(backcast_length, forecast_length, net, norm_constant, test_losses, x_test, y_test, the_lowest_error, device, plot_eval=False, class_dir=name,
while i < 2: #old was 5 #difference > threshold and i += 1 epoch += 1 print("Actual epoch: ", epoch, "\nActual inside file loop: ", i) global_step = train_full_grad_steps( data, device, net, optimiser, test_losses, training_models + training_checkpoint, x_train.shape[0]) train_eval = naf.evaluate_training(backcast_length, forecast_length, net, norm_constant, test_losses, x_train, y_train, the_lowest_error, device, experiment=experiment) experiment.log_metric('train_loss', train_eval) new_eval = naf.evaluate_training(backcast_length, forecast_length, net, norm_constant, test_losses, x_test, y_test, the_lowest_error, device,
iteration += 1 if iteration > 30 or difference < threshold: break data, x_train, y_train, x_test, y_test, norm_constant = naf.one_file_training_data( actual_class_dir, file, forecast_length, backcast_length, batch_size) while difference > threshold and i < limit: i += 1 global_step = naf.train_full_grad_steps( data, device, net, optimiser, test_losses, training_checkpoint, x_train.shape[0]) new_eval = naf.evaluate_training(backcast_length, forecast_length, net, norm_constant, test_losses, x_test, y_test, the_lowest_error, device) print( f"GlobalStep: {global_step}, New evaluation sccore: {new_eval}" ) if new_eval < old_eval: difference = old_eval - new_eval old_eval = new_eval with torch.no_grad(): print("New evaluation value:", new_eval, " iteration:", i) print("Saving...") new_checkpoint_name = str(checkpoint_name[:-3] + str(len(test_losses)) + ".th")