context.set_auto_parallel_context( device_num=device_num, parallel_mode=ParallelMode.DATA_PARALLEL, gradients_mean=True) init() else: rank = 0 device_num = 1 mindrecord_file = args.dataset_path if not os.path.exists(mindrecord_file): print("dataset file {} not exists, please check!".format( mindrecord_file)) raise ValueError(mindrecord_file) dataset = create_gru_dataset(epoch_count=config.num_epochs, batch_size=config.batch_size, dataset_path=mindrecord_file, rank_size=device_num, rank_id=rank) dataset_size = dataset.get_dataset_size() print("dataset size is {}".format(dataset_size)) network = Seq2Seq(config) network = GRUWithLossCell(network) lr = dynamic_lr(config, dataset_size) opt = Adam(network.trainable_params(), learning_rate=lr) scale_manager = DynamicLossScaleManager( init_loss_scale=config.init_loss_scale_value, scale_factor=config.scale_factor, scale_window=config.scale_window) update_cell = scale_manager.get_update_cell() netwithgrads = GRUTrainOneStepWithLossScaleCell(network, opt, update_cell)
type=int, default=1, help='Use device nums, default is 1') parser.add_argument('--result_path', type=str, default='./preprocess_Result/', help='result path') args = parser.parse_args() if __name__ == "__main__": mindrecord_file = args.dataset_path if not os.path.exists(mindrecord_file): print("dataset file {} not exists, please check!".format( mindrecord_file)) raise ValueError(mindrecord_file) dataset = create_gru_dataset(epoch_count=config.num_epochs, batch_size=config.eval_batch_size, \ dataset_path=mindrecord_file, rank_size=args.device_num, rank_id=0, do_shuffle=False, is_training=False) source_ids_path = os.path.join(args.result_path, "00_data") target_ids_path = os.path.join(args.result_path, "01_data") os.makedirs(source_ids_path) os.makedirs(target_ids_path) for i, data in enumerate( dataset.create_dict_iterator(output_numpy=True, num_epochs=1)): file_name = "gru_bs" + str( config.eval_batch_size) + "_" + str(i) + ".bin" data["source_ids"].tofile(os.path.join(source_ids_path, file_name)) data["target_ids"].tofile(os.path.join(target_ids_path, file_name)) print("=" * 20, "export bin files finished", "=" * 20)
str(cb_params.net_outputs[0].asnumpy()), str(cb_params.net_outputs[1].asnumpy()), str(cb_params.net_outputs[2].asnumpy()))) f.write('\n') if __name__ == '__main__': if args.run_distribute: rank = args.rank_id device_num = args.device_num context.set_auto_parallel_context(device_num=device_num, parallel_mode=ParallelMode.DATA_PARALLEL, gradients_mean=True) init() else: rank = 0 device_num = 1 dataset = create_gru_dataset(epoch_count=config.num_epochs, batch_size=config.batch_size, dataset_path=args.dataset_path, rank_size=device_num, rank_id=rank) dataset_size = dataset.get_dataset_size() print("dataset size is {}".format(dataset_size)) network = Seq2Seq(config) network = GRUWithLossCell(network) lr = dynamic_lr(config, dataset_size) opt = Adam(network.trainable_params(), learning_rate=lr) scale_manager = DynamicLossScaleManager(init_loss_scale=config.init_loss_scale_value, scale_factor=config.scale_factor, scale_window=config.scale_window) update_cell = scale_manager.get_update_cell() netwithgrads = GRUTrainOneStepWithLossScaleCell(network, opt, update_cell) time_cb = TimeMonitor(data_size=dataset_size) loss_cb = LossCallBack(rank_id=rank) cb = [time_cb, loss_cb]
def run_gru_eval(): """ Transformer evaluation. """ parser = argparse.ArgumentParser(description='GRU eval') parser.add_argument( "--device_target", type=str, default="Ascend", help="device where the code will be implemented, default is Ascend") parser.add_argument('--device_id', type=int, default=0, help='device id of GPU or Ascend, default is 0') parser.add_argument('--device_num', type=int, default=1, help='Use device nums, default is 1') parser.add_argument('--ckpt_file', type=str, default="", help='ckpt file path') parser.add_argument("--dataset_path", type=str, default="", help="Dataset path, default: f`sns.") args = parser.parse_args() context.set_context(mode=context.GRAPH_MODE, device_target=args.device_target, reserve_class_name_in_scope=False, \ device_id=args.device_id, save_graphs=False) dataset = create_gru_dataset(epoch_count=config.num_epochs, batch_size=config.eval_batch_size, \ dataset_path=args.dataset_path, rank_size=args.device_num, rank_id=0, do_shuffle=False, is_training=False) dataset_size = dataset.get_dataset_size() print("dataset size is {}".format(dataset_size)) network = Seq2Seq(config, is_training=False) network = GRUInferCell(network) network.set_train(False) if args.ckpt_file != "": parameter_dict = load_checkpoint(args.ckpt_file) load_param_into_net(network, parameter_dict) model = Model(network) predictions = [] source_sents = [] target_sents = [] eval_text_len = 0 for batch in dataset.create_dict_iterator(output_numpy=True, num_epochs=1): source_sents.append(batch["source_ids"]) target_sents.append(batch["target_ids"]) source_ids = Tensor(batch["source_ids"], mstype.int32) target_ids = Tensor(batch["target_ids"], mstype.int32) predicted_ids = model.predict(source_ids, target_ids) print("predicts is ", predicted_ids.asnumpy()) print("target_ids is ", target_ids) predictions.append(predicted_ids.asnumpy()) eval_text_len = eval_text_len + 1 f_output = open(config.output_file, 'w') f_target = open(config.target_file, "w") for batch_out, true_sentence in zip(predictions, target_sents): for i in range(config.eval_batch_size): target_ids = [str(x) for x in true_sentence[i].tolist()] f_target.write(" ".join(target_ids) + "\n") token_ids = [str(x) for x in batch_out[i].tolist()] f_output.write(" ".join(token_ids) + "\n") f_output.close() f_target.close()