mindspore/model_zoo/official/nlp/gru

• GRU
• • Paper
• Model Structure
• Dataset
• Environment Requirements
• • Requirements
• Quick Start
• Script Description
• • Dataset Preparation
  • Configuration File
  • Training Process
  • Inference Process
• Model Description
• • Performance
  • • Training Performance
    • Inference Performance
• Random Situation Description
• Others
• ModelZoo HomePage

• GRU
• Model Structure
• Dataset
• Environment Requirements
• Quick Start
• Script Description
  • Dataset Preparation
  • Configuration File
  • Training Process
  • Inference Process
• Model Description
  • Performance
    • Training Performance
    • Inference Performance
• Random Situation Description
• Others
• ModelZoo HomePage

GRU

GRU(Gate Recurrent Unit) is a kind of recurrent neural network algorithm, just like the LSTM(Long-Short Term Memory). It was proposed by Kyunghyun Cho, Bart van Merrienboer etc. in the article "Learning Phrase Representations using RNN Encoder-Decoder for Statistical Machine Translation" in 2014. In this paper, it proposes a novel neural network model called RNN Encoder-Decoder that consists of two recurrent neural networks (RNN).To improve the effect of translation task, we also refer to "Sequence to Sequence Learning with Neural Networks" and "Neural Machine Translation by Jointly Learning to Align and Translate".

Paper

1.Paper: "Learning Phrase Representations using RNN Encoder-Decoder for Statistical Machine Translation", 2014, Kyunghyun Cho, Bart van Merrienboer, Caglar Gulcehre, Dzmitry Bahdanau, Fethi Bougares, Holger Schwenk, Yoshua Bengio

2.Paper: "Sequence to Sequence Learning with Neural Networks", 2014, Ilya Sutskever, Oriol Vinyals, Quoc V. Le

3.Paper: "Neural Machine Translation by Jointly Learning to Align and Translate", 2014, Dzmitry Bahdanau, Kyunghyun Cho, Yoshua Bengio

Model Structure

The GRU model mainly consists of an Encoder and a Decoder.The Encoder is constructed with a bidirection GRU cell.The Decoder mainly contains an attention and a GRU cell.The input of the net is sequence of words (text or sentence), and the output of the net is the probability of each word in vocab, and we choose the maximum probability one as our prediction.

Dataset

In this model, we use the Multi30K dataset as our train and test dataset.As training dataset, it provides 29,000 respectively, each containing an German sentence and its English translation.For testing dataset, it provides 1000 German and English sentences.We also provide a preprocess script to tokenize the dataset and create the vocab file.

Environment Requirements

• Hardware（Ascend）
  • Prepare hardware environment with Ascend processor.
• Framework
  • MindSpore
• For more information, please check the resources below：
  • MindSpore Tutorials
  • MindSpore Python API

Requirements

nltk
numpy

To install nltk, you should install nltk as follow:

pip install nltk

Then you should download extra packages as follow:

import nltk
nltk.download()

Quick Start

After dataset preparation, you can start training and evaluation as follows:

# run training example
cd ./scripts
sh run_standalone_train.sh [TRAIN_DATASET_PATH]

# run distributed training example
sh run_distribute_train_ascend.sh [RANK_TABLE_FILE] [TRAIN_DATASET_PATH]

# run evaluation example
sh run_eval.sh [CKPT_FILE] [DATASET_PATH]

Script Description

The GRU network script and code result are as follows:

├── gru
  ├── README.md                              // Introduction of GRU model.
  ├── src
  |   ├──gru.py                              // gru cell architecture.
  │   ├──config.py                           // Configuration instance definition.
  │   ├──create_data.py                      // Dataset preparation.
  │   ├──dataset.py                          // Dataset loader to feed into model.
  │   ├──gru_for_infer.py                    // GRU eval model architecture.
  │   ├──gru_for_train.py                    // GRU train model architecture.
  │   ├──loss.py                             // Loss architecture.
  │   ├──lr_schedule.py                      // Learning rate scheduler.
  │   ├──parse_output.py                     // Parse output file.
  │   ├──preprocess.py                       // Dataset preprocess.
  │   ├──seq2seq.py                          // Seq2seq architecture.
  │   ├──tokenization.py                     // tokenization for the dataset.
  │   ├──weight_init.py                      // Initialize weights in the net.
  ├── scripts
  │   ├──create_dataset.sh                   // shell script for create dataset.
  │   ├──parse_output.sh                     // shell script for parse eval output file to calculate BLEU.
  │   ├──preprocess.sh                       // shell script for preprocess dataset.
  │   ├──run_distributed_train.sh            // shell script for distributed train on ascend.
  │   ├──run_eval.sh                         // shell script for standalone eval on ascend.
  │   ├──run_standalone_train.sh             // shell script for standalone eval on ascend.
  ├── eval.py                                // Infer API entry.
  ├── requirements.txt                       // Requirements of third party package.
  ├── train.py                               // Train API entry.

Dataset Preparation

Firstly, we should download the dataset from the WMT16 official net.After downloading the Multi30k dataset file, we get six dataset file, which is show as below.And we should in put the in same directory.

train.de
train.en
val.de
val.en
test.de
test.en

Then, we can use the scripts/preprocess.sh to tokenize the dataset file and get the vocab file.

bash preprocess.sh [DATASET_PATH]

After preprocess, we will get the dataset file which is suffix with ".tok" and two vocab file, which are nameed vocab.de and vocab.en.
Then we provided scripts/create_dataset.sh to create the dataset file which format is mindrecord.

bash preprocess.sh [DATASET_PATH] [OUTPUT_PATH]

Finally, we will get multi30k_train_mindrecord_0 ~ multi30k_train_mindrecord_8 as our train dataset, and multi30k_test_mindrecord as our test dataset.

Configuration File

Parameters for both training and evaluation can be set in config.py. All the datasets are using same parameter name, parameters value could be changed according the needs.

• Network Parameters

    "batch_size": 16,                  # batch size of input dataset.
    "src_vocab_size": 8154,            # source dataset vocabulary size.
    "trg_vocab_size": 6113,            # target dataset vocabulary size.
    "encoder_embedding_size": 256,     # encoder embedding size.
    "decoder_embedding_size": 256,     # decoder embedding size.
    "hidden_size": 512,                # hidden size of gru.
    "max_length": 32,                  # max sentence length.
    "num_epochs": 30,                  # total epoch.
    "save_checkpoint": True,           # whether save checkpoint file.
    "ckpt_epoch": 1,                   # frequence to save checkpoint file.
    "target_file": "target.txt",       # the target file.
    "output_file": "output.txt",       # the output file.
    "keep_checkpoint_max": 30,         # the maximum number of checkpoint file.
    "base_lr": 0.001,                  # init learning rate.
    "warmup_step": 300,                # warmup step.
    "momentum": 0.9,                   # momentum in optimizer.
    "init_loss_scale_value": 1024,     # init scale sense.
    'scale_factor': 2,                 # scale factor for dynamic loss scale.
    'scale_window': 2000,              # scale window for dynamic loss scale.
    "warmup_ratio": 1/3.0,             # warmup ratio.
    "teacher_force_ratio": 0.5         # teacher force ratio.
  

Training Process

• Start task training on a single device and run the shell script

  cd ./scripts
  sh run_standalone_train.sh [DATASET_PATH]
  

• Running scripts for distributed training of GRU. Task training on multiple device and run the following command in bash to be executed in scripts/:

  cd ./scripts
  sh run_distributed_train.sh [RANK_TABLE_PATH] [DATASET_PATH]
  

Inference Process

• Running scripts for evaluation of GRU. The commdan as below.

  cd ./scripts
  sh run_eval.sh [CKPT_FILE] [DATASET_PATH]
  

• After evalulation, we will get eval/target.txt and eval/output.txt.Then we can use scripts/parse_output.sh to get the translation.

  cp eval/*.txt ./
  sh parse_output.sh target.txt output.txt /path/vocab.en
  

• After parse output, we will get target.txt.forbleu and output.txt.forbleu.To calculate BLEU score, you may use this perl script and run following command to get the BLEU score.

  perl multi-bleu.perl target.txt.forbleu < output.txt.forbleu
  

Note: The DATASET_PATH is path to mindrecord. eg. train: /dataset_path/multi30k_train_mindrecord_0 eval: /dataset_path/multi30k_test_mindrecord

Model Description

Performance

Training Performance

Parameters	Ascend
Resource	Ascend 910
uploaded Date	01/18/2021 (month/day/year)
MindSpore Version	1.1.0
Dataset	Multi30k Dataset
Training Parameters	epoch=30, batch_size=16
Optimizer	Adam
Loss Function	NLLLoss
outputs	probability
Speed	50ms/step (1pcs)
Epoch Time	13.4s (1pcs)
Loss	2.5984
Params (M)	21
Checkpoint for inference	272M (.ckpt file)
Scripts	gru

Inference Performance

Parameters	Ascend
Resource	Ascend 910
Uploaded Date	01/18/2020 (month/day/year)
MindSpore Version	1.1.0
Dataset	Multi30K
batch_size	1
outputs	label index
Accuracy	BLEU: 30.30
Model for inference	272M (.ckpt file)

Random Situation Description

There only one random situation.

• Initialization of some model weights.

Some seeds have already been set in train.py to avoid the randomness of weight initialization.

Others

This model has been validated in the Ascend environment and is not validated on the CPU and GPU.

ModelZoo HomePage

Please check the official homepage