Skip to content

Latest commit

 

History

History
 
 

bert-base

README.md

Paddle-BERT with Graphcore IPUs

Table of contents

Overview

This project aims to build BERT-Base pre-training and SQuAD fine-tuning task using PaddlePaddle on Graphcore IPU-POD16.

File Structure

File Description
README.md How to run the model.
run_pretrain.py The algorithm script to run pre-training tasks (phase1 and phase2).
run_squad.py The algorithm script to run SQuAD fine-tuning task.
run_squad_infer.py The algorithm script to run SQuAD validation task.
modeling.py The algorithm script to build the Bert-Base model.
dataset_ipu.py The algorithm script to load input data in pre-training.
run_stage.sh Test script to run single stage (phase1, phase2, SQuAD and validation).
run_all.sh Test script to run all of stages.
LICENSE The license of Apache.

Dataset

  1. Pre-training dataset

    Refer to the Wikipedia dataset generator provided by Nvidia (https://github.com/NVIDIA/DeepLearningExamples.git).

    Generate sequence_length=128 and 384 datasets for pre-training phase1 and phase2 respectively.

    Code base：https://github.com/NVIDIA/DeepLearningExamples/tree/88eb3cff2f03dad85035621d041e23a14345999e/TensorFlow/LanguageModeling/BERT

git clone https://github.com/NVIDIA/DeepLearningExamples.git

cd DeepLearningExamples/TensorFlow/LanguageModeling/BERT

bash scripts/docker/build.sh

cd data/

vim create_datasets_from_start.sh

Modified the line 40 `--max_seq_length 512` as `--max_seq_length 384`, and the line 41 `--max_predictions_per_seq 80` as `--max_predictions_per_seq 56`.

cd ../

bash scripts/data_download.sh wiki_only

  2. SQuAD 1.1 dataset

    curl --create-dirs -L https://rajpurkar.github.io/SQuAD-explorer/dataset/train-v1.1.json -o data/squad/train-v1.1.json

curl --create-dirs -L https://rajpurkar.github.io/SQuAD-explorer/dataset/dev-v1.1.json -o data/squad/dev-v1.1.json

Changelog

Changed

Added

  • Added README.md to introduce how to run the Bert-Base model.
  • Added run_squad.py to run the SQuAD fine-tuning task.
  • Added run_squad_infer.py to run the SQuAD validation.
  • Added dataset_ipu.py to load input data in pre-training.
  • Added run_stage.sh to run the single task (phase1, phase2, SQuAD and validation).
  • Added run_all.sh to run the complete process.

Quick Start Guide

The Paddle-Bert project depends on Poplar SDK, Docker, Paddlepaddle and PaddleNLP. Please follow the instructions below to prepare the environment and run the model.

1）Prepare Project Environment

Poplar SDK

Install the Poplar SDK following the instructions in the Getting Started guide for your IPU system. Make sure to source the enable.sh scripts for Poplar and PopART.

SDK version: poplar_sdk-ubuntu_18_04-2.3.0+774-b47c577c2a

Docker

git clone -b paddle_bert_release https://github.com/graphcore/Paddle.git

cd Paddle

# build docker image

docker build -t paddlepaddle/paddle:dev-ipu-2.3.0 -f tools/dockerfile/Dockerfile.ipu .

# The ipu.conf is required here. if the ipu.conf is available, please make sure `${HOST_IPUOF_PATH}` is the right dir of the ipu.conf. 
# If the ipu.conf is not available, please follow the instruction below to generate a ipu.conf.

vipu create partition ipu --size 16

# Then the ipu.conf is able to be found in the dir below.

ls ~/.ipuof.conf.d/

# create container

docker run --ulimit memlock=-1:-1 --net=host --cap-add=IPC_LOCK \
--device=/dev/infiniband/ --ipc=host --name paddle-ipu-dev \
-v ${HOST_IPUOF_PATH}:/ipuof \
-e IPUOF_CONFIG_PATH=/ipuof/ipu.conf \
-it paddlepaddle/paddle:dev-ipu-2.3.0 bash

All of later processes are required to be executed in the container.

Requirements

The following requirements are required by Paddlepaddle and PaddleNLP.

pip3.7 install jieba h5py colorlog colorama seqeval multiprocess numpy==1.19.2 paddlefsl==1.0.0 six==1.13.0

The following requirements are required by the Paddle-Bert.

pip3.7 install wandb

pip3.7 install torch==1.7.0+cpu -f https://download.pytorch.org/whl/cpu/torch_stable.html

pip3.7 install torch-xla@https://storage.googleapis.com/tpu-pytorch/wheels/torch_xla-1.7-cp37-cp37m-linux_x86_64.whl

The Compile and installation of Paddlepaddle

git clone -b paddle_bert_release https://github.com/graphcore/Paddle.git

cd Paddle

# `${POPLAR_DIR}` and `${POPART_DIR} are the directories of the Poplar SDK and the PopART SDK respectively.

cmake -DPYTHON_EXECUTABLE=/usr/bin/python \
-DWITH_PYTHON=ON -DWITH_IPU=ON -DPOPLAR_DIR=`${POPLAR_DIR}` \
-DPOPART_DIR=`${POPART_DIR}` -G "Unix Makefiles" -H`pwd` -B`pwd`/build

cmake --build `pwd`/build --config Release --target paddle_python -j$(nproc)

pip3.7 install -U build/python/dist/paddlepaddle-0.0.0-cp37-cp37m-linux_x86_64.whl

The installation of PaddleNLP

pip3.7 install git+https://github.com/graphcore/PaddleNLP.git@paddle_bert_release

2) Execution

Run the single task (optional)

Please check the --input_dir in run_stage.sh and make sure the dir of input data is right.

The type of the input data in Phase1(Pre-training) is the tfrecord (sequence_length = 128).

The type of the input data in Phase2(Pre-training) is the tfrecord (sequence_length = 384).

The name of the input data in Fine-tuning is the train-v1.1.json.

The name of the input data in Validation is the dev-v1.1.json.

  • Run pre-training phase1 (sequence_length = 128)
./run_stage.sh ipu phase1 _ pretrained_128_model
  • Run pre-training phase2 (sequence_length = 384)
./run_stage.sh ipu phase2 pretrained_128_model pretrained_384_model
  • Run SQuAD fine-tuning task
./run_stage.sh ipu SQuAD pretrained_384_model finetune_model
  • Run SQuAD validation
./run_stage.sh ipu validation finetune_model _

Run the complete process (optional)

./run_all.sh

Parameters

  • model_type The type of the NLP model.
  • model_name_or_path The model configuration.
  • input_dir The directory of the input data.
  • output_dir The directory of the trained models.
  • seq_len The sequence length.
  • max_predictions_per_seq The max number of the masked token each sentence.
  • learning_rate The learning rate for training.
  • weight_decay The weight decay.
  • max_steps The max training steps.
  • warmup_steps The warmup steps used to update learning rate with lr_schedule.
  • logging_steps The gap steps of logging.
  • seed The random seed.
  • device The type of device. 'ipu': Graphcore IPU, 'cpu': CPU.
  • num_ipus The number of IPUs.
  • num_hidden_layers The number of encoder layers.
  • micro_batch_size The batch size of the IPU graph.
  • ipu_enable_fp16 Enable FP16 or not.
  • scale_loss The loss scaling.
  • save_init_onnx Save the initial onnx graph or not.
  • save_per_n_step Sync the weights D2H every n steps.
  • save_steps Save the paddle model every n steps.
  • optimizer_type The type of the optimizer.
  • enable_pipelining Enable pipelining or not.
  • batches_per_step The number of batches per step with pipelining.
  • enable_replica Enable graph replication or not.
  • num_replica The number of the graph replication.
  • enable_grad_acc Enable gradiant accumulation or not.
  • grad_acc_factor Update the weights every n batches.
  • batch_size total batch size (= batches_per_step * num_replica * grad_acc_factor * micro_batch_size).
  • enable_recomputeEnable recompute or not.
  • enable_half_partial Enable matmul fp16 partial or not.
  • available_mem_proportion The available proportion of memory used by conv or matmul.
  • check_data Enable to check input data or not.
  • ignore_index The ignore index for the masked position.
  • hidden_dropout_prob The probability of the hidden dropout.
  • attention_probs_dropout_prob The probability of the attention dropout.
  • is_training Training or inference.

Result

Task Metric Result
Phase1 MLM Loss 1.623
NSP Loss 0.02729
MLM Acc 0.668
NSP Acc 0.9893
tput 9200
Phase2 MLM Loss 1.527
NSP Loss 0.01955
MLM Acc 0.6826
NSP Acc 0.9927
tput 2700
SQuAD EM 80.48249
F1 87.556685

Licensing

The code presented here is licensed under the Apache License Version 2.0, see the LICENSE file in this directory.

This directory includes derived work from the following:

PaddlePaddle/PaddleNLP, https://github.com/PaddlePaddle/PaddleNLP/blob/develop/paddlenlp/transformers/bert/modeling.py

PaddlePaddle/PaddleNLP, https://github.com/PaddlePaddle/PaddleNLP/blob/develop/examples/language_model/bert/static/run_pretrain.py

Copyright (c) 2020 PaddlePaddle Authors. All Rights Reserved.

Licensed under the Apache License, Version 2.0 (the "License"); you may not use this file except in compliance with the License. You may obtain a copy of the License at

http://www.apache.org/licenses/LICENSE-2.0

Unless required by applicable law or agreed to in writing, software distributed under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the specific language governing permissions and limitations under the License.