Skip to content

Distributed training

Introduction

The high performance of distributed training is one of the core advantages of PaddlePaddle. In the classification task, distributed training can achieve almost linear speedup ratio. Generally, OCR training task need massive training data. Such as recognition, PP-OCR v2.0 model is trained based on 1800W dataset, which is very time-consuming if using single machine. Therefore, the distributed training is used in PaddleOCR to speedup the training task. For more information about distributed training, please refer to distributed training quick start tutorial.

Quick Start

Training with single machine

Take recognition as an example. After the data is prepared locally, start the training task with the interface of paddle.distributed.launch. The start command as follows:

1
2
3
4
5
python3 -m paddle.distributed.launch \
    --log_dir=./log/ \
    --gpus "0,1,2,3,4,5,6,7" \
    tools/train.py \
    -c configs/rec/rec_mv3_none_bilstm_ctc.yml

Training with multi machine

Compared with single machine, training with multi machine only needs to add the parameter --ips to start command, which represents the IP list of machines used for distributed training, and the IP of different machines are separated by commas. The start command as follows:

1
2
3
4
5
6
7
ip_list="192.168.0.1,192.168.0.2"
python3 -m paddle.distributed.launch \
    --log_dir=./log/ \
    --ips="${ip_list}" \
    --gpus="0,1,2,3,4,5,6,7" \
    tools/train.py \
    -c configs/rec/rec_mv3_none_bilstm_ctc.yml

Notice:

  • The IP addresses of different machines need to be separated by commas, which can be queried through ifconfig or ipconfig.
  • Different machines need to be set to be secret free and can ping success with others directly, otherwise communication cannot establish between them.
  • The code, data and start command between different machines must be completely consistent and then all machines need to run start command. The first machine in the ip_list is set to trainer0, and so on.

Performance comparison

We conducted model training on 2x8 P40 GPUs. Accuracy, training time, and multi machine acceleration ratio of different models are shown below.

Model Configuration Configuration 8 GPU training time / Accuracy 3x8 GPU training time / Accuracy Acceleration ratio
CRNN rec_chinese_lite_train_v2.0.yml 260k Chinese dataset 2.50d/66.70% 1.67d/67.00% 1.5

We conducted model training on 3x8 V100 GPUs. Accuracy, training time, and multi machine acceleration ratio of different models are shown below.

Model Configuration Configuration 8 GPU training time / Accuracy 3x8 GPU training time / Accuracy Acceleration ratio
SLANet SLANet.yml PubTabNet 49.80h/76.20% 19.75h/74.77% 2.52

Note: when training with 3x8 GPUs, the single card batch size is unchanged compared with the 1x8 GPUs' training process, and the learning rate is multiplied by 2 (if it is multiplied by 3 by default, the accuracy is only 73.42%).

We conducted model training on 4x8 V100 GPUs. Accuracy, training time, and multi machine acceleration ratio of different models are shown below.

Model Configuration Configuration 8 GPU training time / Accuracy 4x8 GPU training time / Accuracy Acceleration ratio
SVTR ch_PP-OCRv3_rec_distillation.yml PP-OCRv3_rec data 10d/- 2.84d/74.00% 3.5

Comments