jingwei-round1-submit (team testB)

Introduction

This is a submission code of team testB in [Tianchi Competition: Jingwei], achieving 0.3484 on leaderboard. We use two models to get the result above, here is the implementation guide.

File structure

The file structure, following jingwei submission code standard, is as follow:

jingwei-round1-submit
|--README.md
|--submit
| |--result_model_2
| |--result_model_1
| |--result_final
|--data
| |--jingwei
| |--train_model_1
| |--train_model_2
| |--jingwei_round1_train_20190619.zip
| |--jingwei_round1_test_a_20190619.zip
|--code
| |--config
| |--dataloader
| |--model_1
| |--model_2
| |--utils

./data/jingwei contains ckpt and log file while training.
./data/train_model_1 is the training data for model_1.
./data/train_model_2 is the training data for model_2.

./code/config contains training config file.
./code/dataloader contains datasets and data augmentation codes.
./code/model_1 and ./code/model_2 contains the two models we use for competition.

./submit/result_model_1 contains test results of model_1.
./submit/result_model_2 contains test results of model_2.
./submit/result_final contains our final results.

Implementation of model_1

Dependencies

We use pytorch framework for training, before training, use requirement.txt to install dependencies.

cd code
pip install -r requirements_1.txt

We also use pretrained resnet50 weight file from https://download.pytorch.org/models/resnet50-19c8e357.pth. The file should be placed under ./data

Three NVIDIA-2080Ti GPUs and 128G RAM are used while training.

CUDA 10.1 cudnn 7.5.0 is needed

Creating training set

We generate our own training set from the official one. To obtain our training set, run python create_model_1_train.py
This would create a train_model_1 dir under ./data

Train and submit

To train model_1, run
python model_1_main.py --exp model_1_exp
To get predictions, run
python model_1_submit.py
Or simply run
./model_1_predict.sh

Until now you should be able to reproduce our result of model_1. Achieving test acc of 0.3230.

Implementation of model_2

Dependencies

We use pytorch framework for training, before training, use requirement.txt to install dependencies.

cd code/model_2
pip install -r requirements_2.txt

We also use pretrained resnet34 weight file from https://download.pytorch.org/models/resnet34-333f7ec4.pth. The file should be placed under ./data

Three NVIDIA-1080Ti GPUs and 64G RAM are used while training.

CUDA 9.1 cudnn 7.1.3 is needed

Train and submit

First move to model_2 directory: cd code/model_2
To train model_2, run
CUDA_VISIBLE_DEVICES=0,1,2 python main.py --backbone resnet --lr 0.007 --workers 12 --epochs 50 --batch-size 12 --gpu-ids 0,1,2 --checkname UNetResNet34 --eval-interval 1 --dataset jingwei --model-name UNetResNet34 --pretrained --base-size 600 --crop-size 512 --loss-type focal
To get predictions, run
python model_2_submit_1.py
and python model_2_submit_2.py

Until now you should be able to reproduce our result of model_2. Achieving test acc of 0.3034.

Combination of model_1 and model_2

To get our final submit, run
python final_result.py
You can find our final results at ./submit/result_final