barbecacov/3rd_place_to_Kaggle_Google_Landmark_Retrieval_2020

This is the source code of the 3rd place to [Kaggle Google Landmark Retrieval 2020]

Introduction

This is the source code of the 3rd place to Kaggle Google Landmark Retrieval 2020.

A detailed description of our solution: https://arxiv.org/abs/2008.10480

Envs

• Pytorch 1.1
• Python 3.6.5
• CUDA 9.0
• Nvidia Tesla P40 * 8

install

pip install -r requirements.txt

apex : Tools for easy mixed precision and distributed training in Pytorch

$ git clone https://github.com/NVIDIA/apex
$ cd apex
$ pip install -v --no-cache-dir --global-option="--cpp_ext" --global-option="--cuda_ext" ./

Data

Data directory structure:

• data_list/label_81313.txt: GLD v2 clean, download from https://drive.google.com/file/d/1IH1nMml9GjIrKVloF9AZO06E28jnXP75/view?usp=sharing
• data_list/label_sub_212843.txt: GLD v2 cluster, download from https://drive.google.com/file/d/1pdlvnzXOt42wVDbxh9tlPtrjYIaIPVby/view?usp=sharing
• train: GLD v2 train, download from https://github.com/cvdfoundation/google-landmark

Download pretrained models (ImageNet)

• resnet-152: https://drive.google.com/file/d/1EqIqTlwaSp9Ta3fBZmzIKw-DkQp22pW-/view?usp=sharing
• resnest-200: https://drive.google.com/file/d/1l0IGCgP8zQuTFPxkIPp5GuX4DP5ltEye/view?usp=sharing

|-- data
  |-- data_list
    |-- label_81313.txt
    |-- label_sub_212843.txt
  |-- train # GLD v2
    |-- 0
    |-- 1
    ...
    |-- f
|-- pretrained_models
  |-- 7x7resnet152-imagenet.pth
  |-- resnest200-75117900.pth
...

Training

GPU_NUM: 8

# ResNest200
sh scripts/cls/resnest200_v2clean.sh 8

sh scripts/cls/resnest200_v2cluster.sh 8

# ResNet152
sh scripts/cls/resnet152_v2clean.sh 8

sh scripts/cls/resnet152_v2cluster.sh 8

Convert

pytorch -> onnx -> tensorflow

cd scripts/interface

sh merge_512.py

sh convert_merge.sh merge_second.onnx 512

final submission: scripts/interface/merge_second