iWildCam_2020 FGVC7

Top 3% (3/126) solution for iWildCam 2020 competition (Categorize animals in the wild), which is a part of the FGVC7 workshop at CVPR 2020

Requirements

• Python 3.6
• pytorch 1.4.0

About the Code

1. Prepare Data

Download the competition data from kaggle website

crop data

python fast_crop_image.py # crop data from images

prepare json for train/val

python prepare_data.py

prepare json for test data

python sort_images.py

2. Train the Model

1. for classification model(e.g. resnet, resnext, efficientnet...)

python train_model224.py -cfg configs/efficientNet.yaml

2. for NTS model

python train.py

3. Prediction

python infer224.py/infer.py

4. Train Cross Model

if you want to train K-cross validation model, and infer it

1. use gen_kcross.py to create kcross train.json/val.json
2. then train and infer it

# first set CROSS_VALIDATION as True in xxx.yaml, then 
python train_model224.py
python infer_crossmodel.py

5. Ensemble models

use model_ensemble.py

What Have not been released

1. data sample used for long-tail methods
2. auxiliary classifier head(for location)

Result Table

	train epochs	Private Score.	Public Score.
EfficientNet-B0	36	83.6	82.6
ResNet-50	36	83.5	82.3
NTS-Net	36	84.6	84.0
SEResnext101	36	82.6	82.8
Ensemble	36	84.7	84.5