iNaturalist 2021 - UFAM Team

6th Place Solution to iNat Challenge 2021: 10,000 Species Recognition Challenge with iNaturalist Data

Requirements

Prepare an environment with python=3.8, tensorflow=2.3.1

Dependencies can be installed using the following command:

pip install -r requirements.txt

Data

Please refer to the iNaturalist 2021 Competition Github page for additional dataset details and download links.

Use the script dataset_tools/create_inat2021_tf_records.py to generate the TFRecords files.

Training

To train a classifier use the script main.py. As long as our final submission has two training stages, you can use the script multi_stage_train.py:

python multi_stage_train.py --training_files=PATH_TO_BE_CONFIGURED/inat_train.record-?????-of-02240 \
    --num_training_instances=2686843 \
    --validation_files=PATH_TO_BE_CONFIGURED/inat_val.record-?????-of-00084 \
    --num_validation_instances=100000 \
    --num_classes=10000 \
    --model_name=efficientnet-b3 \
    --input_size=300 \
    --input_size_stage3=432 \
    --input_scale_mode=uint8 \
    --batch_size=32 \
    --lr_stage1=0.1 \
    --lr_stage2=0.1 \
    --lr_stage3=0.008 \
    --momentum=0.9 \
    --epochs_stage1=0 \
    --epochs_stage2=20 \
    --epochs_stage3=2 \
    --unfreeze_layers=18 \
    --label_smoothing=0.1 \
    --randaug_num_layers=6 \
    --randaug_magnitude=4 \
    --model_dir=PATH_TO_BE_CONFIGURED \
    --random_seed=42

The parameters can also be passed using a config file:

python multi_stage_train.py --flagfile=configs/efficientnet_b3_final_submission_training.config \
    --model_dir=PATH_TO_BE_CONFIGURED

For more parameter information, please refer to multi_stage_train.py or main.py. See configs folder for some training configs examples.

Training Geo Prior Model

To train geo prior model used on our final submission please see our TF implementation.

Prediction

To create a submission for the competition use script predict_main.py:

python predict_main.py --test_files=PATH_TO_BE_CONFIGURED/inat_public_test.record-?????-of-00417 \
    --num_classes=10000 \
    --model_name=efficientnet-b3 \
    --input_size=432 \
    --input_scale_mode=uint8 \
    --batch_size=8 \
    --ckpt_dir=PATH_TO_BE_CONFIGURED/ \
    --geo_prior_ckpt_dir=PATH_TO_BE_CONFIGURED/ \
    --submission_file_path=PATH_TO_BE_CONFIGURED/final_submission.csv \
    --use_tta

Results

Efficientnet-B3 was trained on iNat2021 train set, inference using input of 432x432. Geo Prior model was trained using coordinates and date info from iNat2021 train set.

Model name	Private Score
Efficientnet-B3	0.16756
Efficientnet-B3 + Geo Prior	0.10752
Efficientnet-B3 + Geo Prior + TTA	0.09894

Contact

If you have any questions, feel free to contact Fagner Cunha (e-mail: fagner.cunha@icomp.ufam.edu.br) or Github issues.

License

Apache License 2.0

sodawave / inat2021ufam