https://www.kaggle.com/c/kuzushiji-recognition/
3rd place solution
See overview.md for an overview of my solution.
- python 3.6 or 3.7
- docker
- nvidia-docker
$ pip install -r requirements.txt
Download and place the dataset of this competition as below.
input
├── train_images/
├── test_images/
└── train.csv
and run preprocess.sh
$ bash preprocess.sh
Download the following 8 models and save it in the models directory
- 01_refine_efficientnet_b4_l2softmax_gray190-0060.model
- 02_refine_resnet152_l2softmax_gray112-0069.model
- 03_refine_seresnext101_l2softmax_rgb112-0080.model
- 04_refine_seresnext101_l2softmax_rgb112-0082.model
- 05_refine_resnet152_l2softmax_rgb112-0090.model
Run docker image and run test_detector.sh
$ ./run.sh
# bash test_detector.sh
... wait many hours ## takes about 5 hours on a GCP server with V100x8
Generates 2 prediction results
models/test_060000.pthmodels/test_100000.pth
If you uncommented validation secition of test_detector.sh, it generates more 2 prediction results.
This results is needed when you want to reproduce FalsePositive detection model.
models/val_060000.pthmodels/val_100000.pth
You can also download the generated object detection results from here.
$ bash scripts/auto_generate_per_model_results.sh
Generates 10 results
test_detector_060000_tta7_01_efficientnet_b4
test_detector_060000_tta7_02_resnet152
test_detector_060000_tta7_03_seresnext101
test_detector_060000_tta7_04_seresnext101
test_detector_060000_tta7_05_resnet152
test_detector_100000_tta7_01_efficientnet_b4
test_detector_100000_tta7_02_resnet152
test_detector_100000_tta7_03_seresnext101
test_detector_100000_tta7_04_seresnext101
test_detector_100000_tta7_05_resnet152
$ python scripts/results_fusion_test_060000.py
$ python scripts/results_fusion_test_100000.py
Generates 2 ensemble results
test_detector_060000_tta7_5models_hard_prob
test_detector_100000_tta7_5models_hard_prob
$ python scripts/results_nms_test.py
Generates nms results
test_nms030_tta7_5models_hard_prob
Remove FalsePositive bbox using LightGBM FalsePositive predictor and generate final submission csv.
$ python scripts/remove_false_positive_and_gen_csv.py
Generates final_submission.csv
Place images input/test_images with jpeg format.
$ ls input/test_images
kuzushiji_sample_01.jpg kuzushiji_sample_02.jpg
Replace input/test_images.list include your image filename without extension
$ cat input/test_images.list
kuzushiji_sample_01
kuzushiji_sample_02
$ python scripts/denoising_and_bens_preprocessing.py test
./input/denoised_test/kuzushiji_sample_01.png
./input/denoised_test/kuzushiji_sample_02.png
Remove if object detection results already exists.
$ rm models/test_060000.pth
$ rm models/test_100000.pth
Generate character detection results
$ bash test_detector.sh
Generate recognition results
If FileExistsError occurs, delete the target directory and re-execute
$ bash scripts/auto_generate_per_model_results.sh
$ python scripts/results_fusion_test_060000.py
$ python scripts/results_fusion_test_100000.py
$ python scripts/results_nms_test.py
$ python scripts/remove_false_positive_and_gen_csv.py
$ python scripts/plot_recognition_results.py test_nms030_tta7_5models_hard_prob input/test_images
saved: test_nms030_tta7_5models_hard_prob/kuzushiji_sample_01_with_results.jpg
saved: test_nms030_tta7_5models_hard_prob/kuzushiji_sample_02_with_results.jpg
Images containing recognition results are saved under test_nms030_tta7_5models_hard_prob directory.
Reproducing process of the three models of detection, classification, FalsePositive predictor.
Detection model use Faster R-CNN with:
- ResNet101 backbone
- Multi-scale train&test
- data augmentation (brightness, contrast, saturation, hue, random grayscale)
- no vertical and horizontal flip
use customized maskrcnn_benchmark
Training Detection model with all train images. And validation with public Leaderboard score.
See config for details. e2e_faster_rcnn_R_101_C4_1x_2_gpu_voc.yaml
Use docker, to reproduce character detection model.
$ bash build.sh ### build docker image
$ bash run.sh ### run docker container
# cd /work
# bash train_detector.sh
... wait a few hours
trained model is saved in the kuzushiji_recognition_R101_C4 directory.
To generate detection results, use test_detector.sh inside docker container.
$ ./run.sh
# bash test_detector.sh
... wait a few hours
Use scripts under train_scripts directory
$ bash train_scripts/01_efficientnet_b4_val15779_l2softmax_mixup_re_normalize_gray190.sh
...
For details, see train_scripts/README.md
Train FalsePositive predictor using results of validation data (val_nms030_tta7_5models_hard_prob)
Fix auto_generate_per_model_results.sh for validation (uncomment 6th line and comment out 7th line).
$ bash scripts/auto_generate_per_model_results.sh
Generates 10 validation results
val_detector_060000_tta7_01_efficientnet_b4
val_detector_060000_tta7_02_resnet152
val_detector_060000_tta7_03_seresnext101
val_detector_060000_tta7_04_seresnext101
val_detector_060000_tta7_05_resnet152
val_detector_100000_tta7_01_efficientnet_b4
val_detector_100000_tta7_02_resnet152
val_detector_100000_tta7_03_seresnext101
val_detector_100000_tta7_04_seresnext101
val_detector_100000_tta7_05_resnet152
Ensemble generated results
$ python scripts/results_fusion_val_060000.py
$ python scripts/results_fusion_val_100000.py
Generates 2 ensemble results
val_detector_060000_tta7_5models_hard_prob
val_detector_100000_tta7_5models_hard_prob
nms with 2 ensemble results
$ python scripts/results_nms_val.py
generates nms results
val_nms030_tta7_5models_hard_prob
Search hyper parameter with validation results.
$ python scripts/optuna_search_for_false_positive_detector.py
... wait a few hours
{'lambda_l1': 0.002050689306354841, 'lambda_l2': 0.49425078611198464, 'num_leaves': 203, 'feature_fraction': 0.8606773005600517, 'bagging_fraction': 0.9526576962122715, 'bagging_freq': 2, 'min_child_samples': 66}
and generates extracted feature data
val_nms030_tta7_5models_hard_prob.feather
Train LightGBM FalsePositive predictor reuse generated features at hyper parameter searching.
$ python scripts/gen_false_positive_detector.py`
...
saved: models/booster_for_val_nms030_tta7_5models_hard_prob.pkl
MIT