Intel & MobileODT Cervical Cancer Screening, kaggle competition

Dependencies

• pytorch 0.1.12, installed via pip or from source
• torchvision, installed from source
• torchsample, commit 00a41b05
• pip install tensorboard_logger, tqdm
• tensorflow >= 1.0 to use tensorboard (cpu version is enough)
• other necessary dependencies...

Steps to reproduce our submissions

Prepare data:

The folder ./data_utils contains codes to generate index files (all the idx files used for our models have been generated in ./data_utils) for train/val/test. Note that the images should be put following our folder structure:

# data_root
# ├── additional
# │   ├── Type_1
# │   ├── Type_2
# │   └── Type_3
# ├── test
# │   └── test
# └── train
#     ├── Type_1
#     ├── Type_2
#     └── Type_3

We resize all the original images to 256x256, please use data_utils/resize_convert_all_multi.py and change the paths in it to generate the resized data. We assume the resized_data root is ./data/resized_data/

train:

Use ./scripts/train_10_fold.py and change the arch to resnet101 and inception_v3 to train two kinds of models, respectively.

In this step, we will get model checkpoints in 20 different sub-folders in ./ckpt.

make submission (10-crop, 10-fold ensemble, version 1):

• 10 crop test for each single model: use ./scripts/gen_test_10_fold_cmds.py and change the arch to resnet101 and inception_v3, respectively. We will get 20 different submission files in ./submission, each of which corresponds to a single model.

• ensemble scores of all models (average): put all the paths of the submission files in a file, for example:

ls submission/resnet101_model_epoch_15_fold_* > resnet101_epoch15_10_fold.txt
ls submission/inception_v3_model_epoch_15_fold_* > inception_v3_epoch15_10_fold.txt
cat resnet101_epoch15_10_fold.txt inception_v3_epoch15_10_fold.txt > inception_v3_resnet101_epoch15_10_folds.txt
python ensemble_scores.py inception_v3_resnet101_epoch15_10_folds.txt

The final submission file is like ./submission/ensemble_submission_inception_v3_resnet101_epoch15_10_folds__{date_time}.txt (This scores 0.48662 on public Leaderboard, and 0.85024 on the private Leaderboard.)

make submission (10-crop + 40-crop, 10-fold ensemble, version 2):

• 10 crop test for each single model: use ./scripts/gen_test_10_fold_cmds.py and change the arch to resnet101 and inception_v3, respectively. We will get 20 different submission files in ./submission, each of which corresponds to a single model. (Same as version 1.)

• 40 crop test for each single model: use ./scripts/gen_test_10_fold_cmds_40_crop.py and change the arch to resnet101 and inception_v3, respectively. We will get 20 different submission files in ./submission, each of which corresponds to a single model.

• ensemble scores of all models (average): put all the paths of the submission files in a file, for example:

 # 10 crop
 ls submission/resnet101_model_epoch_15_fold_* > resnet101_epoch15_10_fold.txt
 ls submission/inception_v3_model_epoch_15_fold_* > inception_v3_epoch15_10_fold.txt
 cat resnet101_epoch15_10_fold.txt inception_v3_epoch15_10_fold.txt > inception_v3_resnet101_epoch15_10_folds.txt
 python ensemble_scores.py inception_v3_resnet101_epoch15_10_folds.txt

 # 40 crop
 ls submission/resnet101_model_epoch_15_fold_*_40* > 40crop_resnet101_epoch15_10_fold.txt
 ls submission/inception_v3_model_epoch_15_fold_*_40* > 40crop_inception_v3_epoch15_10_fold.txt
 cat 40crop_resnet101_epoch15_10_fold.txt 40crop_inception_v3_epoch15_10_fold.txt > 40crop_inception_v3_resnet101_epoch15_10_folds.txt
 python ensemble_scores.py 40crop_inception_v3_resnet101_epoch15_10_folds.txt

 # merge 10 and 40 crop (**NB**: 10crop is used twice)
 ls submission/ensemble_submission_inception_v3_resnet101_epoch15_10_folds*  submission/ensemble_submission_inception_v3_resnet101_epoch15_10_folds*  submission/ensemble_submission_40crop_inception_v3_resnet101_epoch15_10_folds* > 10_40_crop_res101_v3_epoch15.txt
 python ensemble_scores.py 10_40_crop_res101_v3_epoch15.txt

The final submission file is like ./submission/ensemble_submission_10_40_crop_res101_v3_epoch15__{date_time}.txt (This scores 0.48592 on the public Leaderboard, and 0.83833 on the private Leaderboard.)

Credits

Written and maintained by:

• Gu Wang
• Yu Yang
• Shi Yan
• Junjie Wu