Code for 3rd place solution in Kaggle Human Protein Atlas Image Classification Challenge.
To read the detailed solution, please, refer to the Kaggle post
The following specs were used to create the original solution.
- Ubuntu 16.04 LTS
- Intel(R) Core(TM) i7-8700 CPU @ 3.20GHz
- 3x NVIDIA TitanX
To reproduct my submission without retrainig, do the following steps:
- Installation
- Download Official Image
- Make RGBY Images for official.
- Download Pretrained models
- Inference
- Make Submission
All requirements should be detailed in requirements.txt. Using Anaconda is strongly recommended.
conda create -n hpa python=3.6
source activate hpa
pip install -r requirements.txt
All required files except images are already in data directory. If you generate CSV files (duplicate image list, split, leak.. ), original files are overwritten. The contents will be changed, but It's not a problem.
After downloading and converting images, the data directory is structured as:
data
+- raw
| +- train
| +- test
| +- external
+- rgby
| +- train
| +- test
| +- external
Download and extract train.zip and test.zip to data/raw directory. If the Kaggle API is installed, run following command.
$ kaggle competitions download -c human-protein-atlas-image-classification -f train.zip
$ kaggle competitions download -c human-protein-atlas-image-classification -f test.zip
$ mkdir -p data/raw
$ unzip train.zip -d data/raw/train
$ unzip test.zip -d data/raw/test
To download external images, run following command. The external images will be located in data/raw/external
$ python tools/download.py
To train or inference, converting to RGBY image is required. Run following commands.
For official:
$ python tools/make_rgby.py --input_dir=data/raw/train --output_dir=data/rgby/train
$ python tools/make_rgby.py --input_dir=data/raw/test --output_dir=data/rgby/test
For external:
$ python tools/make_rgby.py --input_dir=data/raw/external --output_dir=data/rgby/external
You can skip this step. All CSV files are prepared in data directory.
There are duplicated images. To search them, run following commands. duplicates.ahash.csv and duplicates.phash.csv will be generated.
$ python tools/find_duplicate_images.py
Create 5 folds CV set. One for training, the other for searching augmentation. split.stratified.[0-4].csv and split.stratified.small.[0-4].csv will be generated.
$ python stratified_split.py
$ python stratified_split.py --use_external=0
To learn more about data leak, please, refer to this post. Following comand will create data_leak.ahash.csv and data_leak.phash.csv. The other leak is already in data directory.
$ python find_data_leak.py
In configs directory, you can find configurations I used train my final models. My final submission is ensemble of resnet34 x 5, inception-v3 and se-resnext50, but ensemble of inception-v3 and se-resnext50's performance is better.
To find suitable augmentation, 256x256 image and resnet18 are used. It takes about 2 days on TitanX. The result(best_policy.data) will be located in results/search directory. The policy that I used is located in data directory.
$ python train.py --config=configs/search.yml
To train models, run following commands.
$ python train.py --config={config_path}
To train all models, run sh train.sh
The expected training times are:
| Model | GPUs | Image size | Training Epochs | Training Time |
|---|---|---|---|---|
| resnet34 | 1x TitanX | 512 | 40 | 16 hours |
| inception-v3 | 3x TitanX | 1024 | 27 | 1day 15 hours |
| se-resnext50 | 2x TitanX | 1024 | 22 | 2days 15 hours |
To average weights, run following commands.
$ python swa.py --config={config_path}
To average weights of all models, simply run sh swa.sh
The averages weights will be located in results/{train_dir}/checkpoint.
You can download pretrained model that used for my submission from link. Or run following command.
$ wget https://www.dropbox.com/s/qo65gw8kml5hgag/results.tar.gz
$ tar xzvf results.tar.gz
Unzip them into results then you can see following structure:
results
+- resnet34.0.policy
| +- checkpoint
+- resnet34.1.policy
| +- checkpoint
+- resnet34.2.policy
| +- checkpoint
+- resnet34.3.policy
| +- checkpoint
+- resnet34.4.policy
| +- checkpoint
+- inceptionv3.attention.policy.per_image_norm.1024
| +- checkpoint
+- se_resnext50.attention.policy.per_image_norm.1024
| +- checkpoint
If trained weights are prepared, you can create files that contains class probabilities of images.
$ python inference.py \
--config={config_filepath} \
--num_tta={number_of_tta_images, 4 or 8} \
--output={output_filepath} \
--split={test or test_val}
To make submission, you must inference test and test_val splits. For example:
$ python inference.py --config=configs/resnet34.0.policy.yml --num_tta=8 --output=inferences/resnet34.0.test_val.csv --split=test_val
$ python inference.py --config=configs/resnet34.0.policy.yml --num_tta=8 --output=inferences/resnet34.0.test.csv --split=test
To inference all models, simply run sh inference.sh
Following command will ensemble of all models and make submissions.
$ python make_submission.py
If you don't want to use, modify make_submission.py. For example, if you want to use inception-v3 and se-resnext50 then modify test_val_filenames, test_filenames and weights in make_submission.py.
test_val_filenames = ['inferences/inceptionv3.0.test_val.csv',
'inferences/se_resnext50.0.test_val.csv']
test_filenames = ['inferences/inceptionv3.0.test.csv',
'inferences/se_resnext50.0.test.csv']
weights = [1.0, 1.0]
The command generate two files. One for original submission and the other is modified using data leak.
- submissions/submission.csv
- submissions/submission.csv.leak.csv