kaggle-carvana-2017

Team "80 TFlops" solution to Carvana Image Masking Challenge on Kaggle. 4th place with TensorFlow-backed Keras.

Solution overview at Yandex (2017-10-07): YouTube (Russian), pdf (English).

Setup environment

Symlink your dataset folder (optional, you can pass an absolute path with --dataset-dir):

ln -s /path/to/dataset input

All of the data is expected to be in that folder, either as a file in the root or as a subdirectory.

Install all the required packages:

pip install -r requirements.txt

You should have Keras master installed:

pip install git+git://github.com/fchollet/keras.git --upgrade --no-deps

Run training with Resnet50 encoder

python train.py\
  --gpu=1\
  --seed=80\
  --dataset_dir='input'\
  --models_dir='models'\
  --network='resnet50'\
  --preprocessing_function='caffe'\
  --weights='weights/resnet-on-test-combined-8960.000010-0-0.0044540-99.7300751.h5'\
  --fold=1\
  --n_folds=5\
  --folds_source='folds.csv'\
  --val_mask_dir='gif_train_masks'\
  --train_mask_dir='gif_train_masks'

Run training with MobileNet encoder

python train.py\
  --gpu=1\
  --seed=80\
  --dataset_dir='input'\
  --models_dir='models'\
  --network='mobilenet'\
  --preprocessing_function='tf'\
  --weights='weights/mobilenet-on-test-6400.000010-5-0.0054162-99.7713900.h5'\
  --fold=1\
  --n_folds=5\
  --folds_source='folds.csv'\
  --val_mask_dir='gif_train_masks'\
  --train_mask_dir='gif_train_masks'

Run submission encoder

python generate_encoded_submission.py\
  --pred_mask_dir='predicted_masks/subfolder'\
  --submissions_dir='submissions'\
  --pred_threads=32

Run predict on train data

python predict_multithreaded.py \
  --gpu=0,1,2\
  --weights=weights/resnet-refine-fold_3-10240.000010-5-0.0033064-99.7107349.h5\
  --test_data_dir=input/train\
  --pred_mask_dir=predicted_masks/val_from_fold_3\
  --fold=3\
  --folds_source=folds.csv\
  --dataset_dir=input\
  --predict_on_val=True

or, alternatively

sh predict_train.sh

folds.csv

id,fold
424658d9f222,2

The columns should respectively contain a car id and a fold number. Important to notice:

• We had 5 of them to do 5-fold cross-validation.
• Since the cars were of a different size and colors, it was critical to keep each of 16 pictures of a single car in the same fold.
• Folds were created by sorting by the size of masks (i.e. total number of 1s vs 0s in the mask, a simple sum over the mask matrix).