Kaggle Yelp Restaurant Photo Classification

Description of the my approach to this problem was published on the Kaggle's blog.
There is no click-get-submission-file script here, only set of the small scripts for different parts of the final solution.

/model/
folder contains photo-level feature extraction scripts with different pretrained netoworks/layers.

compress.py
Photo-level feature preprocessing: normalization or/and PCA transformation

fisher.cpp
Fisher Vectors computation, input: features generated by compress.py

vlad.cpp
VLAD descriptor computation, input: features generated by compress.py

pool.py
Business-level feature extraction. Input: features generated by compress.py in case of feature averaging, or from fisher.cpp/vlad.cpp in case of FV/VLAD features.

predict_test.py
Model training/prediction/submission file generation.
Input: features generated by different execution of pool.py script.