Kaggle - Yelp Restaurant Photo Classification

Contest details: https://www.kaggle.com/c/yelp-restaurant-photo-classification

Thanks to kaggle competitor Nina Chen for providing starter code for this competition.

Explore raw data

0_explore_data.ipynb (uses R, not python)

Extract features using from pre-trained convolutional neural network

1_feature-extraction-bvlc_reference_caffenet.ipynb

• use Caffe framework
• pre-trained BVLC CaffeNet model, based on Alexnet
• extracts features from fc6, fc7, and softmax (fully-connected) layers

Remove path names from photo ids

2_process_h5_files.ipynb

• better to remove this step entirely by changing the way files are saved in the previous step

Visualize extracted features

3_feature_visualization.ipynb

• visualize accuracy of predicted class (from 1000-dim softmax probability vectors), though these vectors aren't used in the model training
• make t-SNE plots of fc6, fc7, and concatenated fc6+fc7 features, grouping by label
• multi-instance problem - average the fc6 and fc7 photo feature vectors by restaurant

Train SVM on extracted features

4_classification.ipynb

• concatenate fc6 and fc7 layer features
• train SVM model with rbf kernel (seems to perform better than linear SVM)
• multi-label problem - take simplest approach, a.k.a. "binary relevance", by training each label separately
• coarser-grained grid validation was performed, but omitted in this notebook (however, finer-grained validation results remain). Optimize parameters by validation on 20% of training set (more time-consuming cross-validation or optimization per label was not performed)