imclab / yelp

Yelp Restaurant Photo Classification - Kaggle competition

Yelp Restaurant Photo Classification

My solution that scored 0.82246 and finished Yelp Restaurant Photo Classification on 22 position out of 355 teams (top 10%).

Implements the following pipeline:

• Extract image features using four Caffe models. Results in eight 9GB image feature files (two train/test sets for every model)
• Combine image features into business features in two different ways. Results in 16 feature sets, much smaller this time
• Train Logistic Regression classifiers on business features via 10-Fold CV on every business feature set. Stack classifier output into single set of meta features. Train Logistc Regression on meta features and generate an intermediate submission
• Train Neural Net and XGBoost on meta features, predict label probabilities, average predictions, generate submission without F1 Maximization
• Adjust probabilities via Maximum Expected Utility Framework for F-Measure Maximization, generate final submission.

Requires:

• Caffe, deep learning framework
• Scientific Python Stack (including NumPy, SciPy, Pandas. All this can be obtaned with Anaconda distribution)
• XGBoost
• Theano
• Keras
• About 100 GB of free disk space is needed for train/test images, extracted image features, model dumps.

NVIDIA GPU is not required but recommended. Extracting image features on CPU may take several days.

Download:

• The training and test datasets and other data can be downloaded from here
• Get pretrained Caffe models BVLC Reference CaffeNet and BVLC AlexNet as described here. Download the other two models from Places CNN project: Places205-AlexNet, Hybrid-AlexNet

Notice: customized prototxts and mean files already available in the models folder

How to generate the solution(s):

1. After you downloaded and extracted datasets and models, adjust paths in paths.py and set caffe_mode (currently set to CPU)
2. Successively run (make sure you have enough disk space, see above):
   python Stage1_ExtractImageFeatures.py
   python Stage2_CreateBuisnessFeatures.py
   python Stage3_BlendLRModelsCV.py
   python Stage4_KerasXGBoostMEUFsubmission.py
3. You will get three submissions:
   all_models_blendLR_CV.csv
   keras_xgboost_blend_noMEUF.csv
   keras_xgboost_blend_MEUF.csv

Enjoy!

Read my article "What restaurant would your computer like to go to?" and like it if you like it.