Instructions

Setup

The requirments to run the code in this repo are:

0. A valid AWS account
1. awscli installed pip install awscli
2. awscli configured aws configure
3. pywren pip install pywren
4. pywren setup pywren-setup
5. pytorch conda install pytorch

Fisher Vectors

This code is based off work at INRIA by Sanchez et Al. https://hal.inria.fr/hal-00830491v2/document. It is a python port of the pipeline described in their paper.

With 256 GMM centers an ImageNet trained model should get to 55% top-5 accuracy (35.1 top-1 accuracy) With 64 GMM centers an ImageNet trained model should get to 49% top-5 accuracy (28.3 top-1 accuracy) With 16 GMM centers an ImageNet trained model should get to 41.3% top-5 accuracy (21.2 top-1 accuracy)

To train + eval the model

1. Generate features from stored S3 keys usinn PyWren python featurize_fisher_model.py --num_centers 256 (if it errors run the following command to resume) python featurize_fisher_model.py --num_centers 256 --use_cache_gmm_sift --use_cache_gmm_lcs
2. Train model using least squares (file fv_model contains the weights of the model) python train_fisher_model.py fishervector_features.pickle fv_model

Random Features

This code is based off work by Coates & Ng: https://www-cs.stanford.edu/~acoates/papers/coatesng_nntot2012.pdf We remove the kmeans method in their paper and present a purely feedforward random approach.

With 256k random convolutional filters a Cifar-10 trained model should get to 84.2 top-1 accuracy. With 32k random convolutional filters a Cifar-10 trained model should get to 83.3 top-1 accuracy.

To train + eval the model

1. Just run the script python featurize_train_model.py