Learning Deep Parsimonious Representations

This is the code for our NIPS'16 paper:

• Renjie Liao, Alexander Schwing, Richard S. Zemel, Raquel Urtasun. Learning Deep Parsimonious Representations. Neural Information Processing System, 2016.

Please cite the above paper if you use our code.

The code is released under the MIT license.

Data

The configuration of data is as below,

• CIFAR10 and CIFAR100:

  1, Download data from https://www.cs.toronto.edu/~kriz/cifar.html

  2, Change the key data_folder in exp_config.py as the unzipped path of data

• CUB-200-2011:

  1, Download data from http://www.vision.caltech.edu/visipedia/CUB-200-2011.html

  2, Preprocess (Crop + Resize + Subtract Mean) the images following the paper https://people.eecs.berkeley.edu/~nzhang/papers/eccv14_part.pdf

  3, Follow the example files CUB_train_list.txt and CUB_test_list.txt and specify the path of your own images

  4, Convert a pre-trained Alex-Net into Tensorflow format and specify the path in caffe_model_file of exp_config.py (referring to load_caffe_model function in AlexNet.py to convert the model correctly)

Training

Run python run_train_model.py <exp_id> to train a model.

Here exp_id should be one of the function names provided in exp_config.py. For example, setting exp_id to CIFAR10_sample_clustering, it will train a sample clustering model on CIFAR10 dataset.

Testing

Run python run_test_model.py <exp_id> to test a model.

You need to specify the test_model_name and test_folder in exp_config.py before run.

Zero-Shot

Run python run_zero_shot.py CUB_zero_shot to train and test a Struct SVM on top of the learned feature.

We provide the train/val/test split used in our experiment as the file zero_shot_split.npz.

You need to specify the test_model_name and test_folder in exp_config.py before run.

Notes

• The experiment results may differ slightly from what we reported in the paper, as the cross validation is performed based on random split of the training data.