Awesome

Inception Score for MNIST

Train a "perfect" MNIST classifier, and use it to compute inception score (ICP)

With our ICP implementation (pre-trained model saved in directory 'model'), the testing set of MNIST yields a score <img src="https://latex.codecogs.com/gif.latex?$\bf{9.8793~\pm~0.0614}$" />

Note that different pre-trained models may lead to slightly different inception scores.

Prerequisites: Tensorflow 1.0

The Format of Generated Images

The generated images are saved in a mat file, with a tensor named 'images' of size [10000,784], where 10000 is the number of images, and 784 is the dimension of a flattened MNIST image.

If you have multiple checkout points (each is a mat file) saved in a folder, you may specify the directory as

    # folders for generated images
    result_folder = './example_dir/'

    icp = []
    for k in range(50):
        k = k + 1
        mat = scipy.io.loadmat(result_folder+ '{}.mat'.format(str(k).zfill(3)))

If you have one checkout point saved in a mat file, you may specify the file as

    file_name = 'example.mat'
    mat = scipy.io.loadmat(result_folder+ file_name )

How to Use the Code: Evaluation, Re-train and Plot

To evaluate the ICP of generated images, run:

mnist_cnn_icp_eval.py

If you would like to re-train your classifier model, run:

mnist_cnn_train_slim.py

If you would like to plot your inception scores for multiple checkout points, run:

mnist_icp_plot.py

Citation

This code is used in the following paper:

@article{li2017alice,
  title={ALICE: Towards Understanding Adversarial Learning for Joint Distribution Matching},
  author={Li, Chunyuan and Liu, Hao and Chen, Changyou and Pu, Yunchen and Chen, Liqun and Henao, Ricardo and Carin, Lawrence},
  journal={Neural Information Processing Systems (NIPS)},
  year={2017}
}