Experiments on MNIST Database¶

The MNIST database is a large database of handwritten digits commonly used to test image processing systems.

Preparation¶

To download the raw MNIST data set with the default parameters, run:

python -m datasets.mnist.get_raw -m sacred

In this case, the process is very simple. We use Keras to download the MNIST data to ~/.keras/datasets/mnist.npz.

To serialize the raw data use:

python -m datasets.mnist.serialize -m sacred

The following experiment shows how to generate random numbers between 0 and 9 with the GanEstimator module. For implementation details see Components of Reproducible-ML or API Documentation

To run the code with the default parameters defined in the config function just type from the root directory:

python -m exps.mnist.train

Or use the command line to set different parameters:

python -m exps.mnist.train with mnist_feeder.batch_size=16 gan_type="UNCOND" -m sacred

With the default parameters we will obtain the following result for training steps 100, 400 and 2000:

In this experiment, we wish to condition on both the generator and discriminator. We generate MNIST digits conditioned on class labels.

To run the conditioned MNIST GAN use:

python -m exps.mnist.gan with mnist_feeder.batch_size=16 gan_type="COND" -m sacred

With the default parameters we will obtain the following result for training steps 100, 400 and 2000: