search

openai / glow

Code for reproducing results in "Glow: Generative Flow with Invertible 1x1 Convolutions"
https://arxiv.org/abs/1807.03039
MIT License
3.11k stars 515 forks source link

Overparametrization of the MNIST images #70

Open Tal-Golan opened 5 years ago

Tal-Golan commented 5 years ago

It seems that the MNIST images are tiled so they have three (identical) color channels, matching the CIFAR-10 format (data_loaders/get_mnist_cifar, lines 43-44:)

x_train = np.tile(np.reshape(x_train, (-1, 32, 32, 1)), (1, 1, 1, 3)) x_test = np.tile(np.reshape(x_test, (-1, 32, 32, 1)), (1, 1, 1, 3))

Since this triples the size of the latent representation, I wonder whether this is the best way to model MNIST images with Glow. Any reason why I shouldn't disable this tiling by removing the np.tile call?

sophiejwalton commented 5 years ago

Was the reason you used this method is so you can load a model with pretrained weights? Or was it for something else.

Tal-Golan commented 5 years ago

No pre-trained weights, training an MNIST model from scratch.