mwdchang / tfjs-gan

Simple GAN example using tensorflow JS core

TensorflowJS GAN

A Generative Adversarial Network example generating hand-written digits, using just TensorflowJS APIs and vanilla javascript.

Source files

• index.html: UI
• gan.js: GAN code and hyper params
• image-util.js: Image utility
• data.js: Used to load MNIST dataset into compatible tensors, taken from tensorflow examples.
• weights.js: Pre-trained weights/biases for the Generator.

tfjs-gan's Issues

GPU memory crash

Hi @mwdchang, thank you for the awesome repo!

I modified your example to train with some black and white images of hand drawn curves, just for my own learning purposes.

My images are 50x50, so they are a bit bigger than the MNIST images. I'm running into a problem where Chrome's GPU process crashes after a couple of training iterations.

Looking at the Chrome task manager, I noticed that the GPU process retains memory and never garbage collects after training.

Do you have any suggestions? I was thinking about caching the weights but I was not sure how to do that.

Thanks again

[Question]: Make DC GAN

Hi man, great work. I'm trying to make a GAN or a DCGAN for learning.
Taking from your project, should i just replace the gen and disReal and have my deep convolution models there? I'm following a python project but sometimes it's hard to "port" the ideas to tfjs.

I want to input my own images for training and have it output a larger image. How big you would say it's feasible? 128, 256, more? It becomes computationally very expensive right?

Thanks,
Tiago Vasconcelos

Update to TFJS version 1.1.2

Great example. I really appreciate the live githubPages pure javascript format.

Getting zeroslike errors when updating to TFJS v1.1.2 any suggestions?

Works fine on Version 0.15.3, but breaks at Version 1.0.0