How many training data is required for GAN?

Question

I'm beginning to study and implement GAN to generate more datasets. I'll just try to experiment with state-of-the-art GAN models as described here https://paperswithcode.com/sota/image-generation-on-cifar-10.

The problem is I don't have a big dataset (around 1.000) for image classification, I have tried to train and test my dataset with GoogleNet and InceptionV3 and the results are mediocre. I'm afraid that GAN will require a bigger dataset than the usual image classification. I couldn't find any detailed guideline of how to prepare datasets properly for GAN (e.g. minimum images).

So, how many images are required to produce a good GAN model?

Also, I'm curious whether if I can use my image classification dataset directly to train GAN.

Answer 1

There is no one answer to this. In general, the bigger your model is, the more data you will need to train it.

You tak about training a classifier. In this case it also depends on the difficulty (or obviousness) of the task, e.g. classifying white from red images versus detecting smile from frown.

As always with NN, you will need to try training and see if you have enough by looking at the results. If it does not work out, you have options:

• increase your dataset
• perform data augmentation (e.g. scale/rotate/etc.)
• leverage transfer learning. You mention GoogleNet. These were trained on millions of images on the ImageNet task. Leverage this. Use a pretrained model to represent your images. Than use that latent representation as input to your GAN. It will drastically reduce the amount of images that you need.