How to train on a pretrained model?

Question

How to train on a pretrained model?

somiltg opened this issue 5 years ago · 3 comments

We are trying to transfer learn on pix2pix model by using the edges2shoes dataset and using --continue_train to train on our dataset. However, we receive an error regarding Discriminator saying that checkpoint is not present. Why is the checkpoint for discriminator not present in the pretrained model? Is it because it should only be used for testing and not training?

Answer 1 · 2019-12-02T20:00:00.000Z

We didn't save the discriminator. You need to train the discriminator from scratch.

Answer 2 · 2019-12-02T22:20:59.000Z

Do you mean we need to rerun the network on the entire Zappos dataset, because until we provide the checkpoint, we cannot use the pretrained model further. Can you suggest how we can go foorward with the method?

Answer 3 · 2019-12-03T05:54:43.000Z

We did provide the generator's weights. To fine-tune the model on your own dataset, you can initialize the generator with our weights, and train a discriminator from scratch on your own dataset. You don't need the original dataset.