Longer training and higher resolution images? about deep-makeover

[david-gpu]

What would i need to change in the files to let the training run longer, to get better results?

Look at the --train_time option in dm_flags.py. You can change its value through the command line options.

It was also interesting that my gtx 780 (like your 1080) only took 2 hours to complete the training.

It trains for two hours by default. You can change the training time as indicated above.

Does a longer training get you clearer pictures (less artifacts)?

Up to a point. After a while the GAN tends to produce less diverse images. The sweet spot appears to be around 40,000 to 50,000 batches.

Would it be possible to generate images with 2x or larger resolution? I know you would also need high res training images

Right, you would need a larger dataset. The dataset loading code is in dm_celeba.py and dm_input.py. With larger input images you will also need a larger model and more training time.

Pictures with full head, not only cropped below hair

There are good reasons why the images were cropped. You can play with it by modifying dm_input.py.

Input arguments for image generation

Please read dm_flags.py.

[subzerofun]

Thanks for the quick reply! Should have just looked in dm_flags.py 🙃

Looked through the code to see if i can change the training/output resolution myself and the only obvious place i found is:

    size_x, size_y = 80, 100

in dm_input.py. And of course the vars off_x, off_y for cropping.

The dm_model.py code is a little bit too much for me – i don't have that much experience with tensorflow or neural net models in general (i know i could start with the tensorflow beginner tutorials, the docs are quite extensive). But to save time could you please highlight the lines of dm_model.py that i should look into?

---

PS: If you haven't seen this yet, this paper maybe could be interesting for you – it tries to do a similar thing like your srez project:

Pixel Recursive Super Resolution

Abstract: We present a pixel recursive super resolution model that synthesizes realistic details into images while enhancing their resolution. A low resolution image may correspond to multiple plausible high resolution images, thus modeling the super resolution process with a pixel independent conditional model often results in averaging different details--hence blurry edges. By contrast, our model is able to represent a multimodal conditional distribution by properly modeling the statistical dependencies among the high resolution image pixels, conditioned on a low resolution input. We employ a PixelCNN architecture to define a strong prior over natural images and jointly optimize this prior with a deep conditioning convolutional network. Human evaluations indicate that samples from our proposed model look more photo realistic than a strong L2 regression baseline.

https://arxiv.org/pdf/1702.00783v1.pdf
https://arxiv.org/abs/1702.00783