Tiled-mode for project_images.py and resumable training #9
Conversation
|
Yes this both makes sense. |
| if self.tiled: | ||
| dlatent_samples = self._Gs.components.mapping.run(latent_samples, None)[:, :1, :] # [N, 1, 512] | ||
| else: | ||
| dlatent_samples = self._Gs.components.mapping.run(latent_samples, None) # [N, 18, 512] | ||
| self._dlatent_avg = np.mean(dlatent_samples, axis=0, keepdims=True) # [1, 1 or 18, 512] |
There was a problem hiding this comment.
Would you mind explaining the idea behind tiled mode?
Compared to the original repository by Nvidia, I have noticed that @rolux changed this line.
Then you offered to add an argument to toggle between the original line and the line by @rolux.
I suspect tiled mode leads to better projection results, but I don't know why.
|
Would you mind explaining the idea behind tiled mode? I found your PR while looking for information about this exact change. That could explain why I get different projection results with Nvidia's original project and this one. |
|
Alright, I discovered this thread #2 and will read about it tomorrow. I think it might contain the answer. |
This PR contains two improvements:
project_images.py:--no-tiledwhich allows to switch to factory-default projection into1x512dlatents-space (useful for evaluation of "semantic" quality of the model);--result-dir