Fix bug in stage1 (en-/decoder pre-) training

[sRassmann]

Fix the bug of choosing the trained target modality randomly. Instead source and target images are always identical:
Before, the target and source modality were chosen randomly; hence, the generator could not know which target modality to reconstruct leading to in ill-posed loss.

Fix smaller bugs

[jongdory]

I just fixed a bug in stage 1 where the modality of the target and reconstruction were different.

Thank you for your contribution!

[puyln]

I just fixed a bug in stage 1 where the modality of the target and reconstruction were different.

Thank you for your contribution!

Is the 'target' parameter missing from line 130 of the updated vqgan.py file?

[sRassmann]

Yes in the current version the code will not run either, as the forward func expects a target variable.
line 80 should be

def forward(self, input, target=None):

[Niranjin]

Hi Jonghun, love the ALDM repo! Would you be able to share the pre-trained model weights or let me know where to find them? Thanks in advance! I can access the drive link for second model but I couldn't for the first one. Please guide me here.

[jongdory]

Hi Jonghun, love the ALDM repo! Would you be able to share the pre-trained model weights or let me know where to find them? Thanks in advance! I can access the drive link for second model but I couldn't for the first one. Please guide me here.

Hi, thank you for your interest and for pointing this out! You were right; the link to the first model was incorrect. I just fixed it, so you should be able to access it now. Thanks again!