Convolutional Conditional Variational Autoencoder with label in recon loss is bad about deeplearning-models

[rasbt]

I don't know. It works fine for the non-convolutional (fully connected) VAE. I suspect that's because if you add the reconstruction labels in the conv. variant, some of the structure of the embedding is lost.

[Johnson-yue]

the fully connected is very simple and the mnist is also simple dataset，So， you can use any resolution to fix it，but using convolution it is more complex ， so， it should be more tricky to fix it。

Maybe the gradient of network is bad ？

[rasbt]

maybe. but there is no bug in the code, I think. The same architecture works fine when the labels are not concatenated with the image to compute the reconstruction loss. Also, the general approach (concatenating) works fine for fully connected ones.

Maybe this just needs some more hyperparam tuning.

[Johnson-yue]

yes, I think so , would you plan to add other dataset, Maybe like ,CIFAR-10, tiny-Imagenet?

[rasbt]

yeah, one day :). There are many things that could be added. This would be interesting to try to investigate how well this approach works on RGB images

[Johnson-yue]

thanks for your sharing