Before the bolt-related decoders were using the size of the latent space rather than the native size of the output of resnet50's features. This meant that the decoder was trying to build a model different to the encoder. I think this fixes it. Fixed for vq and vae
Before the bolt-related decoders were using the size of the latent space rather than the native size of the output of resnet50's features. This meant that the decoder was trying to build a model different to the encoder. I think this fixes it. Fixed for vq and vae