Open seakforzq opened 11 months ago
And why we need gt_target to prune the upsampled sparsetensor?
# If training, force target shape generation, use net.eval() to disable
if self.training:
keep1 += target
I think the network will not learn how to pruning the output in this way right?
Is your feature request related to a problem? Please describe. How to sample from the vae's latent space?
Describe the solution you'd like The code above only reaches the reconstruction purpose, and target_key from sin is needed, can you help me with sampling from the latent space?
Describe alternatives you've considered NO
Additional context NO