Thank you for your amazing work! When going through your code after reading GAGA and MIOFlow I stubled upon a potential issue.
It seems like you only shuffle the indices once before training the distance preserving autoencoder and then keep fixed batches over the course of the training. Is this the desired behaviour and why would you prefer this over ranom samples in each iteration?
Thank you for your amazing work! When going through your code after reading GAGA and MIOFlow I stubled upon a potential issue.
It seems like you only shuffle the indices once before training the distance preserving autoencoder and then keep fixed batches over the course of the training. Is this the desired behaviour and why would you prefer this over ranom samples in each iteration?