Dataset with incomplete combinations

[brunobian]

Hi! I am using this code to work on a dataset of ~700 words. For each word I am varying several variables (size, font, position, etc) . This results in a too big dataset (+6M instances) to use all the possible combinations during training, so I decided to use a sample of the full dataset. That is pl for the training, but this creates an issue during the evaluation run.

In particular, in evaluate.compute_metrics() I found the first technical issue. To run this method the code tries to reshape samples_zCx and params_zCx tensors using the sizes of the dataset generation factors (lat_sizes) and the latent layer size (latent_dim). This is not a problem when using a dataset with all the possible combinations, but given that I now have a sample of all the possibilities, this is not the case. So, I cannot make the reshape.

I solved this by creating a tensor of np.nan and filling it with the available data in the corresponding cells (using metadata from the dataset that indicates how each instance was created). Technically, this works, but I now have doubts about how this solutions impacts on the following calculations. That is, I now have a tensor with NANs that will be used to compute the conditional entropy H(z|v), is this ok? Would it better to use zeros?

Additionally, computing the conditional entropy with the _estimate_H_zCv() method is pretty computationally expensive given that I have a big tensor full of NANs. Would it be ok to skip the cells with NANs to speedup the process?