Awni Hannun
Awni Hannun
Can you share more details on your use case? Indeed boolean indices are not well supported because we don't know the values when you make the call, and hence we...
Got it, and what do you do with `only_falses`? If it's not too costly you can use a mask instead of selecting the subset? I see it would be nice...
Got it. We need boolean masking for this. But it looks like you don't need to differentiate / transform through this line: ``` non_final_states = state_batch[~terminated_batch] # can use mx.logical_not(terminated_batch)...
It's unlikely going to Numpy will slow things down much for just that step. But you may be able to get this to work using `where`: ``` top_probs = mx.where(cumulative_probs...
You can use `where(cond, x, y)` in c++. But maybe you can you give an example of what you are trying to do?
I recommend trying to find an equivalent computation that does not require you to have an array whose shape depends on another array's data. Based on your comment I can't...
CTC assigns a score to the sequence Y = [y_1, y_2, ..., y_U], namely p(Y | X). This is different than the sequence Z with the blanks inserted. We just...
Thanks for pointing this out! I submitted a PR to fix, hopefully one of the editors will merge it in :). Awni
Take a look at #36.
Yea 💯 . Was just discussing this the other day with @angeloskath (he may already have a diff not sure). I'm happy to have it as a contribution if we...