flomlo

if I understand the comment above https://github.com/rschroll/rmrl/blob/89b5cc38ef45251b96bd3d1c3618429b6b46db92/rmrl/render.py#L442 correctly then we need to load `CropBox` if existent, otherwise `MediaBox`. I would suggest the following fix: ``` bpage_box = list(map(float, basepage.inheritable.CropBox or...

Hello! The way my implementation should solve that (it is currently broken for models which need (batch_size, sequence_length, input_dim), but not if only adding a layer) is via statefulness. The...

If I have a sample A of shape (7,5,3), and we feed it to the *NTM*, the controller will see 5 times in a row input of shape (7,3) (dense...

> there are 7 samples of shape (7,5,3) Do you mean 7 samples, each of batch_size=7 (in which case I would suggest using another prime for the number of samples)?...

Damn, getting those stateful controllers to run again is quite frusttrating. Especially as I know that it once worked! There is a fallback solution of just grabbing the state of...

my code is very, *very* roughly based on the the seya-repository[1]: I translated all his code from Keras 0.3 to Keras 2.0, then realised that I finally understood how an...

Ok, calling the LSTM step-function directly works. But damn, is that ugly! Really not satisfying. And very bad to allow arbitrary controllers with that. That would mean I would have...

Oh man, if I would have been aware of SigmaQuans NTM-Keras solution I wouldnt have build this one :D I agree with you with the stateful + single-step-execution warning: Testing...

update: still working on a keras patch which fixes the behaviour for calling rnn with step-size=1.

Ah yes, the loading/saving problem. Admittedly I havent done that so far via the keras load/save functions. I currently do not understand completely how that works in keras. I have...