What's the difference between readout and decode, conceptually? about recurrentshop HOT 4 CLOSED

Readout is useful when you have a stack of rnn cells, and you want the output of the top cell at time t be available to the bottom cell at t-1.
For initializing the hidden states of a recurrent container, simple create one (rc = RecurrentContainer(...)), build it (rc.build(input_shape)) and reset the states (rc.states=[list of tensors]).
These are stuff that we use internally, will try to add better documentation in the future.

from recurrentshop.

I see, thanks. Doesn't a decoder always need the readout=True flag then?

from recurrentshop.

Not really.. since readout is not the only information being transferred across time. There are states too. The reason readout exists will make more sense when I push the teacher forcing thing.. During train time, instead of readout, the ground truth at time t will be available to the bottom cell at t + 1.

from recurrentshop.

Ah, yes that makes sense. Just out of curiosity: do you transfer the raw class probabilities tensor as input to the next timestep, or the one-hot encoded argmax, i.e. the resulting classification?

Thanks by the way for the updates you've been pushing, this is going in a very nice direction. Unfortunately I have not been able to successfully train a model on my task yet, but I'm hopeful! I like the way the API is structured.

from recurrentshop.