Visualizing activations in RNN while generating network output. about keract

[philipperemy]

@FilipJanitor are you interested in evaluating the network on a sequence step by step?

[philipperemy]

Also activations = outputs by the way here.

[philipperemy]

If you're interested in that, the most straightforward I see is just to split your sequence and feed the cum seq (cf below) to the network:

Your sequence: abcdefg
feed: a -> get activations (outputs of a)
feed: ab -> get activations (outputs of ab)
feed: abc -> get activations (outputs of abc)
...
feed: abcdefg.

I'm sure there are better ways to handle that but it's pretty hacky and straightforward.

[FilipJanitor]

I know for sure that obtaining the next character of the sequence using the model.predict(input) and picking the character using the output probabilities works as expected. I am not sure about the approach you suggested as I don't know how would such iterative feeding affect the internal state of the RNN.

But what you said about the activations and outputs sounds very promising. Assuming that predict returns the outputs of last layer I should be able to reconstruct the output of the last layer from what is returned by get_activations right? The only thing that is a bit confusing for me is this line: https://github.com/philipperemy/keras-activations/blob/e665695e10bf80d77e0bad21319227fb021007f9/keract/keract.py#L30 what would it get in relation to activations if it was uncommented?
Thanks!

[philipperemy]

This func is a Keras function where you define an input tensor and an output tensor. By feeding the input tensor with a numpy matrix, you evaluate the value of the output tensor.

In that case, we list all the layers and we evaluate them one by one. So for your input, you get the output of every layer with this package.

Predict does the same but on the whole model. It will output the outputs of the final layer.

layer_outputs = activations.

Sorry for the confusion ;)

[philipperemy]

LSTM has two components when you evaluate a sequence: States and Outputs. We're only talking about outputs here.

[FilipJanitor]

Yeah, sorry - I made it more complex than it is. It is all clear now :)
Thanks, for help!

[philipperemy]

Ok happy I could help :)