I’ve been experiencing for a while in Keras CNN, and I’ve reached a moment where I need to see what is in the filters and features maps. My models are not improving and I need to dig deeper.
I don’t think there is a simple way of achieving this. It is possible to get more information, e.g. by outputting other layers than the last, though I don’t know of a way to visualize anything helpful.
What do you mean and how do you output the other layers as you suggest?
Is there any possibility to extract the weights on each node? I supose that Knime/Keras has the final weights somewhere in order to apply them once you use the predictor. In the current outputs we users have in the learner I don’t see a way to see anything useful, but maybe (and that is my question) there is another node or python code that we could use to extract those weights. Do we have it? Is it possible to develop it?
Extracting the weights via Keras nodes isn’t possible at the moment. However, a workaround may be to use the DL Python Network Editor node and use a code snippet such as this:
for layer in input_network.layers:
weights = layer.get_weights()
print(weights)
What a pity!! You guys really rock. I am pretty sure you can add such an important information in future updates. Moreover when you already have that data in an internal variable (not visible).
In my case I am not sure if this will be a reallistic option as I do not know Python and the workflow for keras nodes is pretty damn complex, tons of loops and variables.
I will think in ways to solve my problems. If I do not find any, I will have to work on the Python end.
Many thanks! And please, notice me if there is any change in the keras possibilities.
@Apeire Thank you for the kind words! And yes, I will keep my eye out for any KNIME updates regarding visualizing Keras models or anything that I believe would be helpful with your specific issue.