I am trying to visualize classification results from neural network in Keras using TensorBoard Embeddings. If I understand correctly, the Embeddings can be used to visualize any tensor (when TB is being used directly with TensorFlow), not just embedding layer (weights / output).
The Keras TensorBoard callback lets me specify a layer to watch. When I set it to watch the output layer, then the only tensor available in TB is "kernel" (meaning weights?).
So I thought I could use TB directly instead, without the Keras wrapper, but I found no way to obtain the output tensor of Keras layer.
The only workaround I can think of would be to build the tensor (using TF variable) manually. Is this the only way? Or am I missing / misunderstanding something?
Asked
Active
Viewed 676 times
2
Aleš Koblížek
- 831
- 9
- 5
-
Here's [a link](https://stackoverflow.com/questions/45265436/keras-save-image-embedding-of-the-mnist-data-set)! Maybe you can find what you want in this link. – xieydd Dec 10 '17 at 14:40