Enhancements

[michaelklachko]

1. display weights (in addition to activations)

2. display how activations/weights change during training. For example, I should be able to point to checkpoint directory where my model is saved every epoch (or every few iterations), and then hit 'play' button to see how a particular feature map or weights filter is changing as I move between epochs/iterations.

3. display which inputs contribute the most to the activations, at every layer.

These additional features would be very useful for debugging, for model compression, for identifying information flows (e.g. in a DenseNet), for saliency analysis, and many other things as well.

[CharlesLiuyx]

Nice suggestions! Thanks a lot.

The first problem need to be solved is the API of TensorFlow.js is limited which is not powerful compare to TensorFlow. Feel hard to extract every relative weights, parameters and contributors in the models.

Can this features be visualized by TensorBoard?

[michaelklachko]

Can this features be visualized by TensorBoard?

yes, I think so.

[CharlesLiuyx]

Our framework based on TensorFlow.js. So when it comes to the model part, it depends on the implements and APIs of TensorFlow.js because we need to run our model in browser. Our team will try to add this features.

If you have some materials about this topic please refer them to me. Thanks