search

TeamHG-Memex / eli5

A library for debugging/inspecting machine learning classifiers and explaining their predictions
http://eli5.readthedocs.io
MIT License
2.76k stars 334 forks source link

Keras image classifier tutorial issue #384

Open leszekmp opened 4 years ago

leszekmp commented 4 years ago

Copying and pasting the code from the tutorial into a notebook. Works fine through the "#check the top 5 indices" section. In the next section, "eli5.show_prediction(model, doc)" code results in the following error:

RuntimeError: tf.gradients is not supported when eager execution is enabled. Use tf.GradientTape instead.

Tried adding the following two lines:

import tensorflow as tf tf.compat.v1.disable_eager_execution()

Error changes to:

ValueError: Tensor Tensor("Mean:0", shape=(None, 1280), dtype=float32) is not an element of this graph.

Help?!

teabolt commented 4 years ago

Hi @leszekmp,

I can run the entire notebook if I try it with

import tensorflow as tf
tf.compat.v1.disable_eager_execution()

What is your tensorflow version? I am using tensorflow==2.2.0

leszekmp commented 4 years ago

I'm also using tensorflow 2.2.0.

teabolt commented 4 years ago

Could you share more code that you put into your notebook? I'm not sure what's the issue as the notebook works for me when eager execution is disabled.

I think the error is related to the "weights" variable (mean of all the gradients for each activation map) https://github.com/TeamHG-Memex/eli5/blob/master/eli5/keras/gradcam.py#L120. You can search the notebook for 1280 and you'll see that the model has layers with 1280 activation maps near the end of it.