0

I am getting this warning when I am trying to create saliency visualizations.

RuntimeError                              Traceback (most recent call last)

<ipython-input-36-6f13b9abef1d> in <module>()
      8   fig, axes = plt.subplots(1, 2)
      9   # Generate visualization
---> 10   visualize = visualize_saliency(saved_model, layer_index, filter_indices=input_class, seed_input=input_image)
     11   axes[0].imshow(input_image[..., 0])
     12   axes[0].set_title('Original image')


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

This is my code

layer_index = utils.find_layer_idx(saved_model, 'dense_2')
saved_model.layers[layer_index].activation = activations.linear
saved_model = utils.apply_modifications(saved_model)  
indices_to_visualize = [ 0, 12, 38, 83, 112, 74, 190 ]

# Visualize
for index_to_visualize in indices_to_visualize:
  # Get input
  input_image = X_test[index_to_visualize]
  input_class = np.argmax(y_test[index_to_visualize])
  # Matplotlib preparations
  fig, axes = plt.subplots(1, 2)
  # Generate visualization
  visualize = visualize_saliency(saved_model, layer_index, filter_indices=input_class, seed_input=input_image)
  axes[0].imshow(input_image[..., 0]) 
  axes[0].set_title('Original image')
  axes[1].imshow(visualize)
  axes[1].set_title('Saliency map')
  fig.suptitle(f'MNIST target = {input_class}')
  plt.show()

I have tried looking at other similar errors online but I dont understand how those solutions can be implemented in my code. How exactly am I supposed to use tf.GradientTape on my saliency visualization function. I havent come across anyone having the same error for saliency visualization.

Shourya Verma
  • 19
  • 5
  • `keras-vis` is unmaintained and not compatible with TF2. Downgrade your TensorFlow version or use another visualization library. – Lescurel Nov 13 '20 at 07:55
  • @Lescurel Hey! Thanks for the help. Could you recommend some good visualization libraries for 1D images or time series data. I am attempting interpretability of the MIT BIH dataset. – Shourya Verma Nov 14 '20 at 01:37

0 Answers0