GradCAM's heatmap worsened from .1 to .2

sicara / tf-explain

Interpretability Methods for tf.keras models with Tensorflow 2.x

https://tf-explain.readthedocs.io

MIT License

1.02k stars 111 forks source link

GradCAM's heatmap worsened from .1 to .2 #136

Open TedGraham opened 4 years ago

TedGraham commented 4 years ago

I'm using Keras' VGG16 to classify a picture of a dog and am generating a heatmap from GradCAM.explain.

Using .1.0 and specifying the layer_name of "block5_conv3" gives a much nicer heatmap (first image) than using .2.0 and not specifying the layer_name (second image).

the behavior of .2.1 and .2.0 appears the same to me
specifying the layer_name with .2.1 doesn't fix this, the auto choosing of the layer appears to be picking the block5_conv3 layer

explain-0 1 0

explain-0 2 1

RaphaelMeudec commented 4 years ago

Thanks for raising this, will look into it!