search

maikherbig / AIDeveloper

GUI-based software for training, evaluating and applying deep neural nets for image classification
BSD 2-Clause "Simplified" License
109 stars 20 forks source link

Visual features activation visualization in networks #11

Closed alfrem closed 4 years ago

alfrem commented 4 years ago

Woud be nice to be able to see visually what features are being extracted in layers for cnn.

It's possible, too, to see heatmaps over images with those features. Saliencys and Grad-cams.

As seen in: https://raghakot.github.io/keras-vis/vis.visualization/

https://towardsdatascience.com/visualizing-intermediate-activation-in-convolutional-neural-networks-with-keras-260b36d60d0

https://machinelearningmastery.com/how-to-visualize-filters-and-feature-maps-in-convolutional-neural-networks/

https://analyticsindiamag.com/openai-launches-microscope-to-better-understand-neural-networks-in-popular-machine-learning-models/

Its very spectacular when implemented, specially for teaching reasons.

alfrem commented 4 years ago

Great! :D

maikherbig commented 4 years ago

I found a method called Grad-CAM You can now show the activation from an arbitrary layer. To do so, go to the "Assess Model" tab. Load a model and some data. Hit "Update plots" and then double click some item of the confusion matrix. The presented popup window now has options to show heatmaps or to create an overlay of the image and the heatmap. Commits: 76e031e 21ee953 9332f6e