Weights & Biases Callbacks

wandb-callbacks provides some additional Callbacks for Weights & Biases.

Callbacks currently implemented:

• ActivationCallback
  • visualizes the activations of a layer
  • activations are computed for a sample of each class
• DeadReluCallback
  • logs the number of dead relus in each layer
  • prints warning if the percentage is higher than a threshold
• GradCAMCallback
  • Grad-CAM: Visual Explanations from Deep Networks via Gradient-based Localization
  • produces a coarse localization map highlighting the important regions in the image for predicting the class of the image

Sample Implementation

Can be found in notebooks/sample_implementation.ipynb.