Visualize neural networks decisions with Grad-CAM
Last updated
Was this helpful?
Last updated
Was this helpful?
In this tutorial, we will see how Grad-CAM can help you visualize neural network decisions. The output of this tutorial will be a Grad-CAM overlay of your test dataset.
Example of a “car vs not car” model, highlighting the learned “features” of a car:
This tutorial will use a custom deployment block that you can add into your Organization (Enterprise only). This will make the Grad-CAM overlay output available in all your organization's projects.
Note: This currently works for image classification and visual regression models where at least one 2D convolution layer is present in your architecture.
Jupyter notebook:
Grad-CAM (Gradient-weighted Class Activation Mapping) is a technique that helps interpret the predictions of a convolutional neural network (CNN) by visualizing the regions of the input image that most influenced the model's decision. This is achieved by:
Computing gradients of the target output with respect to the feature maps of the last convolutional layer.
Weighting the feature maps by the importance of these gradients.
Generating a heatmap that highlights the areas of interest.
This script extends Grad-CAM for:
Classification Models: Highlights areas contributing to the predicted class.
Visual Regression Models: Highlights areas contributing to the numerical regression output.
Dataset export:
Exports the test dataset from an Edge Impulse project.
Automatically retries until the dataset is ready for download.
Model download:
Retrieves the trained model (.h5 format) for your Edge Impulse project.
Ensures compatibility with models containing 2D convolutional layers.
Grad-CAM visualization:
Applies the Grad-CAM technique to visualize the regions of the input image that contributed most to the model's predictions.
Works for:
Classification models: Highlights regions associated with the predicted class.
Regression models: Highlights regions contributing to the predicted regression value.
Output generation:
Saves Grad-CAM heatmaps overlaid on the original images.
Separates correctly predicted and incorrectly predicted samples into separate directories for easy analysis.
Clone this repository:
Initialize the block:
Push the deployment block to your organization:
Now you should see your custom deployment block in your organization:
Make sure to enable the Privileged mode, this custom block needs to make API requests to retrieve the dataset and the .h5 model
To use your custom block within your project, head to the Deployment page and select the Grad-CAM visualization option, the output will be a .zip file containing your test dataset with the Grad-Cam overlay.
Build:
Run:
You can adjust three parameters to fine-tune the visualization: alpha, pooling-gradients, and heatmap-normalization.
--alpha)The alpha parameter controls the transparency of the Grad-CAM overlay when it is superimposed on the original image.
Default Value: 0.4
Range:
A value between 0 and 1.
0: Fully transparent (only the original image is visible).
1: Fully opaque (only the Grad-CAM overlay is visible).
How to Choose:
Recommended Default (0.4): Provides a balance between showing the original image and highlighting the Grad-CAM heatmap.
Higher Alpha (> 0.5): Use this if you want the Grad-CAM heatmap to dominate the visualization.
Lower Alpha (< 0.4): Use this if you want the original image to be more prominent.
--pooling-gradients)The pooling-gradients parameter determines how gradients (importance signals) are combined across the spatial dimensions of the last convolutional layer to generate the heatmap.
Options:
mean (default):
Averages the gradients across the spatial dimensions.
Smooth out the heatmap, providing a general overview of important regions.
sum_abs:
Takes the sum of the absolute values of the gradients.
Highlights areas with strong activations, which can sometimes create sharper heatmaps.
How to Choose:
Classification Models:
Use mean for a smoother and more generalized heatmap.
Use sum_abs if you want to emphasize the most critical regions (e.g., sharp object boundaries).
Regression Models:
sum_abs is often more useful for regression tasks, as it highlights features contributing to extreme values.
Experiment: Try both options to see which one provides more meaningful visualizations for your model and dataset.
--heatmap-normalization)The heatmap-normalization parameter determines how the heatmap values are scaled for visualization.
Options:
percentile (default):
Ensures the heatmap values are scaled between 0 and 1 based on their maximum value.
Best for emphasizing high-activation areas while normalizing the output globally.
simple:
Normalizes the heatmap by dividing by the global maximum.
A simpler approach that may work well for datasets where most activations are relevant.
How to Choose:
Default (percentile):
Works well in most cases, especially if you expect only certain areas of the image to be significant.
Use simple:
Suitable for models where the activations are uniformly spread, and all areas are equally important.
This script assumes the presence of at least one 2D convolutional layer in the model architecture.
It is designed for image classification and visual regression tasks.
For regression models, the script uses a threshold to determine correctness; adjust this threshold (threshold = 0.1) as needed for your use case.
We also have an option to run it using a . This option will work for non-Enterprise projects.
Source code:
Blog post:
If you want more information on the Grad-CAM technique, we invite you to read this paper .
To test the functionality without setting up locally, use this . It comes pre-configured to run in a browser with no local setup required.
And you can by editing the custom deployment block:
