Explaining a CNN: Grad-CAM and activations

Grad-CAM

Grad-CAM highlights the parts of the image that pushed the model toward a class.

cam = cu.viz.gradcam(model, x)             # x: (1, 3, H, W), as fed to the model
cu.viz.show_cam(x, cam)                    # input and heatmap overlay side by side

cam.class_idx, cam.scores                  # which class was explained, its score
cam.layer                                  # which layer was used, e.g. "layer4"

  • Which class? The predicted one by default; ask for another with class_idx=3. For a batch, pass one index per image.

  • Which layer? The last layer that still has a spatial map, found automatically (layer4 for torchvision ResNets, the body for fastai vision_learner). Choose another with target_layer="layer3".

  • Normalized input? Pass mean=cu.viz.IMAGENET_MEAN, std=cu.viz.IMAGENET_STD to show_cam so the image underneath looks natural.

  • Reading the colors. A color bar next to the overlay maps colors to importance from 0 (ignored) to 1 (most important); the map is scaled per image. Hide it with show_color_bar=False.

With fastai:

x = cu.viz.learner_input(learn, "cat.jpg")       # same transforms as training
cam = cu.viz.gradcam(learn, x)
cu.viz.show_cam(x, cam, mean=cu.viz.IMAGENET_MEAN, std=cu.viz.IMAGENET_STD)

Other CAM methods

After cu.install("cam"), choose a method from pytorch-grad-cam:

cu.viz.gradcam(model, x, method="gradcam++")
# also: "xgradcam", "hirescam", "layercam", "scorecam", "eigencam", "ablationcam"

Layer activations

cu.viz.layer_names(model, torch.nn.Conv2d)              # what can I look at?
acts = cu.viz.activations(model, x, layers=["layer1", "layer3"])
acts["layer1"].shape                                    # (1, 64, 56, 56)
cu.viz.show_feature_maps(acts["layer1"], n=16)          # first 16 channels

Early layers respond to edges and colours; later layers to textures and parts.

None of these functions change your model: training mode, BatchNorm statistics, and parameter gradients are left exactly as they were.