GradCAM

TensorFlow Network layer attention visualizer

GradCAM Principle

GradCAM visualize the saliency map of feature maps to the prediction class. In details, it cals every feature map's weights and its output . Then, apply RELU to filter out negative correlation points.

Then we cals its weights:

where Z is the pixel nums of feature map.

GradCAM Usage

Overide the following part for preprocessing and fetching the input tensor and auxiliary tensors

    # TODO create your own process
    def preprocess(self, x):
        x = x / 255.0
        x = x - np.array([0.485, 0.456, 0.406])
        x = x / np.array([0.229, 0.224, 0.225])
        return x

    # TODO fetch key tensors
    def fetch_key_tensors(self):
        """
            X:input tensor
            logit:logit output tensor
            layer:gradient dependent layer to the logit
            ... :fetch and set other tensors and their default values for inference
        """
        self.X = self.sess.graph.get_tensor_by_name('input:0')
        self.logit = self.sess.graph.get_tensor_by_name("finetune_dense1/BiasAdd:0")
        self.layer = self.sess.graph.get_tensor_by_name("ResNetnSequentialnlayer4nnBasicBlockn1nnReLUnrelun168:0")

        is_training = self.sess.graph.get_tensor_by_name("Placeholder:0")
        self.feed_dict = {is_training:False}

After that, simply use the following function to added saliency map to the images

    visualizer =GradCam(model_dir,nb_classes)
    visualizer.visualize_imgs(imgs_path,out_folder)

Please reference the following for further details:

model explanation

grad-cam.tensorflow