keras_gradcam.py

def grad_cam(img, model):
    """Gradient Activation Map for keras model
    # Arguments
        img: image to plot gradcam
        model: keras model
    # Returns
        gradcam image
    """  
  
    img = np.expand_dims(img, 0)

    last_conv_layer = model.get_layer('leaky_re_lu_17').output # setup the correct name for last conv layer
    grads = K.gradients(model.get_layer('dense_1').output[:,1], last_conv_layer)[0]
    pooled_grads = K.mean(grads, axis=(0, 1, 2))
    out = model.output[:,1]

    iterate = K.function([model.get_layer("input_1").input], 
                    [pooled_grads, last_conv_layer[0], out])

    pooled_grads_value, conv_layer_output_value, output = iterate([img])

    # 2048 for ResNet-50, since the shape of feature maps is (NoneX?x?x2048)
    for i in range(2048):
        conv_layer_output_value[:,:,:,i] *= pooled_grads_value[i]

    gradcam = np.mean(conv_layer_output_value, axis=-1)
    gradcam = np.maximum(gradcam, 0)
    gradcam /= np.max(gradcam)

    import skimage
    import scipy

    cam = scipy.ndimage.zoom(gradcam, 16, mode="nearest")
    gradcams.append(cam)
    print(x, output)
return gradcams