How to use hook to visualise feature maps

1

I’ve written the following code to visualise the feature maps (four from each layer) from a fully-trained CNN network, by passing images forward through the network. I’ve seen examples where layer activations are saved by registering a hook into the network but how do I do this?

def feature_map_visualisation(images, image_index):

    images = images.to(device)

    conv1_activation = model_gpu.first_layer[0](images)

    conv1_active_relu = model_gpu.first_layer[1](conv1_activation)

    conv1_active_pooling = model_gpu.first_layer[2](conv1_active_relu)

    conv1_active_drop = model_gpu.first_layer[3](conv1_active_pooling)

    conv2_activation = model_gpu.second_layer[0](conv1_active_drop)

    figure = plt.figure(figsize = (16,16))

    k = 0

for i in range(0, 4):

      k += 1

      figure.add_subplot(1,4,k)

      img = conv1_activation[image_index][i].cpu().detach().clone().numpy()

      img = rgb2gray(img)

      plt.imshow(img, cmap="gray")

    j=0

for i in range(0,4):

      j+=1

      figure.add_subplot(2,4,j)

      img1=conv2_activation[image_index][i].cpu().detach().clone().numpy()

      img1 = rgb2gray(img1)

      plt.imshow(img1, cmap="gray")
2

Here is a small example of using hooks to get the activations. :wink: