Problem to use 'hook' for extracting the feature maps in Mask-RCNN of Pytorch

Question

I'm using the Mask-RCNN Model that given by Pytorch for my project on Colab.

The model was adjusted to the data that i have(transfer learning) and train process was completed.

To the next step, i wanted to extract the feature maps of some input data from FPN.

So, i've tried to use 'hook' and made a code for that purpose like below.

load the weights of Mask-RCNN

state_dict = torch.load("~~~.path")
models.load_state_dict(state_dict)

function for hook

def get_activation(name):
  def hook(models, input, output):
     activation[name] = output
  return hook

Code to save feature maps of each input data

activation = {}
device = torch.device("cuda:0" if torch.cuda.is_available() else "cpu")
models.to(device)

for i in range(1, 4):
  test_img = Image.open("{path for input data}.png".format(i)).convert('RGB')
  test_img = test_img.resize((800, 800))
  test_img = tr.functional.to_tensor(test_img)
  test_img = test_img.reshape(1, 3, 800, 800)

  with torch.no_grad():
    models.eval()
    activation["fm{}".format(i)] = models.backbone.fpn.extra_blocks.register_forward_hook(get_activation("fm{}".format(i)))
    test_img = test_img.to(device)
    models(test_img)

But the feature map of just one input data was appended to every keys.

Like {"fm1" : a, "fm2" : a, "fm3" : a}

I don't know what is the reason and fault of this code.

Please help me know what should i try.