Empty index_list

[gold-Soul]

I choice the selection scheme “HighestKGradients” and run the code as follows:
net_parser.register_hooks()
target_loss = (criterion(outputs, labels))
target_loss.backward(retain_graph=True)
index_list = net_parser.get_parameters()

But I found that the index_list is none and the function "hook" in function "generated_hook" from the file "HighestKGradients.py" didn't run.

[hslyu]

I wonder how you configure the parameter ratio and the input tensors.
It returns the empty array if hooks are not registered or the parameter ratio is zero.

Can you test whether the hook functions are called during the feed-forward and back-propagation?
It looks like you have customized the code. I wonder the original code returns the same empty array either.

[gold-Soul]

-1. I have test it that the register hook function 'hook_fn = self.generate_hook(start_index)' has been run.
-2. The parameter ratio was set as 0.2.
-3. The input tensors include the loss of a subset of a target class and the loss of a data point.
-4. The hook function didn't run during the feed-forward and back-propagation.
-5. The customized code about GIF compution and hook is highly similar to the original code in GIF-main/scripts/table1-parameter_selection_comparison_cifar10_0.ipynb. But I didn't run this ipynb file

[hslyu]

The hook functions are designed as callback functions that collect parameter indices when layerwise inputs/outputs (or gradient inputs/outputs) goes through their own layer.
Thus, if the hook functions are not running during the feed-forward and back-propagation, the get_parameters method should return the empty list.

It is hard to figure out the exact reason as I do not see the result output and your code snippet.