LayerDeepLift fails when used on a MaxPooling layer?

[Holt59]

I am trying to use LayerDeepLift on multiple layers of a VGG16 model from torchvision.models. It works for all layers except MaxPooling2D layers.

The following (layer 23 is a MaxPool2d layer):

model = torchvision.models.vgg16(pretrained=True)
u = captum.attr.LayerDeepLift(
    model, list(model.features.children())[23]).attribute(
        torch_im[None, ...], target=156)[0]

Raises the following:

---------------------------------------------------------------------------
RuntimeError                              Traceback (most recent call last)
<ipython-input-66-668b5d33db17> in <module>
----> 1 u = captum.attr.LayerDeepLift(model, list(model.features.children())[23]).attribute(torch_im[None, ...], target=156)[0]

i:\languages\python\envs\deel-torch\lib\site-packages\captum\attr\_core\layer\layer_deep_lift.py in attribute(self, inputs, baselines, target, additional_forward_args, return_convergence_delta, attribute_to_layer_input, custom_attribution_func)
    306             inputs,
    307             attribute_to_layer_input=attribute_to_layer_input,
--> 308             output_fn=lambda out: chunk_output_fn(out),
    309         )
    310

i:\languages\python\envs\deel-torch\lib\site-packages\captum\attr\_utils\gradient.py in compute_layer_gradients_and_eval(forward_fn, layer, inputs, target_ind, additional_forward_args, gradient_neuron_index, device_ids, attribute_to_layer_input, output_fn)
    517             for layer_tensor in saved_layer[device_id]
    518         )
--> 519         saved_grads = torch.autograd.grad(torch.unbind(output), grad_inputs)
    520         saved_grads = [
    521             saved_grads[i : i + num_tensors]

i:\languages\python\envs\deel-torch\lib\site-packages\torch\autograd\__init__.py in grad(outputs, inputs, grad_outputs, retain_graph, create_graph, only_inputs, allow_unused)
    155     return Variable._execution_engine.run_backward(
    156         outputs, grad_outputs, retain_graph, create_graph,
--> 157         inputs, allow_unused)
    158
    159

i:\languages\python\envs\deel-torch\lib\site-packages\captum\attr\_core\deep_lift.py in _backward_hook(self, module, grad_input, grad_output, eps)
    461         multipliers = tuple(
    462             SUPPORTED_NON_LINEAR[type(module)](
--> 463                 module, module.input, module.output, grad_input, grad_output, eps=eps
    464             )
    465         )

i:\languages\python\envs\deel-torch\lib\site-packages\captum\attr\_core\deep_lift.py in maxpool2d(module, inputs, outputs, grad_input, grad_output, eps)
    920         grad_input,
    921         grad_output,
--> 922         eps=eps,
    923     )
    924

i:\languages\python\envs\deel-torch\lib\site-packages\captum\attr\_core\deep_lift.py in maxpool(module, pool_func, unpool_func, inputs, outputs, grad_input, grad_output, eps)
   1002
   1003     new_grad_inp = torch.where(
-> 1004         abs(delta_in) < eps, grad_input[0], unpool_grad_out_delta / delta_in
   1005     )
   1006     # If the module is invalid, save the newly computed gradients

RuntimeError: The size of tensor a (28) must match the size of tensor b (14) at non-singleton dimension 3

It works on all layers except the MaxPool2d layers of vgg16.features (it works with the average pooling layer).

I am not sure if this is a restriction of DeepLift or an error in the implementation?

Also, when the error occurs, the model seems to be left in some weird state as re-using it leads to IndexError: tuple index out of range (even with a brand new captum.attr.LayerDeepLift instance).

[NarineK]

Hi @Holt59, thank you for the question. That's interesting! I'll debug it.
In terms of the model state: Those are the hooks that perhaps aren't getting removed. We actually recently fixed it so that we remove the hooks for all cases. For that you might need github version.

[bilalsal]

Hi @Holt59.
While @NarineK looks into the main issue, I thought to refer you to #370 regarding re-using the model after an error occurs.
The new error is likely related to dangling hooks and is fixed in master as explained in the comments.
Hope this helps

[NarineK]

@Holt59, I've been debugging this issue. There seem to be some inconsistencies in the backward pass. In the meanwhile, as a workaround, if you want to attribute to the inputs of the MaxPool2D layer it will work. By default we attribute to the outputs of the layer.

model = torchvision.models.vgg16(pretrained=True)
u = captum.attr.LayerDeepLift(
    model, list(model.features.children())[23]).attribute(
        torch_im[None, ...], target=156, attribute_to_layer_input=True)[0]

[NarineK]

Actually what you were doing will be equivalent to:

u = LayerDeepLift(
    model, list(model.features.children())[24]).attribute(
        torch_im, target=156, attribute_to_layer_input=True)[0]

as a workaround

[NarineK]

@Holt59 , did that workaround work for you?

[Holt59]

The workaround seems to work but I cannot use it in my code base like this since I am trying to compute attributions for multiple layers (and I don't know the following layer). But that's not a big issue, I'm not particularly interested in the MaxPool layers, I can leave them out.

[NarineK]

@Holt59, this PR #390 will fix the problem with MaxPool. To give more context, this problem happened because in the forward_hook we return cloned output tensor and that made the MaxPool modules complex. Since there is a bug related to complex modules in PyTorch and backward_hook, that is, returned input gradients represent only a subset of inputs, it wasn't able to compute the multipliers correctly.

More details about the issue can be found here: https://pytorch.org/docs/stable/nn.html#torch.nn.Module.register_backward_hook

Another point that I wanted to bring up is: In VGG the modules might get reused (you might want to check that). We want to make sure that this isn't happening for the layer algorithms and DeepLift.
If the activations get reused. You can simply redefine the architecture (that's easy to do). More info about it can be found here:

#378 (comment)

[NarineK]

This got fixed through: #390