search

pytorch / captum

Model interpretability and understanding for PyTorch
https://captum.ai
BSD 3-Clause "New" or "Revised" License
4.92k stars 496 forks source link

LRP TypeError: Module type <class 'torch.nn.modules.upsampling.Upsample'> is not supported.No default rule defined. #712

Open bugsuse opened 3 years ago

bugsuse commented 3 years ago

I want to use LRP to explain Unet model, but raise TypeError as follow,

TypeError: Module type <class 'torch.nn.modules.upsampling.Upsample'> is not supported.No default rule defined.

In addition, raise same error when using nn.ConvTranspose2d to upsample.

Any help is appreciated!

NarineK commented 3 years ago

@bugsuse, since nn.ConvTranspose2d is not a very common activation we don't have default rules for it. Have you tried to explicitly attach a rule to it similar to this tutorial ? https://github.com/pytorch/captum/pull/668

bugsuse commented 3 years ago

@NarineK Thanks a lot! I tested it following the tutorial 668. I need to use nn.Upsample or nn.ConvTanspose2d to upsample output to the original size, because I want to use unet model for semantic segmentation task and explain it.

How should I add rules to support nn.Upsample or nn.ConvTranspose2d or skip these unsupported layers when using LRP in captum to calculate attributes?

NarineK commented 3 years ago

@bugsuse for upsampling of ConvTranspose2d you might want to start w/Epsilon rule. Have you tried to set the rules with something like this ?

from captum.attr._utils.lrp_rules import EpsilonRule
model.layer.rule =  EpsilonRule()
bugsuse commented 3 years ago

Hey, @NarineK I'm new to captum and XAI. I don't clear that how to add w/Epsilon rule to nn.ConvTranspose2d? Is there a similar example in captum? I'd like to learn about and try to add? Thanks so much!

bugsuse commented 3 years ago

Hey @NarineK, I have tried add nn.ConvTranspose2d : EpsilonRule to SUPPORTED_LAYERS_WITH_RULES or SUPPORTED_NON_LINEAR_LAYERS into lrp.py, but I get AssertionError due to target dimension. The error information is below,

---------------------------------------------------------------------------
AssertionError                            Traceback (most recent call last)
<ipython-input-20-90353dc30936> in <module>
      1 lrp = LRP(model)
----> 2 attribution = lrp.attribute(vifea.cuda(), target=torch.unsqueeze(vtar, dim=1).cuda(), verbose=True)
      3 attribution = attribution.squeeze().permute(1, 2, 0).detach().numpy()

/data3/works/ai/phd/github/captum/captum/log/__init__.py in wrapper(*args, **kwargs)
     33             @wraps(func)
     34             def wrapper(*args, **kwargs):
---> 35                 return func(*args, **kwargs)
     36 
     37             return wrapper

/data3/works/ai/phd/github/captum/captum/attr/_core/lrp.py in attribute(self, inputs, target, additional_forward_args, return_convergence_delta, verbose)
    191         try:
    192             # 1. Forward pass: Change weights of layers according to selected rules.
--> 193             output = self._compute_output_and_change_weights(
    194                 inputs, target, additional_forward_args
    195             )

/data3/works/ai/phd/github/captum/captum/attr/_core/lrp.py in _compute_output_and_change_weights(self, inputs, target, additional_forward_args)
    341         try:
    342             self._register_weight_hooks()
--> 343             output = _run_forward(self.model, inputs, target, additional_forward_args)
    344         finally:
    345             self._remove_forward_hooks()

/data3/works/ai/phd/github/captum/captum/_utils/common.py in _run_forward(forward_func, inputs, target, additional_forward_args)
    453         else inputs
    454     )
--> 455     return _select_targets(output, target)
    456 
    457 

/data3/works/ai/phd/github/captum/captum/_utils/common.py in _select_targets(output, target)
    472             return torch.gather(output, 1, target.reshape(len(output), 1))
    473         else:
--> 474             raise AssertionError(
    475                 "Tensor target dimension %r is not valid. %r"
    476                 % (target.shape, output.shape)

AssertionError: Tensor target dimension torch.Size([16, 1, 1, 64, 64]) is not valid. torch.Size([16, 1, 1, 64, 64])

The code is below,

lrp = LRP(model) # Unet model for semantic segmentation
attribution = lrp.attribute(vifea.cuda(), target=torch.unsqueeze(vtar, dim=1).cuda(), verbose=True)
attribution = attribution.squeeze().permute(1, 2, 0).detach().numpy()

Is it because LRP of captum does not currently support semantic segmentation models?

NarineK commented 3 years ago

@bugsuse, instead of tensor, do you mind representing target as a list of tuples similar to this ? https://github.com/pytorch/captum/blob/master/tests/attr/helpers/test_config.py#L199

Remember that first dimension corresponds to the example index. Target for each example has to have length 4: s.a. [(0, 0, 0, 0), (0, 0, 24, 24)... ]

bugsuse commented 3 years ago

@NarineK I tried to instead of tensor using the codes below, but raised a new error, 

lrp = LRP(model)
attribution = lrp.attribute(vifea[0:1].cuda(), target=[(0, 0, 0)], verbose=True)
attribution = attribution.squeeze().permute(1, 2, 0).detach().numpy()

The shape of vifea and unet model output is (16, 24, 64, 64) and (16, 1, 64, 64).

The information of error is below, 

Applied <captum.attr._utils.lrp_rules.EpsilonRule object at 0x2b57472e8be0> on layer Conv2d(24, 16, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
Applied <captum.attr._utils.lrp_rules.EpsilonRule object at 0x2b5747309190> on layer BatchNorm2d(16, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
Applied <captum.attr._utils.lrp_rules.EpsilonRule object at 0x2b5747309d60> on layer Conv2d(16, 16, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
Applied <captum.attr._utils.lrp_rules.EpsilonRule object at 0x2b5747232580> on layer BatchNorm2d(16, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
Applied <captum.attr._utils.lrp_rules.EpsilonRule object at 0x2b57473571c0> on layer Conv2d(16, 16, kernel_size=(3, 3), stride=(2, 2), padding=(1, 1))
Applied <captum.attr._utils.lrp_rules.EpsilonRule object at 0x2b5747357a90> on layer Conv2d(16, 32, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
Applied <captum.attr._utils.lrp_rules.EpsilonRule object at 0x2b5747357dc0> on layer BatchNorm2d(32, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
Applied <captum.attr._utils.lrp_rules.EpsilonRule object at 0x2b5747357c10> on layer Conv2d(32, 32, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
Applied <captum.attr._utils.lrp_rules.EpsilonRule object at 0x2b5747357730> on layer BatchNorm2d(32, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
Applied <captum.attr._utils.lrp_rules.EpsilonRule object at 0x2b5747357d30> on layer Conv2d(32, 32, kernel_size=(3, 3), stride=(2, 2), padding=(1, 1))
Applied <captum.attr._utils.lrp_rules.EpsilonRule object at 0x2b5747327df0> on layer Conv2d(32, 64, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
Applied <captum.attr._utils.lrp_rules.EpsilonRule object at 0x2b5747327fd0> on layer BatchNorm2d(64, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
Applied <captum.attr._utils.lrp_rules.EpsilonRule object at 0x2b5747327f10> on layer Conv2d(64, 64, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
Applied <captum.attr._utils.lrp_rules.EpsilonRule object at 0x2b5747327580> on layer BatchNorm2d(64, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
Applied <captum.attr._utils.lrp_rules.EpsilonRule object at 0x2b5747327e20> on layer Conv2d(64, 64, kernel_size=(3, 3), stride=(2, 2), padding=(1, 1))
Applied <captum.attr._utils.lrp_rules.EpsilonRule object at 0x2b5747327f70> on layer Conv2d(64, 128, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
Applied <captum.attr._utils.lrp_rules.EpsilonRule object at 0x2b5747327ee0> on layer BatchNorm2d(128, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
Applied <captum.attr._utils.lrp_rules.EpsilonRule object at 0x2b5747327fa0> on layer Conv2d(128, 128, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
Applied <captum.attr._utils.lrp_rules.EpsilonRule object at 0x2b574735a940> on layer BatchNorm2d(128, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
Applied <captum.attr._utils.lrp_rules.EpsilonRule object at 0x2b574735afa0> on layer Conv2d(128, 128, kernel_size=(3, 3), stride=(2, 2), padding=(1, 1))
Applied <captum.attr._utils.lrp_rules.EpsilonRule object at 0x2b574735aac0> on layer Conv2d(128, 256, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
Applied <captum.attr._utils.lrp_rules.EpsilonRule object at 0x2b574735a2e0> on layer BatchNorm2d(256, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
Applied <captum.attr._utils.lrp_rules.EpsilonRule object at 0x2b574735a790> on layer Conv2d(256, 256, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
Applied <captum.attr._utils.lrp_rules.EpsilonRule object at 0x2b574735abb0> on layer BatchNorm2d(256, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
Applied <captum.attr._utils.lrp_rules.EpsilonRule object at 0x2b574735ad00> on layer Conv2d(256, 256, kernel_size=(3, 3), stride=(2, 2), padding=(1, 1))
Applied <captum.attr._utils.lrp_rules.EpsilonRule object at 0x2b574735a970> on layer Conv2d(256, 256, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
Applied <captum.attr._utils.lrp_rules.EpsilonRule object at 0x2b574735a6d0> on layer BatchNorm2d(256, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
Applied <captum.attr._utils.lrp_rules.EpsilonRule object at 0x2b574735a820> on layer Conv2d(256, 256, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
Applied <captum.attr._utils.lrp_rules.EpsilonRule object at 0x2b574735a6a0> on layer BatchNorm2d(256, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
Applied <captum.attr._utils.lrp_rules.EpsilonRule object at 0x2b574735ae20> on layer Conv2d(512, 256, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
Applied <captum.attr._utils.lrp_rules.EpsilonRule object at 0x2b574735aa90> on layer BatchNorm2d(256, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
Applied <captum.attr._utils.lrp_rules.EpsilonRule object at 0x2b574735a9a0> on layer Conv2d(256, 128, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
Applied <captum.attr._utils.lrp_rules.EpsilonRule object at 0x2b574735a7f0> on layer BatchNorm2d(128, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
Applied <captum.attr._utils.lrp_rules.EpsilonRule object at 0x2b574735a520> on layer Conv2d(256, 128, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
Applied <captum.attr._utils.lrp_rules.EpsilonRule object at 0x2b574735a310> on layer BatchNorm2d(128, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
Applied <captum.attr._utils.lrp_rules.EpsilonRule object at 0x2b574735a430> on layer Conv2d(128, 64, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
Applied <captum.attr._utils.lrp_rules.EpsilonRule object at 0x2b574735a070> on layer BatchNorm2d(64, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
Applied <captum.attr._utils.lrp_rules.EpsilonRule object at 0x2b574735a4f0> on layer Conv2d(128, 64, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
Applied <captum.attr._utils.lrp_rules.EpsilonRule object at 0x2b574735a580> on layer BatchNorm2d(64, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
Applied <captum.attr._utils.lrp_rules.EpsilonRule object at 0x2b574735a7c0> on layer Conv2d(64, 32, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
Applied <captum.attr._utils.lrp_rules.EpsilonRule object at 0x2b574735a910> on layer BatchNorm2d(32, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
Applied <captum.attr._utils.lrp_rules.EpsilonRule object at 0x2b57472426a0> on layer Conv2d(64, 32, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
Applied <captum.attr._utils.lrp_rules.EpsilonRule object at 0x2b57472421f0> on layer BatchNorm2d(32, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
Applied <captum.attr._utils.lrp_rules.EpsilonRule object at 0x2b574711c6d0> on layer Conv2d(32, 16, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
Applied <captum.attr._utils.lrp_rules.EpsilonRule object at 0x2b574711ce80> on layer BatchNorm2d(16, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
Applied <captum.attr._utils.lrp_rules.EpsilonRule object at 0x2b574711c7f0> on layer Conv2d(32, 16, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
Applied <captum.attr._utils.lrp_rules.EpsilonRule object at 0x2b574711c100> on layer BatchNorm2d(16, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
Applied <captum.attr._utils.lrp_rules.EpsilonRule object at 0x2b574711cf70> on layer Conv2d(16, 16, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
Applied <captum.attr._utils.lrp_rules.EpsilonRule object at 0x2b574711cc40> on layer BatchNorm2d(16, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
Applied <captum.attr._utils.lrp_rules.EpsilonRule object at 0x2b574711cd90> on layer Conv2d(16, 1, kernel_size=(1, 1), stride=(1, 1))
---------------------------------------------------------------------------
RuntimeError                              Traceback (most recent call last)
<ipython-input-107-b865fd048fdd> in <module>
      1 lrp = LRP(model)
----> 2 attribution = lrp.attribute(vifea[0:1].cuda(), target=[(0, 0, 0)], verbose=True)
      3 attribution = attribution.squeeze().permute(1, 2, 0).detach().numpy()

/data3/yangl/works/ai/phd/github/captum/captum/log/__init__.py in wrapper(*args, **kwargs)
     33             @wraps(func)
     34             def wrapper(*args, **kwargs):
---> 35                 return func(*args, **kwargs)
     36 
     37             return wrapper

/data3/yangl/works/ai/phd/github/captum/captum/attr/_core/lrp.py in attribute(self, inputs, target, additional_forward_args, return_convergence_delta, verbose)
    197             # propagation and execute back-propagation.
    198             self._register_forward_hooks()
--> 199             normalized_relevances = self.gradient_func(
    200                 self._forward_fn_wrapper, inputs, target, additional_forward_args
    201             )

/data3/yangl/works/ai/phd/github/captum/captum/_utils/gradient.py in compute_gradients(forward_fn, inputs, target_ind, additional_forward_args)
    117         # torch.unbind(forward_out) is a list of scalar tensor tuples and
    118         # contains batch_size * #steps elements
--> 119         grads = torch.autograd.grad(torch.unbind(outputs), inputs)
    120     return grads
    121 

~/tools/miniconda3/envs/pytorch/lib/python3.8/site-packages/torch/autograd/__init__.py in grad(outputs, inputs, grad_outputs, retain_graph, create_graph, only_inputs, allow_unused)
    200         retain_graph = create_graph
    201 
--> 202     return Variable._execution_engine.run_backward(
    203         outputs, grad_outputs_, retain_graph, create_graph,
    204         inputs, allow_unused)

RuntimeError: hook 'backward_hook_activation' has changed the size of value
NarineK commented 3 years ago

@bugsuse, this is interesting. Maybe some activations got reused. Will you be able to share a collab notebooks so that we can debug it ? cc: @nanohanno

bugsuse commented 3 years ago

@NarineK see in colab notebook, the pretrained weights and test data is here. Please let me know if you need more infomation.

nanohanno commented 3 years ago

I have not been able to run the code yet to check the behaviour. Has it maybe something to do with the additional output added in 084d755f6f0e9f950f5966d13fbae15596c193e5 ?

BTW, I just saw that register_backward_hook(hook) is deprecated in the current PyTorch version and is exchanged by register_full_backward_hook(hook).

https://pytorch.org/docs/stable/generated/torch.nn.Module.html?highlight=backward_hook#torch.nn.Module.register_backward_hook

NarineK commented 3 years ago

@nanohanno, register_full_backward_hook isn't fully backward compatible with register_backward_hook. We are waiting for the new version to fix backward compatibility issue.

We were expecting additional arguments to be needed for dropouts. We need to debug this case. @vivekmig and @nanohanno do you have time to debug this issue ?

survivebycoding commented 2 years ago

@NarineK

model.layer.rule = EpsilonRule()

I tried to use this but my layer name is : layer_0/kernel:0 when i am using this name i am getting syntax error. I tried 'layer_0/kernel:0' and "layer_0/kernel:0" as well.