when there are many input features ,how to use the IntegratedGradients?

[Sarah-air]

❓ Questions and Help

the classification model is classification_pred = model([input_x[0], input_x[1],input_x[2], input_x[3]], os_event.size(0),
 {'gex': 0, 'methy': 1, 'mut':2, 'cna': 3}),the input_x of one sample is as the figure showed,how to use the IntegratedGradients to get the four features' attribution?

[khairulislam]

Hi @Sarah-air , if your model uses all 4 inputs, then normal passing is fine. but if it doesn't, you'll have to pass only the inputs that your model uses to the attribute method as input and the rest as additional forward args. The Gradient method throws an error if any input is unused. Check our example here.

[phet1999]

Hi @khairulislam, I have the same problem with this bro and I don't know how to utilize this code for multi-inputs. For example, if I have four different inputs for the model. Is it the code should be like this?

'''ig = IntegratedGradients(model) input1 = torch.Tensor(train_X_hist[:20]) input2 = torch.Tensor(train_X_ex_hist[:20]) input3 = torch.Tensor(train_X_ex_fu[:20]) input4 = torch.Tensor(train_X_static[:20]) input = (input1,input2,input3,input4) baseline1 = torch.zeros_like(input1) baseline2 = torch.zeros_like(input2) baseline3 = torch.zeros_like(input3) baseline4 = torch.zeros_like(input4) baseline = (baseline1,baseline2,baseline3,baseline4) IG, delta = ig.attribute(input,baseline, target=1, return_convergence_delta=True)'''

But it reports that

"AttributeError: 'tuple' object has no attribute 'shape'"

[khairulislam]

Hi @phet1999 , try interpreting one input at a time. Then add the returned interpretations in a tuple. You could have a variable inside the model class to set which input no is being passed. Then the default call method will arrange them 1-4 and call the original call method. the gradient calculation is w.r.t one input only, hence it won't directly support multiple inputs.

additional forward args = (input2,input3,input4)
baseline = (baseline2,baseline3,baseline4)
IG, delta = ig.attribute(
  input1,baseline, target=1, return_convergence_delta=True, 
  additional_forward_args=additional_forward_args
)

[phet1999]

Hi @khairulislam, thank you for your replying, but problems still exist. Here is the code, I first define a new forward function. ''' def forward(self, x_extra_future: torch.Tensor, *args ) -> torch.Tensor: '''

Then, I want to get the attribution of the first input feature ''' additional_forward_args = (input2,input3,input4) baseline = baseline1 IG, delta = ig.attribute( input1,baseline1, target=1, return_convergence_delta=True, additional_forward_args=additional_forward_args ) '''

but it still comes with a problem that ''' AttributeError: 'tuple' object has no attribute 'shape' '''

[phet1999]

Hi, @khairulislam. I am sorry for bothering and I tried another method that sets

model.input1 = torch.Tensor(train_X_hist[:20]).to(device)
model.input2= torch.Tensor(train_X_ex_hist[:20]).to(device)
model.input3 = torch.Tensor(train_X_static[:20]).to(device)

and inside forward function, it acts like this

x_hist,x_extra_hist,x_static = self.input1,self.input2,self.input3

In this situation, there is just one input in the forward function. But the problem is still not solved for this code

ig = IntegratedGradients(model)
    IG, delta = ig.attribute(
    input, baseline, target=1, return_convergence_delta=True
    )

It said that

RuntimeError: Sizes of tensors must match except in dimension 2. Expected size 1000 but got size 20 for tensor number 1 in the list.

It seems that the code repeats my input 50 times at the batch channel(the same as the number of parameters n_steps, and when I set it to 1, the problem is 'AttributeError: 'tuple' object has no attribute 'shape'')