Strange import issue --> AttributeError: module 'captum' has no attribute 'attr'

[adaruna3]

🐛 AttributeError: module 'captum' has no attribute 'attr'

When importing captum with just import captum, the package is missing modules like attr. Running dir(captum) shows that the other modules, like attr are missing. However, if a statement like from captum.attr._utils.input_layer_wrapper import ModelInputWrapper is included after import captum, the captum package then has the missing modules. Again, verified by running a dir(captum). Is this expected behavior?

To Reproduce

Steps to reproduce the behavior:

1. Run the code below:

import captum
print(dir(captum))
print("---------------")
from captum.attr._utils.input_layer_wrapper import ModelInputWrapper
print(dir(captum))
exit()

1. The above should get an output of the following:

['__builtins__', '__cached__', '__doc__', '__file__', '__loader__', '__name__', '__package__', '__path__', '__spec__', '__version__']
---------------
['__builtins__', '__cached__', '__doc__', '__file__', '__loader__', '__name__', '__package__', '__path__', '__spec__', '__version__', '_utils', 'attr', 'log']

Expected behavior

I assume all modules should be available after import captum. If this behavior is not expected, please close and disregard.

Environment

I installed captum using the suggested conda method. See details below. After noticing this issue in an environment I was developing in with many other packages, I made a new, clean environment that only contained captum and any dependencies conda installed. Same issue.

• Captum / PyTorch Version (e.g., 1.0 / 0.4.0): 0.3.1
• OS (e.g., Linux): Ubuntu 18.04.4 LTS
• How you installed Captum / PyTorch (conda, pip, source): conda
• command: conda install captum -c pytorch
• Python version: 3.6.13

[NarineK]

Thank you for raising the issue, @adaruna3 ! We will look into it.

[vivekmig]

Hi @adaruna3 , I think this is likely expected behavior, since importing a module doesn't automatically import all submodules. This would only occur if the main captum init imported the specific submodules, which is not currently the case, so you would need to either do import captum.attr or from captum.attr import * to import the attr submodule. These links also discuss similar issues (1, 2)

[adaruna3]

Hi Vivek, totally understand. Ultimately it depends on the design choices of the developers. I am used to importing the main package captum, torch, or numpy for example, and then expect to be able to access the highly utilized modules intended for public use within that package. Like with torch, I import torch but have access to nn, multiprocessing, optim, etc. and with numpy I have access to random, linalg, etc

I just mentioned it in case that behavior was not caught yet. But if having an import for each module in the package is intended design, that works too. Thanks!!