Resolve softmax issues with pytorch_model.py

[dilyabareeva]

Description

• Extraction of softmax layer from pytorch models sometimes lead to problem.
• We want to update the logic of softmax adjustment as follows:
  

Implemented changes

• Updated logic for softmax adjustments in the model depending on the model structure
• The 6 different cases from the photo represented in get_softmax_arg_model method (see comments)
• A better way to extract Softmax layer from model by replacing it with Identity (instead of rebuilding the model and forward function)

[codecov-commenter]

Codecov Report

Merging #283 (2b85111) into main (c288a7d) will increase coverage by 0.06%.
The diff coverage is 100.00%.

❗ Your organization is not using the GitHub App Integration. As a result you may experience degraded service beginning May 15th. Please install the Github App Integration for your organization. Read more.

@@            Coverage Diff             @@
##             main     #283      +/-   ##
==========================================
+ Coverage   93.29%   93.35%   +0.06%     
==========================================
  Files          62       62              
  Lines        3430     3448      +18     
==========================================
+ Hits         3200     3219      +19     
+ Misses        230      229       -1     

Files Changed	Coverage Δ
quantus/helpers/model/pytorch_model.py	87.05% <100.00%> (+2.75%)	⬆️

[annahedstroem]

A few things! Thanks @dilyabareeva !

[annahedstroem]

Can you double-check the tf implementation so that the function names as well as the logic are replicated there?

[dilyabareeva]

Both use the same ModelInterface - so the public methods are the same. The logic in tf_model might need to be changed similarly: I'm not 100% sure that model.layers[-1] reliably returns the actual output layer in tensorflow. But this PR is about PyTorch, so maybe this is worth opening another issue to further investigate.

[aaarrti]

@dilyabareeva You are modifying the structure of the model provided by the user, it would be great to make it visible to the user what exactly is being done under the hood, and what assumptions were made.
E.g., for the case when the last softmax layer is replaced by a linear one, we could log a message:

<XAI method name> reuires model with linear top, but found softmax activation on top for <model name>. Removed <layer name> at <layer index>

This must not be a warning, since this is expected behaviour. I'd suggest using logging.info, and set log format to something like "[%(filename)s:%(lineno)s->%(funcName)s()]:%(levelname)s: %(message)s", so the user can see where it is coming from.

[annahedstroem]

OK.