IdentityTask applies learnable layers

[niklasmei]

Describe the bug The IdentitiyTask (under graphnet.models.task) applies learnable parameters due to inheritance from LearnedTask and StandardLearnedTask.

To Reproduce Steps to reproduce the behavior:

1. Try and run a script like:

import torch
from graphnet.models.task import IdentityTask`

task1 = IdentityTask()
task2 = IdentityTask()

data = torch.ones(3,4)
print(data)

out1 = task1(data)
out2 = task2(data)

print(out1)
print(out2)

This will throw errors because the inhertance from LearnedTask and StandardLearnedTask require arguments like 'hidden_size'.

2. With a working code like

import torch
from graphnet.models.task import IdentityTask
from graphnet.training.loss_functions import MSELoss


task1 = IdentityTask(nb_outputs = 1, target_labels=['skip'], hidden_size=4, loss_function=MSELoss())
task2 = IdentityTask(nb_outputs = 1, target_labels=['skip'], hidden_size=4, loss_function=MSELoss())

data = torch.ones(3,4)
print(data)

out1 = task1(data)
out2 = task2(data)

print(out1)
print(out2)

you will see that the different tasks return different tensors when actually they should both return the input tensor.

Expected behavior The IdentityTask should return the input data and not apply learnable parameters

Additional context This problem became apparent for me only because, for my application of the EasySyntax base class, I needed a task, although I did not actually use it. When running a distributed training with the 'ddp' strategy I got the error that there are unused parameters and I needed to use 'ddp_find_unused_parameters'. Therefore I assume that this bug is not a high priority and at least for my use case I can get around it. However people who actually use the IdentityTask might take more issue with that.

[Aske-Rosted]

Hi @niklasmei , Thanks for raising this issue.

We discussed your issue at our recent developer meeting and agree that the docstring in IdentityTask is misleading. The IdentityTask has always comprised an affine transformation, and is therefore not an identity in the sense that x - task(x) = 0 . Instead the IdentityTask is an identity in the sense that the _forward implementation contains no augmentation of its input. We propose that this issue is resolved by providing a clear docstring that describes the behavior in higher detail. This is necessary as StandardLearnedTask expects the _forward call to be defined in the daughter classes through @abstractmethod.