[Tests] Add tests for LMS Discrete

[patrickvonplaten]

Similar to other tests we should add tests for the LMS Discrete Scheduler in https://github.com/huggingface/diffusers/blob/main/tests/test_scheduler.py

[patrickvonplaten]

cc @anton-l

[sidthekidder]

Btw, the cmd: pytest tests/test_scheduler.py fails on my system (macOS 11.2.3, Python 3.8.2, Apple clang version 12.0.5) with the error:

FAILED tests/test_scheduler.py::ScoreSdeVeSchedulerTest::test_full_loop_no_noise - AssertionError: assert 2048.0 < 0.01

>       assert abs(result_sum.item() - 14379591680.0) < 1e-2
E       AssertionError: assert 2048.0 < 0.01
E        +  where 2048.0 = abs((14379589632.0 - 14379591680.0))
E        +    where 14379589632.0 = <built-in method item of Tensor object at 0x15f7f0cc0>()
E        +      where <built-in method item of Tensor object at 0x15f7f0cc0> = tensor(1.4380e+10).item

tests/test_scheduler.py:800: AssertionError

[patrickvonplaten]

Btw, the cmd: pytest tests/test_scheduler.py fails on my system (macOS 11.2.3, Python 3.8.2, Apple clang version 12.0.5) with the error:

FAILED tests/test_scheduler.py::ScoreSdeVeSchedulerTest::test_full_loop_no_noise - AssertionError: assert 2048.0 < 0.01

>       assert abs(result_sum.item() - 14379591680.0) < 1e-2
E       AssertionError: assert 2048.0 < 0.01
E        +  where 2048.0 = abs((14379589632.0 - 14379591680.0))
E        +    where 14379589632.0 = <built-in method item of Tensor object at 0x15f7f0cc0>()
E        +      where <built-in method item of Tensor object at 0x15f7f0cc0> = tensor(1.4380e+10).item

tests/test_scheduler.py:800: AssertionError

Yes for us as well @sidthekidder - we're looking into fixing this one in #268

[sidthekidder]

Added some code for the LMS Discrete scheduler similar to the DDIM scheduler test. Though I still haven't understood some parts of this scheduler, so the test_full_loop_no_noise is failing and I'm not sure how to fix it. The value of the residual var here seems to be a nan matrix, leading to the final sample output being nan as well. I don't understand how the dummy_model fn here mocks a real model as well, what does the value being returned mean.
In general, I couldn't find more details of a "LMS discrete scheduler" online, so not sure how to better understand the test_full_loop_no_noise function - would appreciate any tips or help! @patrickvonplaten @patil-suraj

[patrickvonplaten]

@anton-l can you help here?

[anton-l]

Hi @sidthekidder, checking your PR, will comment there soon! 🤗