[Fix] Transformation Matrix Mis-calculation for autoaugmentations

[shijianjian]

Fixes #2843

[shijianjian]

weird. It passed on my local machine, but not here.

[johnnv1]

weird. It passed on my local machine, but not here.

I'm not sure if it's me who added the skip, but if it was, that's why it was like randomness 😢

maybe it's just an adjust in the tolerance?

[shijianjian]

weird. It passed on my local machine, but not here.

I'm not sure if it's me who added the skip, but if it was, that's why it was like randomness 😢

maybe it's just an adjust in the tolerance?

It is added by Dmytro I think. The point is after my fix, it totally works on my machine under Python3.8 and torch1.9.1. I do not know what is going on in the CI. XD

[johnnv1]

It is added by Dmytro I think. The point is after my fix, it totally works on my machine under Python3.8 and torch1.9.1. I do not know what is going on in the CI. XD

in my local it is randomly falling xD

$ conda create -p venv38 python=3.8
$ conda install pytorch=1.9.1 cpuonly -c pytorch -c nvidia
$ pip install accelerate==0.20.3 # otherwise will install new torch
$ pip install -e .[dev,x]
# run it until fail
$ pytest tests/augmentation/test_auto_operation.py --device=cpu --dtype=float64,float32 -vv

some times the AILED tests/augmentation/test_auto_operation.py::TestOperations::test_step_routine[op5] (the op id change in each fail run too) also fail with

E       assert 0.5 != 0.5
E        +  where 0.5 = <built-in method item of Tensor object at 0x7ae54b77a900>()
E        +    where <built-in method item of Tensor object at 0x7ae54b77a900> = tensor([0.5000], grad_fn=<ClampBackward1>).item
E        +      where tensor([0.5000], grad_fn=<ClampBackward1>) = HorizontalFlip(\n  (op): RandomHorizontalFlip(p=0.5, p_batch=1.0, same_on_batch=False)\n).probability

btw, the number of warnings also changes a lot between each run (like between 5 and 13 warning)