Augmentation SpeedUp fix

[sokovninn]

This PR includes:

• Resize() after all other augmentations to ensure the correct input shape
• Optimizing amount of Resize() calls

Some benchmarks on the COCO train dataset (1 epoch, 4 workers, bs 16):

• Mosaic + set of other augmentations (do change shape):

  • resize before augs + resize if needed after augs: 22:26
  • resize after augs: 20:41
  • resize before normalization only + resize if needed after augs: 20:57 - chosen solution

• Only mosaic augmentation (doesn't change shape):

  • resize before augs + resize if needed after augs: 14:53
  • resize after augs: 15:42
  • resize before normalization only: 15:24
  • resize before normalization only + resize if needed after augs: 14:49 - chosen solution

Update:

• Add separate pixel transformations
• Resize if needed before pixel transformations

COCO train dataset (normalize, defocus, sharpen, rotate, perspective, affine, randomscale (all with p=1)):

• Previous approach: 19:33
• batch-> spatial -> resize (if shapes are changed) -> pixel augs: 19:16

Original PR with a bug (reverted)

[github-actions]

Test Results

  1 files    1 suites   4m 28s ⏱️
102 tests 102 ✅ 0 💤 0 ❌
114 runs  114 ✅ 0 💤 0 ❌

Results for commit 178a395.

♻️ This comment has been updated with latest results.

[github-actions]

☂️ Python Coverage

current status: ✅

Overall Coverage

Lines	Covered	Coverage	Threshold	Status
3697	3091	84%	0%	🟢

New Files

No new covered files...

Modified Files

File	Coverage	Status
luxonis_ml/data/augmentations/utils.py	97%	🟢
TOTAL	97%	🟢

updated for commit: 178a395 by action🐍

[kozlov721]

The return type annotation and docs of _parse_cfg are outdated

[kozlov721]

LGTM

[klemen1999]

Left some comments otherwise LGTM