Feat/clip annotation

[HonzaCuhel]

This PR includes the following changes:

• adding CLIP annotator for classification

[sokovninn]

Have you compared clip-vit-large-patch14 and clip-vit-base-patch32?

[github-actions]

☂️ Python Coverage

current status: ✅

Overall Coverage

Lines	Covered	Coverage	Threshold	Status
1015	499	49%	0%	🟢

New Files

File	Coverage	Status
datadreamer/dataset_annotation/clip_annotator.py	52%	🟢
TOTAL	52%	🟢

Modified Files

File	Coverage	Status
datadreamer/dataset_annotation/init.py	100%	🟢
datadreamer/pipelines/generate_dataset_from_scratch.py	41%	🟢
TOTAL	71%	🟢

updated for commit: 5cf88ee by action🐍

[github-actions]

Test Results

  6 files    6 suites   46m 19s ⏱️
 85 tests  36 ✅  49 💤 0 ❌
510 runs  216 ✅ 294 💤 0 ❌

Results for commit 5cf88ee.

♻️ This comment has been updated with latest results.

[HonzaCuhel]

No, I haven't, but I can test it.

[HonzaCuhel]

I've manually compared clip-vit-large-patch14 and clip-vit-base-patch32 annotation on 100 generated images on JP using L4 GPU, I've used batch size 8 for annotation and these are the results:

Latency:

• clip-vit-large-patch14 : 3s
• clip-vit-base-patch32 : 1s

Comparison:

• in 12 cases I found clip-vit-large-patch14's annotation better
• in 7 case I found clip-vit-base-patch32's annotation better

Conclusion

Switched to using clip-vit-base-patch32 for the speed as the performance gap isn't big.

[sokovninn]

Thanks, looks good! I agree with switching to CLIP base.