Fix batching with oriented images and support larger image datasets #178

PaulHax · 2025-01-24T17:57:53Z

The shitspoter dataset has ~12 megapixel images.

Its slow to scroll through the image list. Fixed by serving images via endpoint, not pushing on state.
The image cache was filling up main memory.

Closes #92

- Use generator when computing embeddings. - Ajust image cache size based on image size and available memory.

Don't cache the full transformed image in embedding app. Was running out of memory.

alesgenova · 2025-02-07T16:09:08Z

src/nrtk_explorer/app/images/image_server.py

+    return file_path.split(".")[-1].upper() in COMPATIBLE_FORMATS
+
+
+def ensure_browser_compatable_format(image: Image.Image):


typo "compatable" -> "compatible"

PaulHax requested a review from alesgenova January 24, 2025 19:14

fix(predictor): use image orientation in batch grouping

1c99cfc

Closes #92

PaulHax force-pushed the batch-fix branch from 52c61c5 to 1c99cfc Compare January 24, 2025 20:40

PaulHax changed the title ~~fix(predictor): use image orieantion in batch grouping~~ fix(predictor): use image orientation in batch grouping Jan 24, 2025

fix: avoid out of memory errors with larger images

37672e9

- Use generator when computing embeddings. - Ajust image cache size based on image size and available memory.

PaulHax changed the title ~~fix(predictor): use image orientation in batch grouping~~ Fix batching with oriented images and support larger image datasets Jan 27, 2025

PaulHax added 5 commits January 28, 2025 15:26

perf: use image server rather than base64 images to state

1155250

perf: use image server for transform images

c20e9bd

fix(multiprocess_predictor): cancel requests when task cancelled

a43de11

refactor: store chained transform in images store

7c4a0d6

fix(test_predictor): dont test MultiprocessPredictor

54c0fed

PaulHax force-pushed the batch-fix branch from b35a2ea to 54c0fed Compare January 30, 2025 21:36

PaulHax added 3 commits February 3, 2025 17:02

fix(embeddings): only cache transformed image features

430a8e4

Don't cache the full transformed image in embedding app. Was running out of memory.

fix(multiprocess_predictor): avoid hanging process on ctrl-c

4aca58a

perf(transforms): smaller update images batch size

ee2e927

alesgenova reviewed Feb 7, 2025

View reviewed changes

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Fix batching with oriented images and support larger image datasets #178

Fix batching with oriented images and support larger image datasets #178

PaulHax commented Jan 24, 2025 •

edited

Loading

alesgenova Feb 7, 2025

		return file_path.split(".")[-1].upper() in COMPATIBLE_FORMATS


		def ensure_browser_compatable_format(image: Image.Image):

Fix batching with oriented images and support larger image datasets #178

Are you sure you want to change the base?

Fix batching with oriented images and support larger image datasets #178

Conversation

PaulHax commented Jan 24, 2025 • edited Loading

alesgenova Feb 7, 2025

Choose a reason for hiding this comment

PaulHax commented Jan 24, 2025 •

edited

Loading