ludwig-ai / ludwig

Low-code framework for building custom LLMs, neural networks, and other AI models

http://ludwig.ai

Apache License 2.0

11.22k stars 1.19k forks source link

Add unet encoder-decoder and image output feature #3913

Closed vijayi1 closed 10 months ago

vijayi1 commented 10 months ago

Code Pull Requests

Please provide the following:

a clear explanation of what your code does Added a UNet encoder and decoder stack. Added an image output feature, and image utils to detect pixel classes. Added an image semantic segmentation example using a kaggle dataset. Added test cases for encoder, decoder and dataset.
if applicable, a reference to an issue
a reproducible test for your PR (code, config and data sample) Test example included in PR.

vijayi1 commented 10 months ago

@justinxzhao, I opened this PR and closed the previous PR 3870 due to git rebase issues. I've addressed your comments in this PR.

I'm attaching some sample results. Each result has the input image, the label mask and a predicted mask.

the camseq example included in the PR (https://www.kaggle.com/datasets/carlolepelaars/camseq-semantic-segmentation)
a buildings dataset, with images cropped due to mem limits on my system (https://www.kaggle.com/datasets/sohaibmoradi/segmented-satellite-images-of-buildings)

camseq_0016E5_07959 camseq_0016E5_07959_L camseq_predicted_0016E5_07959_L buildings_img_0 buildings_mask_0 buildings_mask_predicted_0

github-actions[bot] commented 10 months ago

Unit Test Results

  6 files ±0   6 suites ±0 13m 50s :stopwatch: -26s 12 tests ±0   7 :heavy_check_mark: -   2   5 :zzz: +  2 0 :x: ±0 60 runs ±0 30 :heavy_check_mark: - 12 30 :zzz: +12 0 :x: ±0

Results for commit 30feafb2. ± Comparison against base commit 9bb89c6c.

This pull request skips 2 tests.

``` tests.regression_tests.benchmark.test_model_performance ‑ test_performance[ames_housing.gbm.yaml] tests.regression_tests.benchmark.test_model_performance ‑ test_performance[mercedes_benz_greener.gbm.yaml] ```

vijayi1 commented 10 months ago

attached training results for the camseq example above.

camseq_training_log

justinxzhao commented 10 months ago

@vijayi1 This is super cool! Thanks for addressing prior comments and putting together the PR. LGTM!

vijayi1 commented 10 months ago

Thanks for the reviews. @arnavgarg1, I'll create follow up PRs for those two points.

I do not have write access to merge this PR. Please advise on the next step.

arnavgarg1 commented 10 months ago

@vijayi1 - Sounds good! Will keep a lookout for those PRs. We just merged your PR into master!

arnavgarg1 commented 9 months ago

Hey @vijayi1! This was a marquee feature in our Ludwig 0.10.0 release and we'd love to give you a shoutout on LinkedIn. Is there a way to connect? My LinkedIn is https://www.linkedin.com/in/arnavgrg/, feel free to connect!

vijayi1 commented 9 months ago

Hi @arnavgarg1, thanks for the shoutout. I'm not on LinkedIn, but I did lookup the release announcement.

arnavgarg1 commented 9 months ago

No worries! And glad you were able to see it :) Just wanted to make sure you knew and we really appreciate the contribution!

alexsherstinsky commented 9 months ago

@vijayi1 What is your handle in the Ludwig community Slack? Thank you!

vijayi1 commented 9 months ago

"Vijay I", I just joined.