U-net implementation

[annikabrundyn]

Adapting this model from an example in the Lightning repo.

Needs to be tested on Kitti dataset which I've submitted a separate PR for.

[pep8speaks]

Hello @annikabrundyn! Thanks for updating this PR.

• In the file pl_bolts/models/vision/__init__.py:

Line 2:45: W292 no newline at end of file

• In the file tests/models/test_vision_models.py:

Line 56:1: W391 blank line at end of file

Comment last updated at 2020-09-25 00:26:33 UTC

[codecov]

Codecov Report

Merging #247 into master will increase coverage by 0.31%.
The diff coverage is 98.14%.

@@            Coverage Diff             @@
##           master     #247      +/-   ##
==========================================
+ Coverage   88.31%   88.63%   +0.31%     
==========================================
  Files          70       72       +2     
  Lines        3407     3538     +131     
==========================================
+ Hits         3009     3136     +127     
- Misses        398      402       +4     

Flag	Coverage Δ
#cpu	32.35% <33.33%> (-0.27%)	⬇️
#pytest	32.35% <33.33%> (-0.27%)	⬇️
#unittests	87.90% <98.14%> (+0.34%)	⬆️

Flags with carried forward coverage won't be shown. Click here to find out more.

Impacted Files	Coverage Δ
pl_bolts/models/vision/unet.py	98.03% <98.03%> (ø)
pl_bolts/datamodules/__init__.py	100.00% <100.00%> (ø)
pl_bolts/models/__init__.py	100.00% <100.00%> (ø)
pl_bolts/models/vision/__init__.py	100.00% <100.00%> (ø)
pl_bolts/utils/arguments.py	96.15% <0.00%> (ø)

---

Continue to review full report at Codecov.

Legend - Click here to learn more
Δ = absolute <relative> (impact), ø = not affected, ? = missing data
Powered by Codecov. Last update c2cdad7...a3d4991. Read the comment docs.

[oke-aditya]

Very Nice! U-Net was one of the major models.
U-Net can do semantic segmentation as well as work as convolutional autoencoder.

So where does this model exactly go into? Into pl_bolts.models.segmentation or pl_bolts.models.autoencoders
or simply superuseful models go to pl_bolts.models.vision

The original paper introduced this for medical imaging using semantic segmentation. So it is well known for this.

There are weights available for torch hub as well. Here is the repo. This has an implementation of dice loss as well. I'm unsure if it is available in bolts/lightning, and might be good to support.

The implementation is done is very very nice.

[Borda]

can we pls add simple tests for this U-Net?

[annikabrundyn]

can we pls add simple tests for this U-Net?

Yes! Will add simple tests for U-net today. I've also got the segmentation model that uses U-net as backbone and Kitti datamodule that we can add once this is merged.

[Borda]

I've also got the segmentation model that uses U-net as backbone and Kitti datamodule that we can add once this is merged.

cool, then we could upload the weights to S3 :]

[annikabrundyn]

can we pls add simple tests for this U-Net?

I've added a super simple test just to check that the forward gives the expected shape. Once we merge the Kitti datamodule, I'll add the segmentation model (which uses this Unet architecture) and a test to check it's performance on the dataset

Let me know if you'd like me to add any other tests!

[annikabrundyn]

So where does this model exactly go into? Into pl_bolts.models.segmentation or pl_bolts.models.autoencoders
or simply superuseful models go to pl_bolts.models.vision

This is a good question - this is basically just the UNet architecture so I'm adding a semantic segmentation model that uses this as a backbone. But yes - it's a useful model for lots of different applications so also wasn't sure where exactly it belongs! And thanks Aditya 😄