Add depth module

[ejm714]

This implementation takes in a folder of videos (or csv with video filepaths) and outputs a csv of distance estimates for each frame (and detection).

For frames with no detection, distance is null. If there are multiple detections in that frame, there is one line per detection. For example

filepath,time,distance
tests/assets/videos/data/raw/goualougo_2013/gorillas_2013/MPI_FID_20_Duboscia/14-Jun-2013/FID_20_Duboscia_2013-6-14_0083.AVI,0,7.4
tests/assets/videos/data/raw/goualougo_2013/gorillas_2013/MPI_FID_20_Duboscia/14-Jun-2013/FID_20_Duboscia_2013-6-14_0083.AVI,0,7.4
tests/assets/videos/data/raw/goualougo_2013/gorillas_2013/MPI_FID_20_Duboscia/14-Jun-2013/FID_20_Duboscia_2013-6-14_0083.AVI,1,
tests/assets/videos/data/raw/goualougo_2013/gorillas_2013/MPI_FID_20_Duboscia/14-Jun-2013/FID_20_Duboscia_2013-6-14_0083.AVI,2,
tests/assets/videos/data/raw/goualougo_2013/gorillas_2013/MPI_FID_20_Duboscia/14-Jun-2013/FID_20_Duboscia_2013-6-14_0083.AVI,3,

Not included in this PR:

• Masking out other detections where there is more than 1. Under the current implementation, if there are two detections in the frame, the distance estimate will be the same for both detections (since the same input is used for both). This allows us to capture the number of detections in the frame, but not differential distances.

Notes:

• In the original implementation, there was a bug where the order of the first two frames were swapped: [-1, -2, 0, 1, 2]. This bug is maintained to match what the model expects

Closes https://github.com/drivendataorg/pjmf-zamba/issues/57

Remaining to do:

• update docs home page to include depth model (and update CLI output)

[netlify]

✅ Deploy Preview for silly-keller-664934 ready!

Name	Link
🔨 Latest commit	56d9ffd
🔍 Latest deploy log	https://app.netlify.com/sites/silly-keller-664934/deploys/63880f947eb2990009085418
😎 Deploy Preview	https://deploy-preview-247--silly-keller-664934.netlify.app
📱 Preview on mobile	Toggle QR Code... Use your smartphone camera to open QR code link.

---

To edit notification comments on pull requests, go to your Netlify site settings.

[github-actions]

🚀 Deployed on https://deploy-preview-247--silly-keller-664934.netlify.app

[codecov-commenter]

Codecov Report

Merging #247 (56d9ffd) into master (0b9a901) will increase coverage by 0.5%.
The diff coverage is 91.7%.

Additional details and impacted files

@@           Coverage Diff            @@
##           master    #247     +/-   ##
========================================
+ Coverage    87.0%   87.6%   +0.5%     
========================================
  Files          28      31      +3     
  Lines        1964    2164    +200     
========================================
+ Hits         1709    1896    +187     
- Misses        255     268     +13     

Impacted Files	Coverage Δ
zamba/models/publish_models.py	0.0% <0.0%> (ø)
zamba/cli.py	84.8% <74.1%> (-2.0%)	⬇️
zamba/models/config.py	97.4% <95.0%> (+0.5%)	⬆️
zamba/models/depth_estimation/config.py	96.3% <96.3%> (ø)
zamba/models/depth_estimation/depth_manager.py	99.1% <99.1%> (ø)
zamba/models/densepose/config.py	91.8% <100.0%> (-0.7%)	⬇️
zamba/models/depth_estimation/__init__.py	100.0% <100.0%> (ø)
zamba/pytorch/transforms.py	80.3% <100.0%> (+1.1%)	⬆️
zamba/data/video.py	82.4% <0.0%> (+0.7%)	⬆️

[ejm714]

Comments from synchronous code review

[pjbull]

Docs seem great—just one thought on a clarifying intro sentence!

[ejm714]

All PR comments have been addressed. Checked memory usage with scalene and confirmed our predictions are within 0.1 of the competition code, often times closer to a difference of 0.03. This is likely because the competition code does some writing out and reading in of frames with open cv and also uses albumentations to resize (which uses open cv under the hood). We use torchvision transforms instead. Distances are usually measured in increments of 0.5 so these differences don't seem problematic.

Tests are passing locally so I need to look into why they're failing on github.

@pjbull if you have a moment to give the diff a quick read as a backstop, that'd be great. Here's the link