1 scoping

[lizgzil]

---

Description

This is a refactor of the work Adeola did in training the model, but this time using batch.

The full flow can be run by running:

cd asf_floorplan_interpreter/pipeline/
python train_yolo.py --package-suffixes=.txt,.yaml,.jpg --datastore=s3 run --config_file window_door_config.yaml

or the smaller test version:

cd asf_floorplan_interpreter/pipeline/
python train_yolo.py --package-suffixes=.txt,.yaml,.jpg --datastore=s3 run --config_file window_door_test_config.yaml

This will train the model and output the best model in this S3 location.

The metaflow files are in this location: https://s3.console.aws.amazon.com/s3/buckets/open-jobs-lake?prefix=metaflow/FloorPlanYolo/&region=eu-west-1

It is super quick too! took about 5-10 minutes to train.

Still lots to do - e.g. add newly labelled data, make sure the validation is outputting properly? But I think thats for another PR!

Fixes #5

Instructions for Reviewer

Try out

export METAFLOW_PROFILE=floorplan
cd asf_floorplan_interpreter/pipeline/
python train_yolo.py --package-suffixes=.txt,.yaml,.jpg --datastore=s3 run --config_file configs/window_door_test_config.yaml

you will need to create the correct metaflow config (and save it in ~/.metaflowconfig/config_floorplan.json), which I can send you.

Checklist:

• [ ] I have refactored my code out from notebooks/
• [ ] I have checked the code runs
• [ ] I have tested the code
• [ ] I have run pre-commit and addressed any issues not automatically fixed
• [ ] I have merged any new changes from dev
• [ ] I have documented the code
  • [ ] Major functions have docstrings
  • [ ] Appropriate information has been added to READMEs
• [ ] I have explained this PR above
• [ ] I have requested a code review

[lizgzil]

Also includes edits to the prodigy_to_yolo.py script.

As of the 191023 dataset we have:

rooms dataset

Original data was 122 annotations, filtered for accepted labels and deduplicated gives us 105 annotations

• Saving 63 image in the training set
• Saving 21 image in the test set
• Saving 21 image in the val set

window/door/staircase dataset

Original data was 119 annotations, filtered for accepted labels and deduplicated gives us 102 annotations

• Saving 61 image in the training set
• Saving 20 image in the test set
• Saving 21 image in the val set

[lizgzil]

@cmbrennan002 I've addressed your comments in https://github.com/nestauk/asf_floorplan_interpreter/pull/4/commits/200d5113d40c31f85348e4ca20df84b6ae4ee543 🎉 - thanks for the review!

Am I ok to merge now?