This repo is my first look into Data Version Control. It is an attempt at improving reproducibility and rigour for my deep learning experiments and hopefully will enable me to track my experiments in a properly documented fashion going forward in my PhD.
The goal of this mini-project is to segment cats.
The data is private to the University of Sheffield and only staff members can access it. I will likely upload open datasets in the future.
For UoS staff who want to try this out:
-
Clone this repo
-
Create a new conda environment (works on 3.10 for me)
-
Install the required packages
pip install -r requirements.txt -
Download the data automatically using DVC:
dvc pull. This will take a few minutes as Google Drive's servers are slow. -
You can then have a look at the experiments via
dvc exp showanddvc exp list -
You can have a look at the processing pipeline via
dvc dagandcat .dvc.yaml -
Run an experiment simply by running
dvc exp run. It's likely that it will say that there are no updates and so it won't run anything - this is intended! It will only run an experiment if the result will change, so try editing a parameter like batch size or epochs:dvc exp run --set-param "train.batch_size=10" -
Multiple experiments can be queued via
dvc exp run --queue --set-param "train.batch_size=8,16,24"for example. (Not done this yet but it's in the docs). -
Also use the Visual Studio Code DVCLive extension to see the cool graphs and tables comparing experiments!
If you want to upload your own experiment, try creating a branch with a descriptive name and running one, change the code if you like too, DVC seems to be able to handle that. Then message me to add you as a contributor and you can push your experiment.
I'll be looking into using DVC Studio soon to host the experiments. Here is a link to some info on sharing experiments.