AML workflow to train LULC models for PEARL

[srmsoumya]

LULC Pipeline uses AML(Azure Machine Learning) to store data, train models & evaluate scores.

To train a segmentation model for PEARL, we need access to

• AML workspace
• Labeled LULC class dataset as TIF or GeoJSON (we have to make a sample region available)

Our current workflow

1. Create labeled tiffs & train/valid/test CSVs for ML model using naip-utils scripts
2. Move the TIFFs to an Azure Blob storage
3. Move CSVs to data/ inside src
4. Create workspace, CPU & GPU clusters using scripts from train_azure
5. Run run_model.py script to get the most_recent_model.pt file.
6. Download the model & store inside src/model
7. Execute eval, seed_data & class_distribution scripts

These scripts are tied to our AML resources.

cc/ @geohacker