Inference

Obtain Docker Image

Either build the docker image or pull it from quay.io.

Build

Note that the model bundles necessary to build this image are not checked into this repository. The model bundless can be obtained by typing the following.

cd inference/
aws s3 sync s3://azavea-cloud-model/models models --request-payer requester

The docker image can be built by typing the following (with or without the change of directory, as appropriate).

cd inference/
docker build -f Dockerfile -t quay.io/jmcclain/cloud-model:latest .

Pull from quay.io

docker pull quay.io/jmcclain/cloud-model:latest

Perform Inference

cd inference/
docker run -it --rm \
       --runtime=nvidia --shm-size 16G \
       -v $HOME/Desktop/imagery:/input:ro \
       -v /tmp:/output \
       quay.io/jmcclain/cloud-model \
          --infile /input/greenwhich/L2A-0.tif \
          --outfile-final /output/final.tif \
          --outfile-raw /output/raw.tif \
          --level L2A \
          --architectures both

Training

Build Docker Image

cd training/
docker build -t azavea-cloud-model-training -f Dockerfile .

Run Container

cd training/
docker run -it --rm \
       --name azavea-cloud-model-training --runtime=nvidia \
       --shm-size 16G \
       -v $HOME/.aws:/root/.aws:ro \
       azavea-cloud-model-training bash

Invoke Raster-Vision

Local

export AWS_REQUEST_PAYER=requester
export ROOT=/tmp/xxx
rastervision run inprocess /workdir/pipeline.py \
       -a root_uri ${ROOT} \
       -a analyze_uri ${ROOT}/analyze \
       -a chip_uri ${ROOT}/chips \
       -a json catalogs.json \
       -a epochs 2 \
       -a batch_sz 2 \
       -a small_test True \
       chip train

On AWS

It is required to have a compute environment with additional storage for the p3.2xlarge batch instance that is used for training. (The large number of chips will not fit on a volume of the default size.)

Chip

export AWS_REQUEST_PAYER='requester'
export LEVEL='L1C'
export ROOT='s3://bucket/prefix'
rastervision run batch /workdir/pipeline.py \
       -a root_uri ${ROOT}/xxx \
       -a analyze_uri ${ROOT}/${LEVEL}/analyze \
       -a chip_uri ${ROOT}/${LEVEL}/chips \
       -a json catalogs.json \
       -a level ${LEVEL} \
       -s 800 \
       chip

Train

export LEVEL='L1C'
export ARCH='cheaplab'
export ROOT='s3://bucket/prefix'
rastervision run batch /workdir/pipeline.py \
       -a root_uri ${ROOT}/${ARCH}-${LEVEL} \
       -a analyze_uri ${ROOT}/${LEVEL}/analyze \
       -a chip_uri ${ROOT}/${LEVEL}/chips \
       -a json catalogs.json \
       -a level ${LEVEL} \
       -a architecture ${ARCH} \
       train

Licenses

The code in this repository is licensed under the terms given in the LICENSE.md file.

The dataset is licensed under the terms of the Creative Commons Attribution 4.0 International License.