demonstrate a dockerized microservice that uses pytorch with CUDA acceleration

docker-ce v18.06.0-ce + nvidia-docker2

pytorch

pip install docker-compose (not in conda repositories)

create performance tests, and figure out how to perf-test differences between versions

Deploy to GCP, and verify GPU acceleration

install in developer mode from source

python setup.py develop

start the web server manually

python pytorch_server/pytorch_server.py

run tests against the web server

python -m unittest functional_tests/service_test.py
pytest functional_tests/service_test.py

Install the package

python setup.py install
python setup.py develop

start webserver with cli

pytorch_server start

Run functional tests

pytorch_cli test service
pytest functional_tests/service_test.py

to validate the docker container works properly, and docker-compose works

Run the container with docker-compose and test it with service_test

docker-compose build
docker-compose up
pytest integraiton_tests/service_test.py

container_test will run docker-compose before each test case, and execute the test against the running container

python -m unittest container_tests/container_test.py
pytest container_tests/container_test.py

sometimes necessary if container tests are failing

docker-compose down --rmi local
docker-compose build