seaflow-uw/mpm-workflow

SeaFlow matrix population model workflow

SeaFlow matrix population model workflow repository

The basic workflow is to generate SeaFlow particle size distribution and PAR data using the R notebook create_psd.Rmd. Then create plots and run the model using the Python script fit_models.py. To run model days in parallel use the Bash script parallel_run_model.sh.

Install dependencies

R package popcycle

Install the R package popcycle.

Python package PyStan

This workflow has been tested with Python 3.8 and PyStan version 2.19.1.1. Either 1) install PyStan into a Python virtual environment or 2) build/use a Docker image.

Python virtual environment for PyStan

Create a Python virtual environment using whichever tool you'd like and install pystan2 requirements. Here we'll use poetry with Python 3.8.

poetry install
# ... do poetry things
# Export requirements if any have changed
poetry export >requirements-pystan2.txt

Or to use a standard virtual environment

python3 -m venv ./venv
source ./venv/bin/activate
pip3 install -U setuptools wheel pip
pip3 install -r requirements-pystan2.txt
deactivate

#### Docker build PyStan2

To build a Docker image tagged as `seaflow-model-pystan2` with pystan2 installed

```bash
docker build -t seaflow-model-pystan2 --build-arg REQUIREMENTS_FILE=requirements-pystan2.txt -f Dockerfile .

You may want to tag this image with your Docker Hub account and upload to Docker Hub.

docker tag seaflow-model-pystan2 myaccount/seaflow-model-pystan2:version
docker push myaccount/seaflow-model-pystan2:version

Run

Generate model input data

Customize and run the R markdown notebook create_psd.Rmd. This notebook uses renv to manage R package dependencies. If you don't use renv then comment out the invocation near the top of the file.

Run the model

The Stan model and data plots can be run with fit_models.py. For command-line help run fit_models.py --help. There are four subcommands: days, compile, model, and plot-cruise.

days will print a table of cruise days found in either the PSD or PAR Parquet files.

compile will compile the Stan model and cache it in the current working directory. This can save up to a minute per model run. It isn't strictly necessary to run this step separately as the default is to cache the model everytime the model is run.

model will run the model on some or all of the data and at the same time create a plot of input data for every day processed.

plot-cruise will create a plot of input data for all days in a cruise.

Docker run examples

Run fit_models.py in a Docker container for a single day

docker run -it --rm -v $(pwd):/data -w /data seaflow-model-pystan2:latest \
  sh -c 'python3 fit_models.py model \
    --psd-file combined.psd-hourly.parquet \
    --grid-file combined.psd-grid.parquet \
    --par-file combined.par-hourly.parquet \
    --stan-file m_pmb_sigprior_v2.stan --model-name m_pmb \
    --desc Pro \
    --cruise HOT303 \
    --days 2 \
    --output-dir out 2>&1 | tee model-run-HOT303-day2.log'

Run all cruise/day combinations in parallel using GNU parallel

Modify and run parallel_run_model.sh in a Docker container. GNU parallel is already installed in the Docker container.

docker run -it --rm -v $(pwd):/data -w /data seaflow-model-pystan2:latest \
  ./parallel_run_model.sh

Run in EC2

To run models in AWS EC2, use terraform to create an EC2 instance, then use ansible to provision this machine, and finally upload files and log in with SSH to run the model.

Create EC2 instance

First install the command-line tool terraform. Then optionally modify variables.tf in terraform/ondemand to change the type of EC2 instance. Then run terraform.

cd terraform/ondemand
terraform init
terraform apply
cd ../..

This command will start an EC2 instance and print its public IP address. Take this address and create an SSH host entry in ~/.ssh/config that will be used by ansible to provision the EC2 instance. In this example the host is named psd.

Host psd
    Hostname 34.209.151.135
    IdentityFile ~/.ssh/aws.pem
    User ubuntu

Log in to the instance with ssh psd

Provision EC2 instance

Install ansible and make sure it's accessible in your current environment. Then modify ansible/inventories/psd.yml to change the name of the Docker image containing pystan2 to match the Docker Hub location you previously pushed to. Finally provision the instance with ansible. This installs Docker and pulls your pystan2 image.

ansible-playbook -i ansible/inventories/psd.yml ansible/playbook.yml

Run the model

Upload files data and code files to the EC2 instance and run the model with Docker.

scp -r combined.*.parquet m_pmb_sigprior_v2.stan fit_models.py parallel_run_model.sh psd:

Then SSH into the psd EC2 instance and run fit_models.py following the Docker examples earlier in this document.

Terminate the EC2 instance

cd terraform/ondemand
terraform destroy

It's probably a good idea to log into the EC2 web console to confirm the instance has terminated.