GenHack2 - Hackathon for Generative modeling : Simulation of global warming Sea Surface Temperatures (website)

This project is a joint work with @saschagaudlitz and @ziebelle.

Model development

See models/maf.py for a working example.

You need to subclass torch's nn.Module and implement forward, sample and loss methods.

In the end, import contents of your module in the models/__init__.py file and add your model to the models dictionary, for example:

from .maf import *
...

models = {
    ...
    'MAF': MAF,
}

MLFlow

We use MLFlow for model tracking.

pip install mlflow

Then start the MLFlow from the root directory of the project (better specify absolute path due to ray):

mlflow server --backend-store-uri=sqlite:///mlflow.db --default-artifact-root=/Users/konstantin/projects/Flash/mlruns

Then you can access the server under 127.0.0.1:5000.

Note that in mlflow you can run an experiment that has several runs. For example, for hyperparameter tuning or cross-validation, you would run one experiment with multiple runs.

Running models

From command line

You can run a model from the root directory of the project by specifying the name of the class and the config argument as follows:

MLFLOW_TRACKING_URI=sqlite:///mlflow.db python3 run.py --config=configs/vae.yaml

From notebook

You can also run the model from the notebook, see train.ipynb for the example. In the notebook you can monitor 2d-marginals graphically during training.

Hyperparameter tuning

You need to install hyperopt and ray for hyperparameter tuning. You also need to specify the MLFLOW_TRACKING_URI environment variable. Note that you always need to specift the absolute path to the database file.

MLFLOW_TRACKING_URI=sqlite:////Users/konstantin/projects/Flash/mlflow.db pip install hyperopt ray

See hypertune.py for example script.

Testing model

Testing metrics and plots are recorded after the training run. Alternatively, you can run testing separately (for example, with another seed):

python3 run.py --config=configs/maf.yaml --mode=test --run_id=1af75f70444d43999157bbf6c20bccda