simvue 0.0.8

Simulation tracking

Simvue Python client

Collects metadata, metrics and files from simulations in real time.

Configuration

The service URL and token can be defined as environment variables:

export SIMVUE_URL=...
export SIMVUE_TOKEN=...

or a file simvue.ini can be created containing:

[server]
url = ...
token = ...

The exact contents of both of the above options can be obtained directly by clicking the Create new run button on the web UI. Note that the environment variables have preference over the config file.

Usage example

from simvue import Simvue

...

if __name__ == "__main__":

    ...

    # Using a context manager means that the status will be set to completed automatically
    with Simvue() as run:

        # Specify a run name, metadata (dict), tags (list), description, folder
        run.init('example-run-name',
                 {'learning_rate': 0.001, 'training_steps': 2000, 'batch_size': 32}, # Metadaata
                 ['tensorflow'],                                                     # Tags
                 'This is a test.',                                                  # Description
                 '/Project-A/part1')                                                 # Folder full path

        # Set folder details if necessary
        run.set_folder_details('/Project-A/part1',                     # Folder full path
                               metadata={},                            # Metadata
                               tags=['tensorflow'],                    # Tags
                               description='This is part 1 of a test') # Description

        # Upload the code
        run.save('training.py', 'code')

        # Upload an input file
        run.save('params.in', 'input')

        # Add an alert (the alert definition will be created if necessary)
        run.add_alert('loss-too-high',      # Name
                      'is above',           # Type
                      'loss',               # Metric
                      1,                    # Frequency
                      1,                    # Window
                      threshold=10,         # Threshold
                      notification='email') # Notification type

        ...

        while not converged:

            ...

            # Send metrics inside main application loop
            run.log({'loss': 0.5, 'density': 34.4})

            ...

        # Upload an output file
        run.save('output.cdf', 'output')

        # If we weren't using a context manager we'd need to end the run
        # run.close()