nnabla-mlflow

MLflow utilities for NNabla.

dependencies

• nnabla
• mlflow

install

$ pip install git+https://github.com/takuseno/nnabla-mlflow

usage

Calling nnabla_mlflow.autolog monkey-pathches NNabla's monitors to automatically save data as MLflow metrics.

import mlflow
import nnabla_mlflow

from nnabla.monitor import MonitorSeries

nnabla_mlflow.autolog() # monkey-patch Monitor classes

metric_monitor = MonitorSeries('metric', interval=1)

with mlflow.start_run():
    metric_monitor.add(1, 2) # mlflow.log_metric('metric', 1, step=2) is internally called

Currently, the following monitors are supported.

• MonitorSeries
• MonitorTimeElapsed
• MonitorImage
• MonitorImageTile

To automatically save parameters as an artifact, pass with_save_parameters=True.

nnabla_mlflow.autolog(with_save_parameters=True)

TODO

• support MonitorImage and MonitorImageTile
• support log_model and save_model just like mlflow.tensorflow and mlflow.pytorch
• support NNabla's Trainer