Train use mlflow #287
Train use mlflow #287
Conversation
| @@ -0,0 +1,32 @@ | |||
| version: '3.9' | |||
There was a problem hiding this comment.
Please add these ROOT/docker/docker-compose.nlu.yml and ROOT/docker/docker-compose.nlu.dev.yml
nlu/docker-compose.yml
Outdated
| ports: | ||
| - "5002:5000" # Expose MLflow UI | ||
| volumes: | ||
| - ./mlruns:/mlruns # Mount local directory for MLflow artifacts |
There was a problem hiding this comment.
| - ./mlruns:/mlruns # Mount local directory for MLflow artifacts | |
| - mlruns:/mlruns # Mount local directory for MLflow artifacts |
Add "mlruns:" to the volumes
nlu/models/intent_classifier.py
Outdated
| @@ -34,6 +36,7 @@ | |||
| 'fr': "dbmdz/bert-base-french-europeana-cased", | |||
| } | |||
|
|
|||
| mlflow.set_tracking_uri("http://0.0.0.0:5002") | |||
There was a problem hiding this comment.
| mlflow.set_tracking_uri("http://0.0.0.0:5002") | |
| mlflow.set_tracking_uri("http://0.0.0.0:5002") |
This should be the docker hostname
| self.save() | ||
|
|
||
| # Start MLflow run | ||
| with mlflow.start_run() as run: |
There was a problem hiding this comment.
Shouldn't this be implement in the parent class so that it would work for all models ?
There was a problem hiding this comment.
Ideally it should be implemented as a helper class. Otherwise, we're logging metrics during training. The base class doesn't have a fit method.
| @@ -96,7 +96,7 @@ | |||
| if os.path.isfile(os.path.join(model.save_dir, "checkpoint")): | |||
| model.restore() | |||
| else: | |||
| model.save() | |||
| model.save_model() | |||
There was a problem hiding this comment.
Not sure why we need to rename this.
There was a problem hiding this comment.
To avoid confusion with keras' built-in save method. It's been throwing aberrant exceptions. Solved it by renaming the boilerplate method
Motivation
This PR modifies the training process of intent classifier models in order to be able to expose the Mlflow UI and log metrics/artifacts/registered models there.
Type of change:
Please delete options that are not relevant.
Checklist: