https://github.com/chiphuyen/metaflow-transformers-tutorials

Metaflow tutorials for ODSC West 2021
https://github.com/chiphuyen/metaflow-transformers-tutorials

machine-learning metaflow

Last synced: 12 months ago
JSON representation

Metaflow tutorials for ODSC West 2021

• Host: GitHub
• URL: https://github.com/chiphuyen/metaflow-transformers-tutorials
• Owner: chiphuyen
• Created: 2021-11-11T05:21:54.000Z (over 4 years ago)
• Default Branch: main
• Last Pushed: 2021-11-16T19:08:26.000Z (over 4 years ago)
• Last Synced: 2025-03-28T04:41:39.522Z (about 1 year ago)
• Topics: machine-learning, metaflow
• Language: Jupyter Notebook
• Homepage:
• Size: 24.8 MB
• Stars: 64
• Watchers: 4
• Forks: 6
• Open Issues: 0
• Metadata Files:
  • Readme: README.md

Awesome Lists containing this project

• awesome-metaflow - metaflow-transformers-tutorials - HuggingFace DistilBERT fine-tuning tutorials by Chip Huyen. (Examples & Tutorials)

README

          

First, get started with Metaflow by executing these simple flows:

1. `helloworld.py` - a simple hello world flow

2. `counter_branch.py` - test artifacts

3. `parameters.py` - test parameters

4. `foreach.py` - test foreaches (parallel tasks)

After these simple examples, you can take a look at a more realistic case:

In this tutorial, we'll fine-tune a sentiment analysis model on top of

HuggingFace's DistilBERT model with the IMDB dataset.

First, we'll show how to do it without Metaflow.

1. sent_analysis_train.py is the training code (6-7 minutes on the small dataset of 100 samples on my Mac)

2. sent_analysis_predict.py is the prediction code (30 seconds)

We'll do live coding to show how to convert the training code to Metaflow.

See sent_analysis_metaflow.py for instructions.

We'll run `python sent_analysis_metaflow.py --no-pylint run --mode small`

to train a model on 100 samples locally.

We'll show how Metaflow automatically saves trained models which we can access for predictions.

We'll use @batch to train the full dataset (40,000 samples) on AWS.

We'll need GPU since it'll take a while for the full data on CPU.