A system for quickly generating training data with multi-task weak supervision

# Snorkel MeTaL

v0.5.0

ANNOUNCEMENT (3/20): We are excited to have achieved a new state-of-the-art score on the GLUE Benchmark and four of its component tasks using Snorkel MeTaL. Check out the corresponding blog post for an overview of how we did it. The code we used to accomplish this was part of a significant restructuring of multi-task end models in Snorkel MeTaL to make it as easy as possible to perform Massive Multi-Task Learning (MMTL) with supervision at varying levels of granularity and over an arbitrarily large number of tasks. That mmtl package has been released as a part of Snorkel MeTaL v0.5, along with a basic tutorial. Additional tutorials showing more advanced usage (e.g., using a pre-trained BERT network as a shared input module, using multiple label sets, supervising at the token and sentence level simultaneously, etc.) will be released in future minor version updates, though such functionality is already supported.

Stay tuned on other developments in the Snorkel ecosystem at our project landing page: snorkel.stanford.edu.

## Motivation

More concretely, Snorkel MeTaL is a framework for using multi-task weak supervision (MTS), provided by users in the form of labeling functions applied over unlabeled data, to train multi-task models. Snorkel MeTaL can use the output of labeling functions developed and executed in Snorkel, or take in arbitrary label matrices representing weak supervision from multiple sources of unknown quality, and then use this to train auto-compiled MTL networks.

Snorkel MeTaL uses a new matrix approximation approach to learn the accuracies of diverse sources with unknown accuracies, arbitrary dependency structures, and structured multi-task outputs. This makes it significantly more scalable than our previous approaches.

## Q&A

If you are looking for help regarding how to use a particular class or method, the best references are (in order):

• The docstrings for that class
• The MeTaL Commandments
• The corresponding unit tests in tests/
• The Issues page (We tag issues that might be particularly helpful with the "reference question" label)

## Sample Usage

"""
n = # data points
m = # labeling functions
k = cardinality of the classification task

L: an [n,m] scipy.sparse label matrix of noisy labels
Y: an n-dim numpy.ndarray of target labels
X: an n-dim iterable (e.g., a list) of end model inputs
"""

from metal.label_model import LabelModel, EndModel

# Train a label model and generate training labels
label_model = LabelModel(k)
label_model.train_model(L_train)
Y_train_probs = label_model.predict_proba(L_train)

# Train a discriminative end model with the generated labels
end_model = EndModel([1000,10,2])
end_model.train_model(train_data=(X_train, Y_train_probs), valid_data=(X_dev, Y_dev))

# Evaluate performance
score = end_model.score(data=(X_test, Y_test), metric="accuracy")


Note for Snorkel users: Snorkel MeTaL, even in the single-task case, learns a slightly different label model than Snorkel does (e.g. here we learn class-conditional accuracies for each LF, etc.)---so expect slightly different (hopefully better!) results.

## Release Notes

### Major changes in v0.5:

• Introduction of Massive Multi-Task Learning (MMTL) package in metal/mmtl/ with tutorial.
• Additional logging improvements from v0.4

### Major changes in v0.4:

• Improved control over logging/checkpointing/validation
• More modular code, separate Logger, Checkpointer, LogWriter classes
• Support for user-defined metrics for validation/checkpointing
• Logging frequency can now be based on seconds, examples, batches, or epochs
• Naming convention change: hard (int) labels -> preds, soft (float) labels -> probs

## Setup

[1] Install anaconda:

[2] Clone the repository:

git clone https://github.com/HazyResearch/metal.git
cd metal


[3] Create virtual environment:

conda env create -f environment.yml
source activate metal


[4] Run unit tests:

nosetests


If the tests run successfully, you should see 50+ dots followed by "OK".
Check out the tutorials to get familiar with the Snorkel MeTaL codebase!

Or, to use Snorkel Metal in another project, install it with pip:

pip install snorkel-metal


## Developer Guidelines

First, read the MeTaL Commandments, which describe the major design principles, terminology, and style guidelines for Snorkel MeTaL.

If you are interested in contributing to Snorkel MeTaL (and we welcome whole-heartedly contributions via pull requests!), follow the setup guidelines above, then run the following additional command:

make dev


This will install a few additional tools that help to ensure that any commits or pull requests you submit conform with our established standards. We use the following packages:

• isort: import standardization
• black: automatic code formatting
• flake8: PEP8 linting

After running make dev to install the necessary tools, you can run make check to see if any changes you've made violate the repo standards and make fix to fix any related to isort/black. Fixes for flake8 violations will need to be made manually.

### GPU Usage

MeTaL supports GPU usage, but does not include this in automatically-run tests; to run these tests, first install the requirements in tests/gpu/requirements.txt, then run:

nosetests tests/gpu


## Project details

Uploaded source
Uploaded py3