Skip to content
Variational Methods for Pretraining in Resource-limited Environments
Python Jsonnet Jupyter Notebook Other
Branch: master
Clone or download
kernelmachine Merge pull request #38 from allenai/updates
Miscellaneous updates and enhancements
Latest commit e3795dd Jul 31, 2019

README.md

VAMPIRE

VAriational Methods for Pretraining In Resource-limited Environments

Read paper here.

Citation

@inproceedings{vampire,
 author = {Suchin Gururangan and Tam Dang and Dallas Card and Noah A. Smith},
 title = {Variational Pretraining for Semi-supervised Text Classification},
 year = {2019},
 booktitle = {Proceedings of ACL},
}

Installation

Install necessary dependencies via requirements.txt, which will include the latest unreleased install of allennlp (from the master branch).

pip install -r requirements.txt

Install the spacy english model with:

python -m spacy download en

Verify your installation by running:

SEED=42 pytest -v --color=yes vampire

All tests should pass.

Install from Docker

Alternatively, you can install the repository with Docker.

First, build the container:

docker build -f Dockerfile --tag vampire/vampire:latest .

Then, run the container:

docker run -it vampire/vampire:latest

This will open a shell in a docker container that has all the dependencies installed.

Download Data

Download your dataset of interest, and make sure it is made up of json files, where each line of each file corresponds to a separate instance. Each line must contain a text field, and optionally a label field.

In this tutorial we use the AG News dataset hosted on AllenNLP. Download it using the following script:

sh scripts/download_ag.sh

This will make an examples/ag directory with train, dev, test files from the AG News corpus.

Preprocess data

To make pretraining fast, we precompute fixed bag-of-words representations of the data.

python -m scripts.preprocess_data \
            --train-path examples/ag/train.jsonl \
            --dev-path examples/ag/dev.jsonl \
            --tokenize \
            --tokenizer-type spacy \
            --vocab-size 30000 \
            --serialization-dir examples/ag

This script will tokenize your data, and save the resulting output into the specified serialization-dir.

Alternatively, under `https://s3-us-west-2.amazonaws.com/allennlp/datasets/ag-news/preprocessed.tar", we have a tar file containing a pre-processed AG news data (with vocab size set to 30K).

Run

curl -Lo examples/ag/ag.tar https://s3-us-west-2.amazonaws.com/allennlp/datasets/ag-news/vampire_preprocessed_example.tar
tar -xvf examples/ag/ag.tar -C examples/

to access its contents.

In examples/ag (after running the preprocess_data module or unpacking ag.tar), you should see:

  • train.npz - pre-computed bag of word representations of the training data
  • dev.npz - pre-computed bag of word representations of the dev data
  • vampire.bgfreq - background word frequencies
  • vocabulary/ - AllenNLP vocabulary directory

This script also creates a reference corpus to calcuate NPMI (normalized pointwise mutual information), a measure of topical coherence that we use for early stopping. By default, we use the validation data as our reference corpus. You can supply a --reference-corpus-path to the preprocessing script to use your own reference corpus.

In examples/ag/reference, you should see:

  • ref.npz - pre-computed bag of word representations of the reference corpus (the dev data)
  • ref.vocab.json - the reference corpus vocabulary

Pretrain VAMPIRE

Set your data directory and vocabulary size as environment variables:

export DATA_DIR="$(pwd)/examples/ag"
export VOCAB_SIZE=30000

If you're training on a dataset that's to large to fit into RAM, run VAMPIRE in lazy mode by additionally exporting:

export LAZY=1

Then train VAMPIRE:

python -m scripts.train \
            --config training_config/vampire.jsonnet \
            --serialization-dir model_logs/vampire \
            --environment VAMPIRE \
            --device -1

This model can be run on a CPU (--device -1). To run on a GPU instead, run with --device 0 (or any other available CUDA device number).

This command will output training logs at model_logs/vampire.

For convenience, we include the --override flag to remove the previous experiment at the same serialization directory.

Inspect topics learned

During training, we output the learned topics after each epoch in the serialization directory, under model_logs/vampire.

After your model is finished training, check out the best_epoch field in model_logs/vampire/metrics.json, which corresponds to the training epoch at which NPMI is highest.

Then open up the corresponding epoch's file in model_logs/vampire/topics/.

Use VAMPIRE with a downstream classifier

Using VAMPIRE with a downstream classifier is essentially the same as using regular ELMo. See this documentation for details on how to do that.

This library has some convenience functions for including VAMPIRE with a downstream classifier.

First, set some environment variables:

  • VAMPIRE_DIR: path to newly trained VAMPIRE
  • VAMPIRE_DIM: dimensionality of the newly trained VAMPIRE (the token embedder needs it explicitly)
  • THROTTLE: the sample size of the data we want to train on.
  • EVALUATE_ON_TEST: whether or not you would like to evaluate on test
export VAMPIRE_DIR="$(pwd)/model_logs/vampire"
export VAMPIRE_DIM=81
export THROTTLE=200
export EVALUATE_ON_TEST=0

Then, you can run the classifier:

python -m scripts.train \
            --config training_config/classifier.jsonnet \
            --serialization-dir model_logs/clf \
            --environment CLASSIFIER \
            --device -1

As with VAMPIRE, this model can be run on a CPU (--device -1). To run on a GPU instead, run with --device 0 (or any other available CUDA device number)

This command will output training logs at model_logs/clf.

The dataset sample (specified by THROTTLE) is governed by the global seed supplied to the trainer; the same seed will result in the same subsampling of training data. You can set an explicit seed by passing the additional flag --seed to the train module.

With 200 examples, we report a test accuracy of 83.9 +- 0.9 over 5 random seeds on the AG dataset. Note that your results may vary beyond these bounds under the low-resource setting.

You can’t perform that action at this time.