DEMix

This repository contains modeling utilities for "DEMix Layers: Disentangling Domains for Modular Language Modeling" (Gururangan et. al, 2021).

This code is a fork of Fairseq. It is based on Python 3.8, CUDA 11 and includes PyTorch 1.8.0, NCCL 2.8.4 and apex.

Dataset

The multidomain dataset scripts are housed in another repository, located here. Clone that repository and follow instructions to setup data to train on.

Follow that tutorial to generate data-bins on eight (small) example domains.

Make sure to set the DATA_DIR accordingly.

Fairseq Installation

If you've already made an environment from the dataset creation phase, just use that. Otherwise:

conda create env --name demix
cd demix/
pip install --editable .

Additionally, please make sure you have the dependencies above installed (check Fairseq documentation for more information).

Tutorial

Here we will follow a tutorial to train on the example domains from the tutorial in the DEMix-data repository. Note that the model that results from this tutorial is pretty bad, because we're working with very small amounts of data and also a small LM. This tutorial is there to help you quickly understand the pipeline, and ensure that each script completes successfully.

To replicate the DEMix paper, with a GPT-3 model, follow the instructions here.

Basic Training

After setting up the example domains, run the following to train a small language model. Note that the scripts in this paper assume you are running on a multi-node GPU cluster with SLURM.

First, allocate some nodes, with GPUs with at least 32GB of RAM. Here we allocate 1 node with 8 volta32GB GPUs.

salloc --gpus-per-node 8 --nodes 1  -C 'volta32gb' --ntasks-per-node 8 --cpus-per-task 10 --mem 400G --time XXX --partition YYY

Then run:

export NUM_GPUS=8
export DISTRIBUTED_PORT=12345
export MODEL=transformer_lm
export EXPERIMENT=demix
# $DATA_DIR was set in DEMix-data tutorial.
export DATA_BIN=${DATA_DIR}/data-bin/
export EXPERIMENT_SUFFIX=tutorial
export SERIALIZATION_DIR=$(pwd)/demix_tutorial_model
bash tutorial/train.sh $NUM_GPUS \
                    $DISTRIBUTED_PORT \
                    $MODEL \
                    $EXPERIMENT \
                    $DATA_BIN \
                    $SERIALIZATION_DIR \
                    $EXPERIMENT_SUFFIX

This will output a trained language model in ${SERIALIZATION_DIR}

To train balanced dense LM, set export EXPERIMENT=dense, to train unbalanced dense LM, set export EXPERIMENT=unbalanced, to train "+Domain Token" LM , set export EXPERIMENT=domain_token.

We have provided a simple script demix/train.sh, with the same interface, with all hyperparameter preset to help replicate results in the paper.

Evaluation

We have two ways to evaluate the demix language model: with and without mixing experts.

Evaluating without mixing experts

To evaluate the language model without mixing experts, you can supply the checkpoint from a GPU on a particular rank (to specify the use of the domain expert that was trained on that GPU):

export DATA_BIN=${DATA_DIR}/data-bin/
export GPU_RANK=0
export PATH_TO_CHECKPOINT=${SERIALIZATION_DIR}/checkpoint_last-rank-${GPU_RANK}.pt
export OUTPUT_PATH=eval_output.jsonl
export SPLIT=valid
export DOMAIN=imdb
bash tutorial/eval_lm.sh $DATA_BIN $PATH_TO_CHECKPOINT $OUTPUT_PATH $SPLIT $DOMAIN

To evaluate on test data, set export SPLIT=test

The same script is used for the other baselines.

For the +domain token model, you can additionally supply a domain token to use at test time:

export DOMAIN_TOKEN=XXX
bash tutorial/eval_lm.sh $DATA_BIN $PATH_TO_CHECKPOINT $OUTPUT_PATH $SPLIT $DOMAIN $DOMAIN_TOKEN

Evaluating with mixing experts

First, we estimate the posterior distribution on 100 sequences of validation data of the domain using the following command:

export DATA_BIN=${DATA_DIR}/data-bin
export DOMAIN=imdb
export DEV_POSTERIOR_OUTPUT=dev_posteriors.jsonl
# set NUM_EVALUATION_GPUS equal to the number of experts you'd like to ensemble.
export NUM_EVALUATION_GPUS=8;
bash tutorial/mix_eval_lm.sh $NUM_EVALUATION_GPUS $DATA_BIN  ${SERIALIZATION_DIR}/checkpoint_last-rank-0.pt:${SERIALIZATION_DIR}/checkpoint_last-rank-1.pt:${SERIALIZATION_DIR}/checkpoint_last-rank-2.pt:${SERIALIZATION_DIR}/checkpoint_last-rank-3.pt:${SERIALIZATION_DIR}/checkpoint_last-rank-4.pt:${SERIALIZATION_DIR}/checkpoint_last-rank-5.pt:${SERIALIZATION_DIR}/checkpoint_last-rank-6.pt:${SERIALIZATION_DIR}/checkpoint_last-rank-7.pt $DOMAIN $DEV_POSTERIOR_OUTPUT estimate;

Then, we open $POSTERIOR_OUTPUT, extracting the exp_avg_posterior value of the last line in that file:

export POSTERIOR=$(tail -n 1 $DEV_POSTERIOR_OUTPUT | jq -rc '.exp_avg_posterior | join(",")')

We use this posterior as the domain prior (supplied as a string) when evaluating on test data, like so:

bash tutorial/mix_eval_lm.sh $NUM_EVALUATION_GPUS $DATA_BIN  ${SERIALIZATION_DIR}/checkpoint_last-rank-0.pt:${SERIALIZATION_DIR}/checkpoint_last-rank-1.pt:${SERIALIZATION_DIR}/checkpoint_last-rank-2.pt:${SERIALIZATION_DIR}/checkpoint_last-rank-3.pt:${SERIALIZATION_DIR}/checkpoint_last-rank-4.pt:${SERIALIZATION_DIR}/checkpoint_last-rank-5.pt:${SERIALIZATION_DIR}/checkpoint_last-rank-6.pt:${SERIALIZATION_DIR}/checkpoint_last-rank-7.pt $DOMAIN $DEV_POSTERIOR_OUTPUT eval $POSTERIOR cached_prior;

Adapting the Language Model

We additionally provide scripts to adapt the language model to a new domain.

DEMix DAPT

In this tutorial, we just adapt one of the existing experts to a new example domain in the demix-data project, located in /path/to/demix-data/new_example_domains.

First, we need to figure out which domain expert has the most affinity to the target domain we want to adapt to:

export NEW_DATA_BIN=/private/home/suching/demix-data/new_example_domains/data-bin/
export NEW_DOMAIN=acl_papers
export DEV_POSTERIOR_OUTPUT=${NEW_DOMAIN}_posterior.jsonl
# set NUM_EVALUATION_GPUS equal to the number of experts you'd like to ensemble.
export NUM_EVALUATION_GPUS=8;
bash tutorial/mix_eval_lm.sh $NUM_EVALUATION_GPUS $NEW_DATA_BIN  ${SERIALIZATION_DIR}/checkpoint_last-rank-0.pt:${SERIALIZATION_DIR}/checkpoint_last-rank-1.pt:${SERIALIZATION_DIR}/checkpoint_last-rank-2.pt:${SERIALIZATION_DIR}/checkpoint_last-rank-3.pt:${SERIALIZATION_DIR}/checkpoint_last-rank-4.pt:${SERIALIZATION_DIR}/checkpoint_last-rank-5.pt:${SERIALIZATION_DIR}/checkpoint_last-rank-6.pt:${SERIALIZATION_DIR}/checkpoint_last-rank-7.pt $NEW_DOMAIN $DEV_POSTERIOR_OUTPUT estimate;
export POSTERIOR=$(tail -n 1 $DEV_POSTERIOR_OUTPUT | jq -rc '.exp_avg_posterior | join(",")')

Here, we find that the most likely expert is expert number 5.

export POSTERIOR=$(tail -n 1 $DEV_POSTERIOR_OUTPUT | jq -rc '.exp_avg_posterior | join(",")')
echo $POSTERIOR

We then adapt expert 5 to the target domain using the tutorial/dapt.sh script, using DEMix DAPT:

export PATH_TO_CHECKPOINT=${SERIALIZATION_DIR}/checkpoint_last-rank-5.pt
export UNFREEZE_PARAMETERS=feedforward
export NEW_SERIALIZATION_DIR=$(pwd)/${NEW_DOMAIN}_demix_dapt
export EXPERIMENT_SUFFIX=test
bash tutorial/dapt.sh $NEW_DATA_BIN $NEW_DOMAIN $PATH_TO_CHECKPOINT $UNFREEZE_PARAMETERS $NEW_SERIALIZATION_DIR $EXPERIMENT_SUFFIX

Once this is trained, you can add that expert to your ensemble when evaluating on new data:

export NEW_DATA_BIN=/path/to/demix-data/new_example_domains/data-bin/
export NEW_DOMAIN=acl_papers
export DEV_POSTERIOR_OUTPUT=${NEW_DOMAIN}_posterior.jsonl
# set NUM_EVALUATION_GPUS equal to the number of experts you'd like to ensemble.
export NUM_EVALUATION_GPUS=8;
export PATH_TO_NEW_EXPERT=${NEW_SERIALIZATION_DIR}/checkpoint_last-rank-0.pt
bash tutorial/mix_eval_lm.sh $NUM_EVALUATION_GPUS $NEW_DATA_BIN  ${SERIALIZATION_DIR}/checkpoint_last-rank-0.pt:${SERIALIZATION_DIR}/checkpoint_last-rank-1.pt:${SERIALIZATION_DIR}/checkpoint_last-rank-2.pt:${SERIALIZATION_DIR}/checkpoint_last-rank-3.pt:${SERIALIZATION_DIR}/checkpoint_last-rank-4.pt:${SERIALIZATION_DIR}/checkpoint_last-rank-5.pt:${SERIALIZATION_DIR}/checkpoint_last-rank-6.pt:${PATH_TO_NEW_EXPERT} $NEW_DOMAIN $DEV_POSTERIOR_OUTPUT estimate;
export POSTERIOR=$(tail -n 1 $DEV_POSTERIOR_OUTPUT | jq -rc '.exp_avg_posterior | join(",")')

Dense DAPT

If you wanted to do Dense DAPT instead, just change the environment variables:

export PATH_TO_CHECKPOINT=/path/to/dense/model/checkpoint_last.pt
export FEEDFORWARD_OR_FULL=full
export SERIALIZATION_DIR=$(pwd)/${NEW_DOMAIN}_dense_dapt
export EXPERIMENT_SUFFIX=test
bash tutorial/dapt.sh $NEW_DATA_BIN $NEW_DOMAIN $PATH_TO_CHECKPOINT $FEEDFORWARD_OR_FULL $SERIALIZATION_DIR $EXPERIMENT_SUFFIX

DEMix Layers for Modular Language Modeling

Related tags

Overview

DEMix

Dataset

Fairseq Installation

Tutorial

Basic Training

Evaluation

Evaluating without mixing experts

Evaluating with mixing experts

Adapting the Language Model

DEMix DAPT

Dense DAPT

Owner

Suchin

Code for NeurIPS 2020 article "Contrastive learning of global and local features for medical image segmentation with limited annotations"

An implementation of DeepMind's Relational Recurrent Neural Networks in PyTorch.

A minimalist environment for decision-making in autonomous driving

Lolviz - A simple Python data-structure visualization tool for lists of lists, lists, dictionaries; primarily for use in Jupyter notebooks / presentations

Temporal Segment Networks (TSN) in PyTorch

Official Implementation and Dataset of "PPR10K: A Large-Scale Portrait Photo Retouching Dataset with Human-Region Mask and Group-Level Consistency", CVPR 2021

DexterRedTool - Dexter's Red Team Tool that creates cronjob/task scheduler to consistently creates users

Source code for Fixed-Point GAN for Cloud Detection

A colab notebook for training Stylegan2-ada on colab, transfer learning onto your own dataset.

GuideDog is an AI/ML-based mobile app designed to assist the lives of the visually impaired, 100% voice-controlled

This repository accompanies the ACM TOIS paper "What can I cook with these ingredients?" - Understanding cooking-related information needs in conversational search

Official Pytorch implementation of "Beyond Static Features for Temporally Consistent 3D Human Pose and Shape from a Video", CVPR 2021

Fast and Simple Neural Vocoder, the Multiband RNNMS

Code for the Interspeech 2021 paper "AST: Audio Spectrogram Transformer".

MiniHack the Planet: A Sandbox for Open-Ended Reinforcement Learning Research

Source codes for "Structure-Aware Abstractive Conversation Summarization via Discourse and Action Graphs"

CLIP (Contrastive Language–Image Pre-training) for Italian

Sharpened cosine similarity torch - A Sharpened Cosine Similarity layer for PyTorch

High-fidelity 3D Model Compression based on Key Spheres

Official code for 'Pixel-wise Energy-biased Abstention Learning for Anomaly Segmentationon Complex Urban Driving Scenes'