Discourse Mutual Information (DMI)

This repository hosts the PyTorch-based implementation for the DMI model proposed in Representation Learning for Conversational Data using Discourse Mutual Information Maximization.

Requirements

wandb
transformers
datasets
torch 1.8.2 (lts)

Getting Access to the Source Code or Pretrained Models

To get access to the source-code or pretrained-model checkpoints, please send a request to AcadGrants@service.microsoft.com and cc to pawang.iitk [at] iitkgp.ac.in and bsantraigi [at] gmail.com.

Note: The requesting third party (1) can download and use these deliverables for research as well as commercial use, (2) modify it as they like but should include citation to our work and include this readme, and (3) cannot redistribute strictly to any other organization.

Cite As

@inproceedings{santra2022representation,
  title={Representation Learning for Conversational Data using Discourse Mutual Information Maximization},
  author={Santra, Bishal and Roychowdhury, Sumegh and Mandal, Aishik and Gurram, Vasu and Naik, Atharva and Gupta, Manish and Goyal, Pawan},
  booktitle={Proceedings of the 2022 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies},
  year={2022}
}

How to run?

Loading and Finetuning the model for a task

For finetuning the model on the tasks mentioned in the paper, or on a new task, use the run_finetune.py script or modify it according to your requirements. Example commands for launching finetuning based on some DMI checkpoints can be found in the auto_eval directory.

For example, if you have downloaded the checkpoint DMI_small_model.pth in checkpoints/ directory, you can launch one of these auto_eval scripts as:

MODEL_NAME_PATH="checkpoints/DMI_small_model.pth" bash long_eval/probe_part1_rob.sh

Special Note

For running experiments with the checkpoints of different sizes, use the corresponding code branches as directed below -

DMI_Base: master branch
- Finetuning scripts in auto_eval/
DMI_Medium: berty branch
- Finetuning scripts in auto_eval/8L/
DMI_Small: berty branch
- Finetuning scripts in auto_eval/

The main difference among these models is the DMI_Base model uses the Roberta-base architecture as the core for its encoder, whereas the other two uses bert-8L and bert-6L respectively.

Pretraining Dataset

There are two types of dataset structure that are available for model pretraining.

In case of smaller or "Normal" datasets, a single train_dialog file contains all the training data and is consumed fully during each epoch.

In case of "Large" datasets, the files are split into smaller shards and saved as .json files.

Normal Datasets: For example of this, check the data/dailydialog or data/reddit_1M directories.

data/reddit_1M
├── test_dialogues.txt
├── train_dialogues.txt
└── val_dialogues.txt

Large Datasets: This mode can be activated by setting the --dataset argument to rMax, i.e., --dataset rMax or -dd rMax. This also require you to provide the -rmp argument for the directory path containing the json files. For validation during pretraining, this model uses the DailyDialog validation set by default.

data/rMax-subset
├── test-00000-of-01000.json
├── test-00001-of-01000.json
├── test-00002-of-01000.json
├── test-00003-of-01000.json
├── ...
├── train-00000-of-01000.json
├── train-00001-of-01000.json
├── train-00002-of-01000.json
├── train-00003-of-01000.json
└── ...

For training a model

To train a new model, it can be started using the pretrain.py script.

Example:

For training from scratch:

python pretrain.py \
  -dd rMax -voc roberta \
  --roberta_init \
  -sym \
  -bs 64 -ep 1000 -vi 400 -li 50 -lr 5e-5 -scdl \
  --data_path ./data \
  -rmp /disk2/infonce-dialog/data/r727m/ \
  -t 1 \
  -ddp --world_size 6 \
  -ntq

To resume training from an existing checkpoint: This example shows resuming training from a checkpoint saved under checkpoints/DMI-Small_BERT-26Jan/. Also note how we specify a name an existing BERT/RoBERTa model which defines the architecture and the original initialization of the model weights.

python pretrain.py \
  -dd rMax -voc bert \
  --roberta_init \
  -robname google/bert_uncased_L-8_H-768_A-12 \
  -sym -bs 130 -lr 1e-5 -scdl -ep 1000 -vi 400 -li 50 \
  --data_path ./data \
  -rmp /disk2/infonce-dialog/data/r727m/ \
  -ddp --world_size 4 \
  -ntq -t 1 \
  -re -rept checkpoints/DMI-Small_BERT-26Jan/model_current.pth

It accepts the following arguments.

  -h, --help            show this help message and exit
  -dd {dd,r5k,r100k,r1M,r1M/cc,rMax,rMax++,paa,WoW}, --dataset {dd,r5k,r100k,r1M,r1M/cc,rMax,rMax++,paa,WoW}
                        which dataset to use for pretraining.
  -rf, --reddit_filter_enabled
                        Enable reddit data filter for removing low quality dialogs.
  -rmp RMAX_PATH, --rmax_path RMAX_PATH
                        path to dir for r727m (.json) data files.
  -dp DATA_PATH, --data_path DATA_PATH
                        path to the root data folder.
  -op OUTPUT_PATH, --output_path OUTPUT_PATH
                        Path to store the output ``model.pth'' files
  -voc {bert,blender,roberta,dgpt-m}, --vocab {bert,blender,roberta,dgpt-m}
                        mention which tokenizer was used for pretraining? bert or blender
  -rob, --roberta_init  Initialize transformer-encoder with roberta weights?
  -robname ROBERTA_NAME, --roberta_name ROBERTA_NAME
                        name of checkpoint from huggingface
  -d D_MODEL, --d_model D_MODEL
                        size of transformer encoders' hidden representation
  -d_ff DIM_FEEDFORWARD, --dim_feedforward DIM_FEEDFORWARD
                        dim_feedforward for transformer encoder.
  -p PROJECTION, --projection PROJECTION
                        size of projection layer output
  -el ENCODER_LAYERS, --encoder_layers ENCODER_LAYERS
                        number of layers in transformer encoder
  -eh ENCODER_HEADS, --encoder_heads ENCODER_HEADS
                        number of heads in tformer enc
  -sym, --symmetric_loss
                        whether to train using symmetric infonce
  -udrl, --unsupervised_discourse_losses
                        Additional unsupervised discourse-relation loss components
  -sdrl, --supervised_discourse_losses
                        Additional supervised discourse-relation loss components
  -es {infonce,jsd,nwj,tuba,dv,smile,infonce/td}, --estimator {infonce,jsd,nwj,tuba,dv,smile,infonce/td}
                        which MI estimator is used as the loss function.
  -bs BATCH_SIZE, --batch_size BATCH_SIZE
                        batch size during pretraining
  -ep EPOCHS, --epochs EPOCHS
                        epochs for pretraining
  -vi VAL_INTERVAL, --val_interval VAL_INTERVAL
                        validation interval during training
  -li LOG_INTERVAL, --log_interval LOG_INTERVAL
                        logging interval during training
  -lr LEARNING_RATE, --learning_rate LEARNING_RATE
                        set learning rate
  -lrc, --learning_rate_control
                        LRC: outer layer and projection layer will have faster LR and rest will be LR/10
  -t {0,1}, --tracking {0,1}
                        whether to track training+validation loss wandb
  -scdl, --use_scheduler
                        whether to use a warmup+decay schedule for LR
  -ntq, --no_tqdm       disable tqdm to create concise log files!
  -ddp, --distdp        Should it use pytorch Distributed dataparallel?
  -ws WORLD_SIZE, --world_size WORLD_SIZE
                        world size when using DDP with pytorch.
  -re, --resume         2-stage pretrain: Resume training from a previous checkpoint?
  -rept RESUME_MODEL_PATH, --resume_model_path RESUME_MODEL_PATH
                        If ``Resuming'', path to ckpt file.