Awesome

<img src="img/logo.jpg" width="8%" alt="" align=center /> DiffuSeq

Official Codebase for __DiffuSeq__: Sequence to Sequence Text Generation With Diffusion Models and __DiffuSeq-v2__: Bridging Discrete and Continuous Text Spaces for Accelerated Seq2Seq Diffusion Models.

Highlights

• Our proposed DiffuSeq as a conditional language model is trained end-to-end in a classifier-free manner.
• We establish a theoretical connection among AR, NAR and DiffuSeq models (refer to our original paper).
• DiffuSeq is a powerful model for text generation, matching or even surpassing competitive AR, iterative NAR, and large-PLMs on quality and diversity.

Our study addresses promising achievements by such a new sequence-to-sequence learning paradigm.

Update: Our enhanced version effectively accelerates the training convergence by 4x and generates samples of similar quality 800x faster, rendering it significantly closer to practical application.

Setup:

The code is based on PyTorch and HuggingFace transformers.

pip install -r requirements.txt 

Datasets

Prepare datasets and put them under the datasets folder. Take datasets/CommonsenseConversation/train.jsonl as an example. We use four datasets in our paper.

DiffuSeq Training

cd scripts
bash train.sh

Arguments explanation:

• --dataset: the name of datasets, just for notation
• --data_dir: the path to the saved datasets folder, containing train.jsonl,test.jsonl,valid.jsonl
• --seq_len: the max length of sequence $z$ ($x\oplus y$)
• --resume_checkpoint: if not none, restore this checkpoint and continue training
• --vocab: the tokenizer is initialized using bert or load your own preprocessed vocab dictionary (e.g. using BPE)

It will take 2 more days to train a DiffuSeq model on 4 NVIDIA A100 80G GPUs for QG and QQP, and the training steps should be increased accordingly along with the size of the training set. To reproduce the results of Table 1 in our paper, we suggest the following configuration for each dataset when training.

Update:

Additional argument:

• --learned_mean_embed: set whether to use the learned soft absorbing state.
• --denoise: set whether to add discrete noise
• --use_fp16: set whether to use mixed precision training
• --denoise_rate: set the denoise rate, with 0.5 as the default

It only take around 11 hours to train a model on 2 NVIDIA A100 80G GPUs for QQP.

python -m torch.distributed.launch --nproc_per_node=4 --master_port=12233 --use_env run_train.py --diff_steps 2000 --lr 0.0001 --learning_steps 50000 --save_interval 10000 --seed 102 --noise_schedule sqrt --hidden_dim 128 --bsz 2048 --dataset qqp --data_dir {datasets/QQP} --vocab bert --seq_len 128 --schedule_sampler lossaware --notes qqp

python -m torch.distributed.launch --nproc_per_node=4 --master_port=12233 --use_env run_train.py --diff_steps 2000 --lr 0.0001 --learning_steps 40000 --save_interval 2000 --seed 102 --noise_schedule sqrt --hidden_dim 128 --bsz 2048 --microbatch 64 --dataset qg --data_dir {datasets/QG} --vocab bert --seq_len 128 --schedule_sampler lossaware --notes qg

python -m torch.distributed.launch --nproc_per_node=7 --master_port=12233 --use_env run_train.py --diff_steps 2000 --lr 0.0001 --learning_steps 140000 --save_interval 20000 --seed 102 --noise_schedule sqrt --hidden_dim 128 --bsz 2048 --microbatch 64 --dataset dialogue --data_dir {datasets/Conversation} --vocab bert --seq_len 128 --schedule_sampler lossaware --notes dialogue

python -m torch.distributed.launch --nproc_per_node=8 --master_port=12233 --use_env run_train.py --diff_steps 2000 --lr 0.0001 --learning_steps 80000 --save_interval 20000 --seed 102 --noise_schedule sqrt --hidden_dim 128 --bsz 2048 --microbatch 64 --dataset dialogue --data_dir {datasets/TS} --vocab bert --seq_len 128 --schedule_sampler lossaware --notes ts

Empirically, larger batchsize (larger microbatch here) can achieve higher BLEU score (without MBR). If you want to sync training loss to wandb, please customize your wandb setting in train.py (add your own API KEY).

DiffuSeq Decoding

You need to modify the path to model_dir, which is obtained in the training stage.

cd scripts
bash run_decode.sh

To reproduce the results of Table 1 in our paper, we suggest the size of MBR candidate set to be 10 (run 10 times using different seeds). Empirically, larger size can achieve higher BLEU score. For diversity metrics, the size of MBR candidate set is 3 when computing.

Speed-up Decoding

We customize the implementation of DPM-Solver++ to DiffuSeq to accelerate its sampling speed.

cd scripts
bash run_decode_solver.sh

Evaluation & MBR

You need to specify the folder of decoded texts. This folder should contain the decoded files from the same model but sampling with different random seeds. If mbr is not attached, we will compute the diversity score from the files in the folder, otherwise we will do MBR decoding:

cd scripts
python eval_seq2seq.py --folder ../{your-path-to-outputs} --mbr

Note: if you want to use this evaluation script for output files from other models, please make sure the same line from these output files refers to the same piece of data. Otherwise the diversity score could be incorrect.

Update

• Update 10 Oct 2023: We update the DiffuSeq-v2, targeting the training/sampling speed up. Details in new branch diffuseq-v2.
• Update 22 May 2023: We prepare the checkpoint and sampling results for remaining tasks in this link.
• Update 28 Nov 2022: We prepare the checkpoint and sampling results of 10 seeds for QQP dataset in this link.
• Update 14 Feb 2023: We update the evaluation scripts and camera ready version of the paper.

Welcome to discuss if you have any questions.

Citation

Please add the citation if our paper or code helps you.

@inproceedings{gong2022diffuseq,
  author = {Gong, Shansan and Li, Mukai and Feng, Jiangtao and Wu, Zhiyong and Kong, Lingpeng},
  booktitle = {International Conference on Learning Representations, ICLR},
  title = {{DiffuSeq}: Sequence to Sequence Text Generation with Diffusion Models},
  year = 2023
}

@article{gong2023diffuseqv2,
  title={DiffuSeq-v2: Bridging Discrete and Continuous Text Spaces for Accelerated Seq2Seq Diffusion Models},
  author={Gong, Shansan and Li, Mukai and Feng, Jiangtao and Wu, Zhiyong and Kong, Lingpeng},
  journal={arXiv preprint arXiv:2310.05793},
  year={2023}
}