Repository files navigation

This repository contains source code necessary to reproduce the results presented in the EMNLP 2020 paperOptimus: Organizing Sentences via Pre-trained Modeling of a Latent Space.

The network architecture of Optimus: encoder for representation learning and decoder for generation	Sentences are organized and manipulated in a pre-trained compact and smooth latent space

For more on this project, see theMicrosoft Research Blog post.

May 21, 2020: Releasing ademo for latent space manipulation, including sentence interpolation and analogy. Check out thewebsite.

May 20, 2020: The latent space manipulation code is cleaned and released. See instructions atoptimius_for_snli.md.

May 13, 2020: The fine-tuning code for langauge modeling is released. See instructions atoptimus_finetune_language_models.md

There are four steps to use this codebase to reproduce the results in the paper.

1. Dependencies
2. Prepare datasets
3. Model training
   1. Pre-training on setences in Wikipedia
   2. Languange Modeling
   3. Guided Language Generation
   4. Low-resource Language Understanding
4. Collect and plot results

Pull docker from Docker Hub at:chunyl/pytorch-transformers:v2. Please see the instruction atdoc/env.md

The project is organized into the following structures, with ensential files & folders visualized.output saves the models checkpoints.

├── Optimus   └── code       ├── examples           ├── big_ae               ├── modules                   ├── vae.py                   └── ...               ├── run_lm_vae_pretraining_phdist_beta.py               ├── run_lm_vae_training.py               └── ...   ├── pytorch_transformers               ├── modeling_bert.py               ├── modeling_gpt2.py               └── ...       ├── scripts           ├── scripts_docker   ├── scripts_local   ├── scripts_philly   └── data       └── datasets           ├── wikipedia_json_64_filtered               └── ...   ├── snli_data           └── ...   └── output       ├── pretrain       ├── LM       └── ...

Please download or preparation the data via following the instructions atdata/download_datasets.md.

1. Pre-training on setences in Wikipedia

We pre-trained our models on Philly (a Microsoft internal compute cluster), the code is specialized for multi-node multi-GPU compute on this platform. The pre-training main python isrun_lm_vae_pretraining_phdist_beta.py. You may need to adjust the distributed training scripts.

2. Languange Modeling

To have a fair comparison with existing VAE languange models, we consider a model with latent dimension 32. The pre-trained model is fine-tuned on four commonly datasets for one epoch. Please see the details atdoc/optimus_finetune_language_models.md

3. Guided Language Generation

Latent Space Manipulation To ensure good performance, we consider a model with latent dimension 768. The pre-trained model is fine-tuned on SNLI dataset, where sentences show related patterns. Please see the details atPlease see the details atdoc/optimius_for_snli.md

4. Low-resource Language Understanding

Once the networks are trained and the results are saved, we extracted key results using Python script. The results can be plotted using the included IPython notebookplots/main_plots.ipynb.Start the IPython Notebook server:

$ cd plots$ ipython notebook

Select themain_plots.ipynb notebook and execute the includedcode. Note that without modification, we have copyed our extracted results into the notebook, and script will output figures in the paper. If you've run your own training and wish to plot results, you'll have to organize your results in the same format instead.

Please drop me (Chunyuan) a line if you have any questions.

@inproceedings{li2020_Optimus,  title={Optimus: Organizing Sentences via Pre-trained Modeling of a Latent Space},  author={Li, Chunyuan and Gao, Xiang and Li, Yuan and Li, Xiujun and Peng, Baolin and Zhang, Yizhe and Gao, Jianfeng},  booktitle={EMNLP},  year={2020}}