Styled Augmented Translation

Last update: Dec 29, 2022

Related tags

Overview

SAT

Style Augmented Translation

Introduction

By collecting high-quality data, we were able to train a model that outperforms Google Translate on 6 different domains of English-Vietnamese Translation.

English to Vietnamese Translation (BLEU score)

Vietnamese to English Translation (BLEU score)

Get data and model at Google Cloud Storage

Check out our demo web app

Visit our blog post for more details.

Using the code

This code is build on top of vietai/dab:

To prepare for training, generate tfrecords from raw text files:

python t2t_datagen.py \
--data_dir=$path_to_folder_contains_vocab_file \
--tmp_dir=$path_to_folder_that_contains_training_data \
--problem=$problem

To train a Transformer model on the generated tfrecords

python t2t_trainer.py \
--data_dir=$path_to_folder_contains_vocab_file_and_tf_records \
--problem=$problem \
--hparams_set=$hparams_set \
--model=transformer \
--output_dir=$path_to_folder_to_save_checkpoints

To run inference on the trained model:

python t2t_decoder.py \
--data_dir=$path_to_folde_contains_vocab_file_and_tf_records \
--problem=$problem \
--hparams_set=$hparams_set \
--model=transformer \
--output_dir=$path_to_folder_contains_checkpoints

In this colab, we demonstrated how to run these three phases in the context of hosting data/model on Google Cloud Storage.

Dataset

Our data contains roughly 3.3 million pairs of texts. After augmentation, the data is of size 26.7 million pairs of texts. A more detail breakdown of our data is shown in the table below.

	Pure	Augmented
Fictional Books	333,189	2,516,787
Legal Document	1,150,266	3,450,801
Medical Publication	5,861	27,588
Movies Subtitles	250,000	3,698,046
Software	79,912	239,745
TED Talk	352,652	4,983,294
Wikipedia	645,326	1,935,981
News	18,449	139,341
Religious texts	124,389	1,182,726
Educational content	397,008	8,475,342
No tag	5,517	66,299
Total	3,362,569	26,715,950

Data sources is described in more details here.

Comments

Data leakage issue in evaluation?

Hi team @lmthang @thtrieu @heraclex12 @hqphat @KienHuynh

The obtained results of a Transformer-based model on the PhoMT test set surprised me. My first thought was that as VietAI and PhoMT datasets have several overlapping domains (e.g. Wikihow, TED talks, Opensubtitles, news..): whether there might be a potential data leakage issue in your evaluation (e.g. PhoMT English-Vietnamese test pairs appear in the VietAI training set)?

In particular, we find that 6294/19151 PhoMT English-Vietnamese test pairs appear in the VietAI training set (v2). When evaluating your model on the PhoMT test set, did you guys retrain the model on a VietAI training set variant that does not contain PhoMT English-Vietnamese test pairs?

Cheers, Dat.

opened by datquocnguyen 3

Demo website is not working

Hi, seems like the easiest to reach out here but https://demo.vietai.org/ is down, looks like the page tried to serve a 404 error page.

Connection failed with status 404, and response "<!DOCTYPE html> <html lang=en> <meta charset=utf-8> <meta name=viewport content="initial-scale=1, minimum-scale=1, width=device-width"> <title>Error 404 (Not Found)!!1</title> <style> *{margin:0;padding:0}html,code{font:15px/22px arial,sans-serif}html{background:#fff;color:#222;padding:15px}body{margin:7% auto 0;max-width:390px;min-height:180px;padding:30px 0 15px}* > body{background:url(//www.google.com/images/errors/robot.png) 100% 5px no-repeat;padding-right:205px}p{margin:11px 0 22px;overflow:hidden}ins{color:#777;text-decoration:none}a img{border:0}@media screen and (max-width:772px){body{background:none;margin-top:0;max-width:none;padding-right:0}}#logo{background:url(//www.google.com/images/branding/googlelogo/1x/googlelogo_color_150x54dp.png) no-repeat;margin-left:-5px}@media only screen and (min-resolution:192dpi){#logo{background:url(//www.google.com/images/branding/googlelogo/2x/googlelogo_color_150x54dp.png) no-repeat 0% 0%/100% 100%;-moz-border-image:url(//www.google.com/images/branding/googlelogo/2x/googlelogo_color_150x54dp.png) 0}}@media only screen and (-webkit-min-device-pixel-ratio:2){#logo{background:url(//www.google.com/images/branding/googlelogo/2x/googlelogo_color_150x54dp.png) no-repeat;-webkit-background-size:100% 100%}}#logo{display:inline-block;height:54px;width:150px} </style> <a href=//www.google.com/><span id=logo aria-label=Google></span></a> <p><b>404.</b> <ins>That’s an error.</ins> <p>The requested URL <code>/healthz</code> was not found on this server. <ins>That’s all we know.</ins> ".

opened by VietThan 1

Got RuntimeError when run on Google Colab

I ran the Readme.md samples on Google Colab with GPU and got this Error "RuntimeError: Expected all tensors to be on the same device, but found at least two devices, cpu and cuda:0! (when checking argument for argument index in method wrapper__index_select)".

Error code: outputs = model.generate(tokenizer(inputs, return_tensors="pt", padding=True).input_ids.to('cuda'), max_length=512)

opened by kietbg0079 0
Got error 'AssertionError: Torch not compiled with CUDA enabled' on Macbook M1 pro

I have tried the example on my Macbook M1 pro but got this error: =>outputs = model.generate(tokenizer(inputs, return_tensors="pt", padding=True).input_ids.to('cuda'), max_length=512) raise AssertionError("Torch not compiled with CUDA enabled") AssertionError: Torch not compiled with CUDA enabled

Please help!

opened by htnha 4
Question about loading model

I have a question about loading model. I have trained a Russian-to-Vietnamese model base on your code and tensor2tensor. Every time I want to predict a new sentence, it always load the model again, even before that I have already predicted another sentence. I want to ask that if there is a way not to have reload the model when predict a new sentence. Thank you very much

opened by hieunguyenquoc 1
I have a issue about running decoder

Data loss: Unable to open table file /content/drive/MyDrive/SAT/checkpoint: Failed precondition: /content/drive/MyDrive/SAT/checkpoint; Is a directory: perhaps your file is in a different file format and you need to use a different restore operator?

I used a pretrain model : model.augmented.envi.ckpt-1415000.data-00000-of-00001, model.augmented.envi.ckpt-1415000.index, model.augmented.envi.ckpt-1415000.meta. All 3 file are put in checkpoint

Could somebody help me with this issue ?

opened by hieunguyenquoc 6

Releases(v1.0)

v1.0(Oct 2, 2021)

First version.

Trained on 3.3M training data points. Transformer with 9-layer encoder and 9-layer decoder. Tested on a multi-domain dataset, outperforming Google Translate. Experiments with style-tagging and data appending.
Source code(tar.gz)
Source code(zip)

Owner

GitHub Repository

Styled Augmented Translation

Related tags

Overview

SAT

Style Augmented Translation

Introduction

Using the code

Dataset

Comments

Data leakage issue in evaluation?

Demo website is not working

Got RuntimeError when run on Google Colab

Got error 'AssertionError: Torch not compiled with CUDA enabled' on Macbook M1 pro

Question about loading model

I have a issue about running decoder

Releases(v1.0)

v1.0(Oct 2, 2021)

Owner

A PyTorch implementation for PyramidNets (Deep Pyramidal Residual Networks)

This is an official implementation for "Video Swin Transformers".

Accuracy Aligned. Concise Implementation of Swin Transformer

[WACV 2022] Contextual Gradient Scaling for Few-Shot Learning

A set of tests for evaluating large-scale algorithms for Wasserstein-2 transport maps computation.

Some code of the implements of Geological Modeling Using 3D Pixel-Adaptive and Deformable Convolutional Neural Network

Code for our CVPR 2022 Paper "GEN-VLKT: Simplify Association and Enhance Interaction Understanding for HOI Detection"

PanopticBEV - Bird's-Eye-View Panoptic Segmentation Using Monocular Frontal View Images

Pytorch cuda extension of grid_sample1d

CoCosNet v2: Full-Resolution Correspondence Learning for Image Translation

DeepMind's software stack for physics-based simulation and Reinforcement Learning environments, using MuJoCo.

Repository for the NeurIPS 2021 paper: "Exploiting Domain-Specific Features to Enhance Domain Generalization".

This repo contains the implementation of the algorithm proposed in Off-Belief Learning, ICML 2021.

Code for Discriminative Sounding Objects Localization (NeurIPS 2020)

Code from Daniel Lemire, A Better Alternative to Piecewise Linear Time Series Segmentation

PyTorch Implementation of Vector Quantized Variational AutoEncoders.

Code for the RA-L (ICRA) 2021 paper "SeqNet: Learning Descriptors for Sequence-Based Hierarchical Place Recognition"

Symbolic Parallel Adaptive Importance Sampling for Probabilistic Program Analysis in JAX

Towards Multi-Camera 3D Human Pose Estimation in Wild Environment

Implementation of our paper 'RESA: Recurrent Feature-Shift Aggregator for Lane Detection' in AAAI2021.