Contents

Getting Started

• Overview
  • Installation
  • Citation
  • Additional resources
• Quickstart
  • Step 0: Install OpenNMT-py
  • Step 1: Prepare the data
  • Step 2: Train the model
  • Step 3: Translate
• Versions
  • Breaking changes
  • Performance tips
• Contributors
  • Guidelines
• References

Frequently Asked Questions

• How do I use my v2 models in v3 ?
• How do I train the Transformer model?
• Performance tips
• Position encoding: Absolute vs Relative vs Rotary Embeddings vs Alibi
• Do you support multi-gpu?
• How do I use Pretrained embeddings (e.g. GloVe)?
  • Example
• How can I ensemble Models at inference?
• How can I weight different corpora at training?
  • Example
• What special tokens does OpenNMT-py use?
• How can I apply on-the-fly tokenization and subword regularization when training?
  • Example
• What are the readily available on-the-fly data transforms?
  • General purpose
  • Tokenization
  • BART-style noise
  • SwitchOut and sampling
• How can I create custom on-the-fly data transforms?
• How to use LoRa and 8bit loading to finetune a big model ?
• Can I get word alignments while translating?
  • Raw alignments from averaging Transformer attention heads
  • Supervised learning on a specific head
• How can I update a checkpoint’s vocabulary?
• How can I use source word features?
• How can I set up a translation server ?
  • I. How it works?
  • II. How to start the server without Docker ?
  • III. How to start the server with Docker ?
  • IV. How to use the API ?

Examples

• Translation WMT17 en-de
  • Dependencies
  • Running WMT17 EN-DE
• Language Model Wiki-103
  • Step 0: Download and clean the data
  • Step 1: Prepare the subword model - BPE with pyonmttok
  • Step 2: Build the vocabulary
  • Step 3: Train the model
  • Step 4: Generate output
• Summarization CNN/DM
  • Preparing the data and vocab
  • Training
  • Inference
  • Evaluation
  • Scores and Models
  • References
• Gated Graph Neural Networks
  • Dependencies
  • Quick Start
  • Graph data format
  • Vocabulary notes
  • Options
  • Acknowledgement

Scripts

• Build Vocab
  • Configuration
  • Data
  • Vocab
  • Features
  • Transform/Filter
  • Transform/Prefix
  • Transform/Suffix
  • Transform/FuzzyMatching
  • Transform/InferFeats
  • Transform/InlineTags
  • Transform/Normalize
  • Transform/BART
  • Transform/Docify
  • Transform/SwitchOut
  • Transform/Token_Drop
  • Transform/Token_Mask
  • Transform/Uppercase
  • Transform/Subword/Common
  • Transform/Subword/ONMTTOK
  • Transform/Clean
  • Reproducibility
• Train
  • Configuration
  • Data
  • Vocab
  • Features
  • Pruning
  • Embeddings
  • Transform/Filter
  • Transform/Prefix
  • Transform/Suffix
  • Transform/FuzzyMatching
  • Transform/InferFeats
  • Transform/InlineTags
  • Transform/Normalize
  • Transform/BART
  • Transform/Docify
  • Transform/SwitchOut
  • Transform/Token_Drop
  • Transform/Token_Mask
  • Transform/Uppercase
  • Transform/Subword/Common
  • Transform/Subword/ONMTTOK
  • Transform/Clean
  • Model-Embeddings
  • Model-Embedding Features
  • Model- Task
  • Model- Encoder-Decoder
  • Model- Attention
  • Model - Alignement
  • Generator
  • General
  • Reproducibility
  • Initialization
  • Optimization- Type
  • Optimization- Rate
  • Logging
  • Dynamic data
• Translate
  • Configuration
  • Model
  • Data
  • Features
  • Beam Search
  • Random Sampling
  • Reproducibility
  • Penalties
  • Decoding tricks
  • Logging
  • Efficiency
  • Transform/Filter
  • Transform/Prefix
  • Transform/Suffix
  • Transform/FuzzyMatching
  • Transform/InferFeats
  • Transform/InlineTags
  • Transform/Normalize
  • Transform/BART
  • Transform/Docify
  • Transform/SwitchOut
  • Transform/Token_Drop
  • Transform/Token_Mask
  • Transform/Uppercase
  • Transform/Subword/Common
  • Transform/Subword/ONMTTOK
  • Transform/Clean
• Server
  • Named Arguments

API

• Framework
  • Model
  • Trainer
  • Loss
  • Optimizer
• Modules
  • Embeddings
  • Encoders
  • Decoders
  • Attention
• Translation
  • Translations
  • Translator Class
  • Decoding Strategies
  • Scoring
• Server
  • Models
  • Core Server
• Data Loaders
  • Data Iterator
  • Dataset

Legacy

• FAQ (Legacy version)
  • How do I use Pretrained embeddings (e.g. GloVe)?
  • How do I use the Transformer model?
  • Do you support multi-gpu?
  • How can I ensemble Models at inference?
  • How can I weight different corpora at training?
  • Can I get word alignment while translating?
• Image to Text
  • Dependencies
  • Quick Start
  • Options
• Speech to Text
  • Dependencies
  • Quick Start
  • Options
  • Acknowledgement
• Video to Text
  • Recurrent
  • Transformer