How graviraja/seq2seq works
This repository is a collection of standalone PyTorch sequence-model implementations, not a shared application. Its principal workflow is to train and evaluate a translation model, then use it for translation; the scripts implement alternative recurrent, attention-based, and Transformer approaches. BERT is a separate pretraining example. Training and model components are separated below only to show the calls visible in the excerpts; most behavior lives together in each script. The README documents all six capabilities, but does not establish shared orchestration between them.
Main entry points: simple_seq2seq.py Train and Evaluate · phrase_seq2seq.py Train and Evaluate · attention_seq2seq.py Train and Evaluate · padded_seq2seq.py Train and Evaluate
Components 12
Recurrent Translation
- Seq2Seq Model
simple_seq2seq.py - Train and Evaluate
simple_seq2seq.py - Phrase Seq2Seq
phrase_seq2seq.py - Train and Evaluate
phrase_seq2seq.py
Attention Translation
- Attention Seq2Seq
attention_seq2seq.py - Train and Evaluate
attention_seq2seq.py - Padded Attention
padded_seq2seq.py - Train and Evaluate
padded_seq2seq.py
Transformer Models
- Transformer Translation
transformer.py - Train and Evaluate
transformer.py - BERT Pretraining Model
BERT_basic.py - Pretraining Example
BERT_basic.py
Connections 5
- Train and Evaluate → Seq2Seq Model: calls model
simple_seq2seq.py - Train and Evaluate → Phrase Seq2Seq: calls model
phrase_seq2seq.py - Train and Evaluate → Attention Seq2Seq: calls model
attention_seq2seq.py - Train and Evaluate → Padded Attention: calls model
padded_seq2seq.py - Train and Evaluate → Transformer Translation: calls model
transformer.py