Seq2Seq in PyTorch
This is a complete suite for training sequence-to-sequence models in PyTorch. It consists of several models and code to both train and infer using them.
Using this code you can train:
- Neural-machine-translation (NMT) models
- Language models
- Image to caption generation
- Skip-thought sentence representations
- And more...
Installation
git clone --recursive https://github.com/eladhoffer/seq2seq.pytorch
cd seq2seq.pytorch; python setup.py develop
Models
Models currently available:
- Simple Seq2Seq recurrent model
- Recurrent Seq2Seq with attentional decoder
- Google neural machine translation (GNMT) recurrent model
- Transformer - attention-only model from "Attention Is All You Need"
Datasets
Datasets currently available:
- WMT16
- WMT17
- OpenSubtitles 2016
- COCO image captions
- Conceptual captions
All datasets can be tokenized using 3 available segmentation methods:
- Character based segmentation
- Word based segmentation
- Byte-pair-encoding (BPE) as suggested by bpe with selectable number of tokens.
After choosing a tokenization method, a vocabulary will be generated and saved for future inference.
Training methods
The models can be trained using several methods:
- Basic Seq2Seq - given encoded sequence, generate (decode) output sequence. Training is done with teacher-forcing.
- Multi Seq2Seq - where several tasks (such as multiple languages) are trained simultaneously by using the data sequences as both input to the encoder and output for decoder.
- Image2Seq - used to train image to caption generators.
Usage
Example training scripts are available in scripts folder. Inference examples are available in examples folder.
- example for training a transformer on WMT16 according to original paper regime:
DATASET=${1:-"WMT16_de_en"}
DATASET_DIR=${2:-"./data/wmt16_de_en"}
OUTPUT_DIR=${3:-"./results"}
WARMUP="4000"
LR0="512**(-0.5)"
python main.py \
--save transformer \
--dataset ${DATASET} \
--dataset-dir ${DATASET_DIR} \
--results-dir ${OUTPUT_DIR} \
--model Transformer \
--model-config "{'num_layers': 6, 'hidden_size': 512, 'num_heads': 8, 'inner_linear': 2048}" \
--data-config "{'moses_pretok': True, 'tokenization':'bpe', 'num_symbols':32000, 'shared_vocab':True}" \
--b 128 \
--max-length 100 \
--device-ids 0 \
--label-smoothing 0.1 \
--trainer Seq2SeqTrainer \
--optimization-config "[{'step_lambda':
\"lambda t: { \
'optimizer': 'Adam', \
'lr': ${LR0} * min(t ** -0.5, t * ${WARMUP} ** -1.5), \
'betas': (0.9, 0.98), 'eps':1e-9}\"
}]"
- example for training attentional LSTM based model with 3 layers in both encoder and decoder:
python main.py \
--save de_en_wmt17 \
--dataset ${DATASET} \
--dataset-dir ${DATASET_DIR} \
--results-dir ${OUTPUT_DIR} \
--model RecurrentAttentionSeq2Seq \
--model-config "{'hidden_size': 512, 'dropout': 0.2, \
'tie_embedding': True, 'transfer_hidden': False, \
'encoder': {'num_layers': 3, 'bidirectional': True, 'num_bidirectional': 1, 'context_transform': 512}, \
'decoder': {'num_layers': 3, 'concat_attention': True,\
'attention': {'mode': 'dot_prod', 'dropout': 0, 'output_transform': True, 'output_nonlinearity': 'relu'}}}" \
--data-config "{'moses_pretok': True, 'tokenization':'bpe', 'num_symbols':32000, 'shared_vocab':True}" \
--b 128 \
--max-length 80 \
--device-ids 0 \
--trainer Seq2SeqTrainer \
--optimization-config "[{'epoch': 0, 'optimizer': 'Adam', 'lr': 1e-3},
{'epoch': 6, 'lr': 5e-4},
{'epoch': 8, 'lr':1e-4},
{'epoch': 10, 'lr': 5e-5},
{'epoch': 12, 'lr': 1e-5}]" \
61 Dec 28, 2022
16 Dec 21, 2022
800 Jan 03, 2023
46 Sep 29, 2022
85 Dec 30, 2022
58 Dec 02, 2022
7 Nov 30, 2022
21 Dec 26, 2022
5 Oct 20, 2022
0 Jun 01, 2022
54 Dec 07, 2022
13 Aug 01, 2022
126 Dec 20, 2022
92 Dec 21, 2022
42 Dec 10, 2022
40 Aug 14, 2022
3 Jan 27, 2022
455 Jan 08, 2023
23 Nov 11, 2022
1.1k Dec 12, 2022