/encoder-agnostic-adaptation

Encoder-Agnostic Adaptation for Conditional Language Generation

Primary LanguagePythonMIT LicenseMIT

Encoder-Agnostic Adaptation for Conditional Language Generation

This repo contains the code used in Encoder-Agnostic Adaptation for Conditional Language Generation, Zachary M. Ziegler, Luke Melas-Kyriazi, Sebastian Gehrmann and Alexander M. Rush. It extends OpenNMT-py.

This code was tested with pytorch 1.0.1. See requirements.txt for a complete list of dependencies.

Download GPT2 weights

cd gpt2 && python download_model.py 124M

General notes

All experiments use gradient accumulation to mimic the large batch sizes these hyperparameter settings were optimized for by e.g. Facebook. If you run into GPU memory issues simply reduce the batch size and increase the accum_count to keep the effective batch size the same.

Data

The BPEized data used in the experiments in the paper can be found here. To run any of these models with your own data you should first BPEize it with python gpt2/encode_text.py <filename>. Before training the raw data is preprocessed into binary data shards with the commands below.

Class-conditional generation

Preprocess

python preprocess.py -train_src data/imdb/train.src.bpe -train_tgt data/imdb/train.tgt.bpe -valid_src data/imdb/valid.src.bpe -valid_tgt data/imdb/valid.tgt.bpe -save_data data/imdb/IMDB_BPETGT -tgt_seq_length_trunc 400 -tgt_vocab gpt2/vocab.txt -fixed_vocab -free_src

Train

Baseline: python train.py -config config/imdb/transformer_imdb_cond.yml -run_name baseline

Simple fusion: python train.py -config config/imdb/transformer_imdb_cond.yml -run_name simple_fusion -gpt2_params_path gpt2/models/124M/ -simple_fusion -dropout 0.1 -accum_count 30 -batch_size 1000 -valid_batch_size 16

Repr-transformer: python train.py -config config/imdb/transformer_imdb_cond.yml -run_name repr_trans -GPT_representation_loc tgt -GPT_representation_mode elmo -gpt2_params_path gpt2/models/124M/ -position_encoding_learned_dec -word_vec_size 768 -rnn_size 768

Context attention: python train.py -config config/imdb/transformer_imdb_ctxattn.yml -run_name ctxattn -gpt2_params_path gpt2/models/124M/ -gpt2_init_embanddec

Pseudo self attention: python train.py -config config/imdb/transformer_imdb_psa.yml -run_name psa -gpt2_params_path gpt2/models/124M/ -gpt2_init_embanddec

Generation

Generation is performed via random sampling.

python translate.py -beam_size 1 -random_sampling_topk -1 -random_sampling_temp 0.7 -model <path/to/model.pt> -src data/imdb/test.src.bpe -min_length 1 -max_length 400 -verbose

Summarization

Preprocess

python preprocess.py -train_src data/cnndm/train.txt.src.bpe -train_tgt data/cnndm/train.txt.tgt.bpe -valid_src data/cnndm/val.txt.src.bpe -valid_tgt data/cnndm/val.txt.tgt.bpe -save_data data/cnndm/CNNDM_BPE_COPY -src_seq_length_trunc 400 -tgt_seq_length_trunc 100 -src_vocab gpt2/vocab.txt -tgt_vocab gpt2/vocab.txt -dynamic_dict -fixed_vocab

Train

The default settings use 4 GPUs (see config files). If using more GPUs or fewer GPUs, modify the world_size and gpu_ranks values in the config file and adjust accum_count so the effective batch size remains the same.

Baseline: python train.py -config config/cnndm/transformer_cnndm_baseline.yml -run_name baseline

Repr-transformer: python train.py -config config/cnndm/transformer_cnndm_baseline.yml -run_name repr_trans -GPT_representation_loc tgt -GPT_representation_mode elmo -gpt2_params_path gpt2/models/124M/ -position_encoding_learned_dec -word_vec_size 768 -rnn_size 768 -train_steps 50000

Context attention: python train.py -config config/cnndm/transformer_cnndm_ctxattn.yml -run_name ctxattn -gpt2_params_path gpt2/models/124M/ -gpt2_init_embanddec

Pseudo self attention: python train.py -config config/cnndm/transformer_cnndm_psa.yml -run_name psa -gpt2_params_path gpt2/models/124M/ -gpt2_init_embanddec

Generation

Generation is performed via beam search.

python translate.py -beam_size 5 -model <path/to/model.pt> -src data/cnndm/test.txt.src.bpe -min_length 60 -verbose -block_ngram_repeat 3

Story generation

The default settings use 4 GPUs (see config files). If using more GPUs or fewer GPUs, modify the world_size and gpu_ranks values in the config file and adjust accum_count so the effective batch size remains the same.

Preprocess

python preprocess.py -train_src data/stories/train.wp_source.bpe -train_tgt data/stories/train.wp_target.bpe -valid_src data/stories/valid.wp_source.bpe -valid_tgt data/stories/valid.wp_target.bpe -save_data data/stories/STORIES_BPE -src_vocab gpt2/vocab.txt -tgt_vocab gpt2/vocab.txt -fixed_vocab

Train

Baseline: python train.py -config config/story_gen/transformer_story_baseline.yml -run_name baseline

Repr-transformer: python train.py -config config/story_gen/transformer_story_baseline.yml -run_name repr_trans -GPT_representation_loc tgt -GPT_representation_mode elmo -gpt2_params_path gpt2/models/124M/ -position_encoding_learned_dec -word_vec_size 768 -rnn_size 768

Context attention: python train.py -config config/story_gen/transformer_story_ctxattn.yml -run_name ctxattn -gpt2_params_path gpt2/models/124M/ -gpt2_init_embanddec

Pseudo self attention: python train.py -config config/story_gen/transformer_story_psa.yml -run_name psa -gpt2_params_path gpt2/models/124M/ -gpt2_init_embanddec

Generation

Generation is performed via top-k/random sampling.

python translate.py -beam_size 1 -random_sampling_topk 100 -random_sampling_temp 0.9 -model <path/to/model.pt> -src data/stories/test.wp_source.bpe -max_length 1000 -verbose

Image captioning

Coming soon...