Awesome

Robust Navigation with Language Pretraining and Stochastic Sampling

Introduction

This repository contains source code and trained checkpoint to reproduce the results presented in the paper Robust Navigation with Language Pretraining and Stochastic Sampling.

Download

We provide two trained model checkpoints of Bert-base.

wget https://xiuldlstorage.blob.core.windows.net/r2r/public/emnlp/$MODEL_NAME.zip
unzip $MODEL_NAME.zip -d $MODEL_DIR

MODEL_NAME could be spl_53, spl_54.4.

Download the generated paths for data augmentation.

wget https://xiuldlstorage.blob.core.windows.net/r2r/public/emnlp/data/R2R_bi_12700_seed10-60_literal_speaker_data_aug_paths_unk.json
wget https://xiuldlstorage.blob.core.windows.net/r2r/public/emnlp/data/R2R_bi_12700_seed10-60_literal_speaker_data_aug_paths_unk_bert.txt

Download the bi-directional speaker generated paths for data augmentation.

wget https://xiuldlstorage.blob.core.windows.net/r2r/public/emnlp/data/R2R_bi_12700_seed10-60_literal_speaker_data_augmentation_paths.json
wget https://xiuldlstorage.blob.core.windows.net/r2r/public/emnlp/data/R2R_bi_12700_seed10-60_literal_speaker_data_augmentation_paths_bert.txt

Download some pre-trained Bert base and large checkpoints for play.

wget https://xiuldlstorage.blob.core.windows.net/r2r/public/emnlp/pretrain.zip

Installation

Please follow R2R for the environment setup.

# install bert package
pip install pytorch-pretrained-bert

Run Training

Script to run training (not update Bert, transformer_update=False).

# For example, best val_unseen spl = 0.53, best val_unseen sr = 0.58
CUDA_VISIBLE_DEVICES=0 python ./tasks/R2R/train.py --feedback_method teacher --bidirectional True --encoder_type bert --top_lstm True --transformer_update False --batch_size 20 --log_every 40 --pretrain_n_sentences 6 --pretrain_splits bi_12700_seed10-60_literal_speaker_data_aug_paths_unk --save_ckpt 10000 --ss_n_pretrain_iters 50000 --pretrain_n_iters 60000 --ss_n_iters 60000 --n_iters 70000 --dropout_ratio 0.4 --dec_h_type vc --schedule_ratio 0.4 --optm Adamax --att_ctx_merge mean --clip_gradient_norm 0 --clip_gradient 0.1 --use_pretrain --action_space -1 --pretrain_score_name sr_unseen --train_score_name sr_unseen --enc_hidden_size 1024 --hidden_size 1024 --result_dir ./base/results/ --snapshot_dir ./base/snapshots/ --plot_dir ./base/plots/

Script to run finetuning (update Bert, transformer_update=True).

CUDA_VISIBLE_DEVICES=0 python ./tasks/R2R/train.py --feedback_method teacher --dropout_ratio 0.4 --dec_h_type vc --optm Adamax --schedule_ratio 0.2 --att_ctx_merge mean --clip_gradient_norm 0 --clip_gradient 0.1 --log_every 32 --action_space -1 --n_iters 34000 --train_score_name sr_unseen --enc_hidden_size 1024 --hidden_size 1024 --result_dir ./base/results/ --snapshot_dir ./base/snapshots/ --plot_dir ./base/plots/ --n_iters_resume N --ss_n_iters N+10000 --save_ckpt 512 --bidirectional True --encoder_type bert --top_lstm True --transformer_update True --batch_size 16 --learning_rate 5e-5

N is the iteration of the best checkpoint from the above training.

# For example, with a pre-train base checkpoint, N = 61040, best val_unseen spl = 0.556, best val_unseen sr = 0.602
CUDA_VISIBLE_DEVICES=0 python ./tasks/R2R/train.py --feedback_method teacher --n_iters_pretrain_resume 61040 --learning_rate 0.0001 --pretrain_model_path path_to/pretrain/bert_s/ --save_ckpt 9600 --ss_n_pretrain_iters 71000 --pretrain_n_iters 81000 --ss_n_iters 81000 --n_iters 91000 --bidirectional True --encoder_type bert --top_lstm True --bert_n_layers 1 --transformer_update True --batch_size 16 --log_every 48 --pretrain_n_sentences 6 --pretrain_splits bi_12700_seed10-60_literal_speaker_data_augmentation_paths --dropout_ratio 0.4 --dec_h_type vc --schedule_ratio 0.3 --optm Adamax --att_ctx_merge mean --clip_gradient_norm 0 --clip_gradient 0.1 --use_pretrain --action_space -1 --pretrain_score_name sr_unseen --train_score_name spl_unseen --enc_hidden_size 1024 --hidden_size 1024

Run Validation and Test

Script to play with the checkpoint on the paper (val unseen spl=55).

CUDA_VISIBLE_DEVICES=0 python ./tasks/R2R/train.py --panoramic True --result_dir ./test --snapshot_dir ./snapshots --plot_dir ./plot --action_space -1 --n_iters 10 --att_ctx_merge mean --n_iters_resume 63480 --sc_after 0 --sc_score_name sr_unseen --train False --val_splits val_seen,val_unseen,test --enc_hidden_size 1024 --hidden_size 1024 --feedback_method teacher --clip_gradient 0.1 --clip_gradient_norm 0 --dec_h_type vc --schedule_ratio -1.0 --dump_result --bidirectional True --optm Adamax --encoder_type bert --top_lstm True --transformer_update False --batch_size 24 --pretrain_model_path path_to/spl_53/snapshots/

Script to play with a better checkpoint (val unseen spl=56.2).

CUDA_VISIBLE_DEVICES=0 python ./tasks/R2R/train.py --panoramic True --result_dir ./test --snapshot_dir ./snapshots --plot_dir ./plot --action_space -1 --n_iters 10 --att_ctx_merge mean --n_iters_resume 68576 --sc_after 0 --sc_score_name sr_unseen --train False --val_splits val_seen,val_unseen,test --enc_hidden_size 1024 --hidden_size 1024 --feedback_method teacher --clip_gradient 0.1 --clip_gradient_norm 0 --dec_h_type vc --schedule_ratio -1.0 --dump_result --bidirectional True --optm Adamax --encoder_type bert --top_lstm True --transformer_update False --batch_size 24 --pretrain_model_path path_to/spl_54.4/snapshots/

Citations

Please consider citing this paper if you use the code:

@article{li2019robust,
  title={Robust Navigation with Language Pretraining and Stochastic Sampling},
  author={Li, Xiujun and Li, Chunyuan and Xia, Qiaolin and Bisk, Yonatan and Celikyilmaz, Asli and Gao, Jianfeng and Smith, Noah and Choi, Yejin},
  conference={EMNLP},
  year={2019}
}