Episodic Transformer for Vision-and-Language Navigation
Alexander Pashevich, Cordelia Schmid, Chen Sun
Episodic Transformer (E.T.) is a novel attention-based architecture for vision-and-language navigation. E.T. is based on a multimodal transformer that encodes language inputs and the full episode history of visual observations and actions. This code reproduces the results obtained with E.T. on ALFRED benchmark. To learn more about the benchmark and the original code, please refer to ALFRED repository.
Clone repo:
$ git clone https://github.com/alexpashevich/E.T..git ET
$ export ET_ROOT=$(pwd)/ET
$ export ET_LOGS=$ET_ROOT/logs
$ export ET_DATA=$ET_ROOT/data
$ export PYTHONPATH=$PYTHONPATH:$ET_ROOT
Install requirements:
$ virtualenv -p $(which python3.7) et_env
$ source et_env/bin/activate
$ cd $ET_ROOT
$ pip install --upgrade pip
$ pip install -r requirements.txt
Download ALFRED dataset:
$ cd $ET_DATA
$ sh download_data.sh json_feat
Copy pretrained checkpoints:
$ wget http://pascal.inrialpes.fr/data2/apashevi/et_checkpoints.zip
$ unzip et_checkpoints.zip
$ mv pretrained $ET_LOGS/
Render PNG images and create an LMDB dataset with natural language annotations:
$ python -m alfred.gen.render_trajs
$ python -m alfred.data.create_lmdb with args.visual_checkpoint=$ET_LOGS/pretrained/fasterrcnn_model.pth args.data_output=lmdb_human args.vocab_path=$ET_ROOT/files/human.vocab
Note #1: For rendering, you may need to configure args.x_display
to correspond to an X server number running on your machine.
Note #2: We do not use JPG images from the full
dataset as they would differ from the images rendered during evaluation due to the JPG compression.
Evaluate an E.T. agent trained on human data only:
$ python -m alfred.eval.eval_agent with eval.exp=pretrained eval.checkpoint=et_human_pretrained.pth eval.object_predictor=$ET_LOGS/pretrained/maskrcnn_model.pth exp.num_workers=5 eval.eval_range=None exp.data.valid=lmdb_human
Note: make sure that your LMDB database is called exactly lmdb_human
as the word embedding won't be loaded otherwise.
Evaluate an E.T. agent trained on human and synthetic data:
$ python -m alfred.eval.eval_agent with eval.exp=pretrained eval.checkpoint=et_human_synth_pretrained.pth eval.object_predictor=$ET_LOGS/pretrained/maskrcnn_model.pth exp.num_workers=5 eval.eval_range=None exp.data.valid=lmdb_human
Note: For evaluation, you may need to configure eval.x_display
to correspond to an X server number running on your machine.
Train an E.T. agent:
$ python -m alfred.model.train with exp.model=transformer exp.name=et_s1 exp.data.train=lmdb_human train.seed=1
Evaluate the trained E.T. agent:
$ python -m alfred.eval.eval_agent with eval.exp=et_s1 eval.object_predictor=$ET_LOGS/pretrained/maskrcnn_model.pth exp.num_workers=5
Note: you may need to train up to 5 agents using different random seeds to reproduce the results of the paper.
Language encoder pretraining with the translation objective:
$ python -m alfred.model.train with exp.model=speaker exp.name=translator exp.data.train=lmdb_human
Train an E.T. agent with the language pretraining:
$ python -m alfred.model.train with exp.model=transformer exp.name=et_synth_s1 exp.data.train=lmdb_human train.seed=1 exp.pretrained_path=translator
Evaluate the trained E.T. agent:
$ python -m alfred.eval.eval_agent with eval.exp=et_synth_s1 eval.object_predictor=$ET_LOGS/pretrained/maskrcnn_model.pth exp.num_workers=5
Note: you may need to train up to 5 agents using different random seeds to reproduce the results of the paper.
You can also generate more synthetic trajectories using generate_trajs.py, create an LMDB and jointly train a model on it. Please refer to the original ALFRED code to know more the data generation. The steps to reproduce the results are the following:
- Generate 45K trajectories with
alfred.gen.generate_trajs
. - Create a synthetic LMDB dataset called
lmdb_synth_45K
usingargs.visual_checkpoint=$ET_LOGS/pretrained/fasterrcnn_model.pth
andargs.vocab_path=$ET_ROOT/files/synth.vocab
. - Train an E.T. agent using
exp.data.train=lmdb_human,lmdb_synth_45K
.
If you find this repository useful, please cite our work:
@inproceedings{pashevich2021episodic,
title = {{Episodic Transformer for Vision-and-Language Navigation}},
author = {Alexander Pashevich and Cordelia Schmid and Chen Sun},
booktitle = {ICCV},
year = {2021},
}