Skip to content

Latest commit

 

History

History
34 lines (27 loc) · 1.3 KB

README.md

File metadata and controls

34 lines (27 loc) · 1.3 KB
Raw

Installation

This was tested on an AWS EC2 p3.8xlarge instance with Ubuntu 18.04. The instance has 4 Tesla V100 GPUs, 32 vCPUs and 244 GB Memory.

  1. Create a new conda environment and activate it.
conda create -n ns_model python=3.8
conda activate ns_model
  1. Install pytorch
pip install torch==1.12.1+cu116 torchvision==0.13.1+cu116 --extra-index-url https://download.pytorch.org/whl/cu116
  1. Install all other requirements
pip install -r modeling/ns_model/requirements.txt

Training

  1. Run the script to download the trajectory data and the vision model checkpoints.
scripts/fetch_vision_model.sh
scripts/fetch_trajectory_data.sh
  1. The entry point to the training and evaluation code is in modeling/ns_model/train_eval.py. The model is an episodic transformer model trained on the training data data/trajectory-data/train.json. The evaluation for action prediction accuracy is done on the validation data data/trajectory-data/valid.json.
export ALEXA_ARENA_DIR="$HOME/AlexaArena"
CUDA_VISIBLE_DEVICES=0 python -m modeling.ns_model.train_eval
  1. The trained models are stored in --checkpt-dir. By default it is stored in logs/ns_model_checkpt/. This can be used for end-to-end mission level evaluation following the instructions in modeling/inference/README.md