Home

Awesome

Curriculum Learning For LVN

License: MIT

This is the PyTorch implementation of our paper:<br> Curriculum Learning for Vision-and-Language Navigation [arxiv]<br> Jiwen Zhang, Zhongyu Wei, Jianqing Fan, Jiajie Peng<br> 35th Conference on Neural Information Processing Systems (NeurIPS 2021)

Most Recent Events

Model architectures

This repository includes several SOTA navigation agents previously released. They are

and a path-instruction scorer

Installation

Setting up Environments

  1. Install Python 3.6 (Anaconda recommended: https://docs.anaconda.com/anaconda/install/index.html).

  2. Install PyTorch following the instructions on https://pytorch.org/ (in our experiments, it isPyTorch 1.5.1+cu101).

  3. Following build instructions in this github to build up a v0.1 Matterport3D simulator.

    Besides, just in case you have an error when compiling the simulator, you can try this

    mkdir build && cd build
    cmake -D CUDA_TOOLKIT_ROOT_DIR=path/to/yout/cuda
    make
    cd ../
    

    For more details on the Matterport3D Simulator, you can refer to README_Matterport3DSimulator.md.

Dataset Download

Luckily, this repository contains the R2R dataset and CLR2R dataset, so you ONLY have to download precomputing ResNet image features from Matterport3DSimulator.

Download and extract the tsv files into the img_features directory. You will only need the ImageNet features to replicate our results.

Clone Repo

Clone (or just download) this reposiroty and replace <code>tasks</code> directory in original Matterport3D simulator with the one in this reposiroty.

After following the steps above the your file directory should look like this:

Matterport3D/
    build/            # should be complied in your machine
    cmake/
    connectivity/     # store Json connecivity graphs for each scan
    img_features/     # store precomputed image features, i.e. ResNet-152 features
    include/
    pybind11/         # a dependency of Matterport3D Simulator
    ...
    tasks/R2R-judy/   # replace it with the one in this directory
    ...

Usage Instructions

To replicate the Table 3 in our paper, try the following command in shell.

CONFIG_PATH="path-to-config-file"
CL_MODE="" # "" / "NAIVE" / "SELF-PACE"

python tasks/R2R-judy/main.py \
--config-file $CONFIG_PATH \
TRAIN.DEVICE your_device_id \
TRAIN.CLMODE $CL_MODE \
...

You can refer to <code>task/tasks/R2R-judy/runner</code> for more details.