代码：Zero-Shot Visual Imitation

https://github.com/pathak22/zeroshot-imitation

Zero-Shot Visual Imitation

In ICLR 2018 [Project Website] [Videos]

Deepak Pathak*, Parsa Mahmoudieh*, Guanghao Luo*, Pulkit Agrawal*, Dian Chen, 
Yide Shentu, Evan Shelhamer, Jitendra Malik, Alexei A. Efros, Trevor Darrell
University of California, Berkeley

 

This is the implementation for the ICLR 2018 paper Zero Shot Visual Imitation. We propose an alternative paradigm wherein an agent first explores the world without any expert supervision and then distills its experience into a goal-conditioned skill policy with a novel forward consistency loss. The key insight is the intuition that, for most tasks, reaching the goal is more important than how it is reached.

@inproceedings{pathakICLR18zeroshot,   Author = {Pathak, Deepak and   Mahmoudieh, Parsa and Luo, Guanghao and   Agrawal, Pulkit and Chen, Dian and   Shentu, Yide and Shelhamer, Evan and   Malik, Jitendra and Efros, Alexei A. and   Darrell, Trevor},   Title = {Zero-Shot Visual Imitation},   Booktitle = {ICLR},   Year = {2018}}

1) Installation and Usage

Requirements

git clone -b master --single-branch https://github.com/pathak22/zeroshot-imitation.gitcd zeroshot-imitation/# (1) Install requirements:sudo apt-get install python-tkvirtualenv venvsource $PWD/venv/bin/activatepip install --upgrade pippip install numpypip install -r src/requirements.txt# (2) Install Caffe: http://caffe.berkeleyvision.org/install_apt.htmlgit clone https://github.com/BVLC/caffe.gitsudo apt-get install libprotobuf-dev libleveldb-dev libsnappy-dev libopencv-dev libhdf5-serial-dev protobuf-compilersudo apt-get install libatlas-base-devsudo apt-get install libgflags-dev libgoogle-glog-dev liblmdb-devsudo apt-get install --no-install-recommends libboost-all-devcd caffe/  # edit Makefile.configmake all -jmake pycaffemake test -jmake runtest -j# Note: If you are using conda, then its easy:# $ conda install -c conda-forge caffe# $ conda install -c conda-forge opencv=3.2.0

Data setup

Data can be downloaded at google drive link. This is the same data as used in Combining Self-Supervised Learning and Imitation for Vision-Based Rope Manipulation.

You will need the rope9 dataset and img_mean.npy from this download.

Then, download the AlexNet weights, bvlc_alexnet.npy from here

• put rope9 data in data/datasets/rope9

• put img_mean.npy in data/img_mean.npy

• put bvlc_alexnet.npy in nets/bvlc_alexnet.npy

Training

python -i train.py# fwd_consist=True to turn foward consistency loss on,# or leave it False for to just learn the inverse modelr = RopeImitator('name', fwd_consist=True)# to train baseline, turn baseline_reg=True. note that fwd_consist should be turned on as well (historical accident)r = RopeImitator('name', fwd_consist=True, baseline_reg=True)# Restore old models, if any. default of model_name is just current model namer.restore(iteration, model_name='name of old model')# trainingr.train(num_iters)

Note that the accuracies presented is not a good measure of real world performance. The purpose of forward consistency is to learn actions consistent with state transistions, which don't necessarily have to be the ground truth actions.

2) Other resources

• Paper

• Project Website

• Videos