This is the official project page for the NoScope project.
Please read the blog post and paper for more details!
This repository contains the code for the optimization step in the paper. The inference code is here.
You will need the following installed:
Your machine will need at least:
To set up the inference engine, do the following: Note: It is recommended that you create a folder that contains this repository, the tensorflow-noscope repository, and the data folder referred to below.
git clone https://github.com/stanford-futuredata/tensorflow-noscope.git
cd tensorflow-noscope
git checkout speedhax
git submodule init
git submodule update
./configure
cd tensorflow
bazel build -c opt --copt=-mavx2 --config=cuda noscope
The build will fail. To fix this, update the BUILD file to point towards your OpenCV install and add this directory to your PATH environmental variable. The BUILD file is in the tensorflow-noscope git repository at tensorflow/noscope/BUILD. You will need to change all references to "/lfs/0/ddkang/". You will probably need to change these to "/usr/" if you installed OpenCV using the directions above.
Please encourage the Tensorflow developers to support non-bazel building and linking. Due to a quirk in bazel, it may occasionally "forget" that tensorflow-noscope was built. If this happens, rebuild.
To set up the optimization engine, install the NoScope python package by going to the root directory of where you checked out https://github.com/stanford-futuredata/noscope and running "pip install -e ./"
Once you have inference engine set up, the example/
subfolder within this repository contains the
script to reproduce Figure 5d in the paper.
In order to run this:
wget https://storage.googleapis.com/noscope-data/csvs-yolo/coral-reef-long.csv
wget https://storage.googleapis.com/noscope-data/videos/coral-reef-long.mp4
code
and data
paths in example/run.sh. code
should point to the folder that contains
both the noscope and tensorflow-noscope folders. This value is how the optimization and inference engines
find eachother. data
should point to the data folder created in this section.The datasets that are currently available are coral-reef-long
and jackson-town-square
. Due to the expense of hosting these files, we have turned on requester pays for download. Please use an authenticated gsutil
to download the files.
The mp4 video files are available at https://storage.googleapis.com/noscope-data/videos/VIDEO_NAME.mp4
The CSVs with ground truth labels are available at
https://storage.googleapis.com/noscope-data/csvs-yolo/VIDEO_NAME.csv
If the above links do not work, you can download the data on Google drive here.