Robustar

Interactive Toolbox for Precise Data Annotation and Robust Vision Learning

This repository is for end users. For source code, please visit our source code repository.

For a user-oriented introduction, please refer to our documentation and tutorial pages

Robustar Docs

For a more academic-style introduction, one can also refer to our workshop paper released recently:

Robustar: Interactive Toolbox Supporting Precise Data Annotation for Robust Vision Learning

Main Use Case (Workflow):

the model can be trained elsewhere and then fed into the software;
With new test samples, the model can help identify the samples that are responsible for the prediction through influence function;
the software offers saliency map to help the user know which part of the features the model are paying necessary attention;
the users can use the drawing tools to brush out the superficial pixels;
new annotation of these images will serve as the role as augmented images for continued training.

Video Summary _{^{_{(click to watch the video)}}}

Installation and Usage

To install Robustar, please first intall Docker locally. If you wish to use GPU, please install a few more dependencies withinstall-nvidia-toolkit.sh .
Then, one can start the robustar with robustar.sh included in this repository. You may need to use sudo if you are on UNIX systems, or open run a terminal as system admin on Windows.
- First time users please run ./robustar.sh -m setup -a <version> to pull docker image. Please visit our DockerHub page for a complete list of versions.
- One can run ./robustar.sh -m run <options> to start robustar. Please read the next section for detailed options. You can also refer to example.sh for sample running configurations.
Directly run ./robustar.sh will display the help message.

Click to Expand Built-in Help Message

Help documentation for robustar.

Basic usage: robustar -m [command] [options]

[command] can be one of the following: setup, run.

setup will prepare and pull the docker image, and create a new container for it.

run will start to run the system.

Command line switches [options] are optional. The following switches are recognized.
-p --Sets the value for the port docker forwards to. Default is 8000.
-a --Sets the value for the tag of the image. Default is latest.
-n --Sets the value for the name of the docker container. Default is robustar.
-t --Sets the path of training images folder. Currently only supports the PyTorch DataLoader folder structure as following
-----images/
----------dogs/
---------------1.png
---------------2.png
----------cats/
---------------adc.png
---------------eqx.png
-e --Sets the path of testing images folder. Currently only supports the PyTorch DataLoader folder structure
-i --Sets the path of the calculation result of the influence function.
-c --Sets the path of model check points folder.
-o --Sets the path of configuration file. Default is configs.json.
-h --Displays this help message. No further functions are performed.