Dash-deep

A pure python web service (based on plotly/dash) that is capable of

1. Creation, annotation and updating of image segmentation datasets.
2. Running training and inference jobs.
3. Monitoring the progress of training jobs and related accuracy metrics.
4. Monitoring and halting of running jobs.
5. GPU memory and utilization monitoring.

In other words, it is a pure python customizable version of Nvidia DIGITS with a focus on image segmentation task.

You can see a video of example usage here.

So far supports only Python 2 -- porting to Python 3 is in progress. The project is under development and will most probably change a lot in a future.

Image annotation

We have a support for simple web based image uploading and annotation. This can be used to create a new dataset, make your annotations and run the training code. So far the tool supports:

1. Zooming in and out
2. Magic lasso
3. Rectangle selection
4. Cropping.

See a simple example below:

Training accuracy/loss monitoring

We support real-time monitoring of training metrics which can be customized, but by default support validation accuracy and training loss.

Updating existing dataset

You can update saved dataset at any time:

Running

To start the server run:

 python -m dash_deep.index run_server <ip_address> <port>

You can use ip address of 0.0.0.0 which will automatically use the network ip of the machine that you are running the server on.

Open browser and navigate to:

<ip_address>:<port>/

To get the help page, call:

 python -m dash_deep.index --help

The default login name and pass are hello and world.