usage_instructions

Usage Instructions

Environment requirement

Frontend

Bazel 3.2.0 [check here for more information]

Server

flask 1.1.2

Backend

Please check deepvisualinsight/requirement.txt

Prepare data for visualizing

please save the subject model and related data in ~/DVI_exp_data

Check your local IP address

On the machine you would like to launch the server, please check your local IP address. You may refer to here for help. Then, please change the IP address in the file tensorboard/tensorboard/plugins/projector/vz_projector/standalone_projector_config.json to your own local IP address:

1. "serverIp": "xxx.xxx.xxx.xxx" or "serverIp": "localhost". By default, "serverIp" is "localhost"

Set up environment

On Windows OS, you may need to close the firewall.

1. run conda create -n DVI python=3.6
pip install -r requirements.txt // Launch the server
2. run cd ~/DeepVisualInsight
3. run cd server
4. run python server.py
   To check if your server works normally, you may use another machine(e.g your phone) that connects to the same Local Area Network, then try to open http://xxx.xxx.xxx.xxx:5000/ where xxx.xxx.xxx.xxx is your local IP address. If successful, you can see a message Index Page

Launch the frontend

1. run cd ~/DeepVisualInsight/tensorboard
2. run bazel run tensorboard/plugins/projector/vz_projector:standalone. You may take several minutes at the first time.
3. Open http://localhost:6006/standalone.html.
   http://localhost:6006/standalone.html.
4. May wait few seconds for launching. Then Input Model Pathand input resolution. Then click on Next button.
   
   You may need to wait for a few seconds (or few minutes if you use online computation) for server side to retrieve information. At this time, you may switch the color map.
   
   Now you can enjoy DeepVisualInsight :)
   

Filtering

We support keywords below. If you input multiple keywords, they are considered to be combined by an and operation.

• label: (support multi choice)
• prediction: (support multi choice)
• is_training: true/false
• is_correct_prediction: true/false
• new_selection: true/false
• is_noisy: true/false
• active_learning: true => (training or new selection)
• noisy_type: original/flipped/others
• original_or_flipped: true => (original or flipped noisy data)

Hyperparameters

The following hyperparameters control the result of our tools.

Variable	Meaning
encoder	Dimension Reduction network
decoder	...

inspector panel

Search by metadata field

Field	type	predicate
type	str	"train", "test", "unlabel"
new_selection	int	iteration
prediction	str	"T", "W"
label	str	label name
uncertainty	int	ranking
diversity	int	ranking
tot	int	ranking