Chinese

TS-VIS(Tianshu Visualization) is a visualization tool kit of Tianshu AI Platform. , which support visualization of the most popular deep learning frameworks, such as TensorFlow, PyTorch, OneFlow, etc.

Document (Chinese): https://feyily.github.io/tsvis-document/

HighLights

• Framework-independent, support visualization of the most popular deep learning frameworks, such as TensorFlow, PyTorch, OneFlow, etc.
• Faster response speed
• Support the visualization of large-scale data
• Support real-time visualization during training
• Support embedding sample visualization
• Support neural network exception visualization
• Supports multiple feature map visualizations
• Support Transformer visualization
• Support RNN hidden state visualization

Features

• Graph: Visualize neural network structure, including computational graph and structure graph
• FeatureMap：Visualize convolutional features of convolutional neural networks
• Scalar: Visualize arbitrary scalar data including accuary and loss
• Media: Visualize media data including images, text, and audio
• Distribution: Visualize the distribution of weights, biases, etc. in neural network
• Embedding: Visualize arbitrary high-dimensional data through dimensionality reduction algorithm
• Hyperparameter: Visualize neural network indicators under different hyperparameters
• Exception: Map neural network tensor data to two dimensions, visualize tensor data statistics
• Custom: Move the charts in Scalar, Media, and Distribution to this module for comparison and viewing
• Attention:Visualize Attention Data for Image and Text Transformer Models
• HiddenState:Visualize RNN hidden state values and perform matching analysis

Install

We provide two installation methods: install by pip and install from source. No matter which method you pick, you need to make sure that your Python version is 3.6 or higher, otherwise please upgrade Python first.

Install by pip

pip install tsvis

Install from source

TS-VIS adopts the architecture of separation of frontend and backend, so you need to build the frontend and backend separately

• Build frontend from source:

  cd webapp
  

  Install dependencies first

  npm install
  

  Package frontend to generate static files

  npm run build
  

• Build backend from source:

  To install the backend, you need to first move the static files generated by previous step to tsvis/server/frontend folder

  Then install the Python dependency package setuptools

  pip install setuptools
  

  Run setup.py to install TS-VIS to your Python environment

  python setup.py install
  

Run

After installation, you can run the following command. If the version information is output in the console, it means that you have installed TS-VIS correctly.

tsvis -v

Then you can run the visualization with the following command

tsvis --logdir path/to/logdir/

By default, the visualization service will start at http://127.0.0.1:9898, open the browser to access the visualization content.