Data Science Workshop (DSW) of Platform for AI (PAI) provides the TensorBoard plug-in developed by the TensorFlow community as a visualization tool for deep learning training. You can use TensorBoard to view information about TensorFlow tasks from task logs in a visualized manner and understand the performance metrics of model trainings.
Features
TensorBoard provides the following features:
Training metrics monitoring
You can view the time-varying curves of metrics, such as loss and accuracy, in real time to obtain the model training progress.
Graph structure visualization
You can view the graph structure of a TensorFlow model, such as operations and layers.
Histogram visualization
You can view information such as histograms of weights, biases, and other tensors to analyze the distribution characteristics of model parameters, such as weights and biases. This helps you optimize the model and identify potential issues.
Image, audio, and text embedding visualization
You can visualize an embedding by projecting the embedding onto a low-dimensional space.
Other performance metrics
You can view metrics related to hardware resource utilization, such as GPU, CPU, memory, and GPU memory, and other custom metrics.
Procedure
For more information, see Introduction to use TensorBoard in DSW.