table of Contents
2 Introduction of Tensortboard startup and interface meaning
3 Tensorboard visualization of custom variables
3.1 Custom Training Tensorboard visualization
3.2 We can customize the learning rate:
3.3 Image of custom learning rate:
0 Preface
TensorBoard is a visualization tool released to facilitate the understanding, debugging and optimization of TensorFlow programs .
1 Use the callback function
TensorBoard runs by reading TensorFlow event files . The TensorFlow event file contains the main data you will be involved in the operation of TensorFlow .
2 Introduction of Tensortboard startup and interface meaning
- SCALARS panels are mainly used as the recording accuracy, and the learning rate loss and other trends single value . Use tf.summary.scalar() in the code to record it in the file: there are 3 small icons in the lower right corner of each graph, the first is to view the large graph, and the second is whether to logarithm to the y axis The third is if you drag or zoom the coordinate axis, and then return to the original position.
- The GRAPHS panel displays the overall structure of the network you have built , displays the direction and size of the data flow, and can also display the time spent on each node during training , the amount of memory consumed, and the number of parameters. The graph displayed by default is divided into two parts: Main Graph and Auxiliary Nodes. The main diagram shows the network structure, and the auxiliary nodes show the initialization , training, and saving nodes. We can double-click a node or click + in the upper right corner of the node to expand and view the situation inside, or align to zoom
- DISTRIBUTIONS is mainly used to show the changes of various parameters in the network with the increase of the number of training steps , which can be said to be a stack of multi-quantile line graphs.
- HISTOGRAMS and DISTRIBUTIONS are different ways of displaying the same data. The difference is that with DISTRIBUTIONS, HISTOGRAMS can be said that the frequency distribution histogram of the stack .
3 Tensorboard visualization of custom variables
3.1 Tensorboard visualization in custom training
