Logging MXNet Data for Visualization in TensorBoard
MXBoard provides a set of APIs for logging MXNet data for visualization in TensorBoard. The idea of this project comes from discussions with Zihao Zheng, the author of dmlc/tensorboard, on delivering a visualization solution for MXNet users. We aim at providing the logging APIs that can process MXNet data efficiently and supporting most of the data types for visualization in the TensorBoard GUI. We adapted the following low-level logging components from their Python and C++ implementations in TensorFlow:
_EventLoggerThread. We also adapted the user-level logging APIs defined in
SummaryWriter from tensorboard-pytorch. The encoding algorithm used in writing protobuf objects into event files is directly borrowed from TeamHG-Memex/tensorboard_logger.
MXBoard supports a set of Python APIs for logging the following data types for TensorBoard to render. Logging APIs for other languages may be added in the future.
The corresponding Python APIs are accessible through a class called
SummaryWriter as follows:
mxboard.SummaryWriter.add_graph mxboard.SummaryWriter.add_scalar mxboard.SummaryWriter.add_histogram mxboard.SummaryWriter.add_embedding mxboard.SummaryWriter.add_image mxboard.SummaryWriter.add_text mxboard.SummaryWriter.add_pr_curve mxboard.SummaryWriter.add_audio
Install MXBoard from PyPI
pip install mxboard
Install MXBoard Python package from source
git clone https://github.com/awslabs/mxboard.git cd mxboard/python python setup.py install
Install TensorBoard from PyPI
MXBoard is a logger for writing MXNet data to event files. To visualize those data in browsers, users still have to install TensorBoard separately.
pip install tensorboard
Use the following to verify that the TensorBoard binary has been installed correctly.
Other required packages
MXBoard relies on the following packages for data logging.
Please note that you need to install MXNet manually before using MXBoard. The other packages will be installed automatically when you install MXBoard via pip or building from source. If you want to build from source, please make sure that protobuf compiler is installed. Check this page for downloading the protobuf compiler whose file name starts with “protoc”.
Visualizing MXNet data in 30 seconds
Now that you have installed all of the required packages, let’s walk through a simple visualization example. You will see how MXBoard enables visualizing MXNet
NDArrays with histograms.
Step 1. Logging event data to a file.
Prepare a Python script for writing data generated by the
normal operator to an event file. The data is generated ten times with decreasing standard deviation and written to the event file each time. It’s expected to see the data distribution gradually become more centered around the mean value. Note that here we specify creating the event file in the folder
logs under the current directory. We will need to pass this folder path to the TensorBoard binary.
import mxnet as mx from mxboard import SummaryWriter with SummaryWriter(logdir='./logs') as sw: for i in range(10): # create a normal distribution with fixed mean and decreasing std data = mx.nd.normal(loc=0, scale=10.0/(i+1), shape=(10, 3, 8, 8)) sw.add_histogram(tag='norml_dist', values=data, bins=200, global_step=i)
Step 2. Launch TensorBoard to load the event file generated above.
Use the following command to start the TensorBoard server. It will use the logs that were generated in the current directory’s
tensorboard --logdir=./logs --host=127.0.0.1 --port=8888
Note that in some situations, the port number
8888 may be occupied by other applications and launching TensorBoard may fail. You may choose a different available port number.
Step 3. Open TensorBoard in your browser.
In the browser, enter the address
127.0.0.1:8888, and click the tab HISTOGRAMS in the TensorBoard GUI. You will see data distribution changing as time progresses.
- Quick start for logging data of various types
- Monitoring training an MNIST model with MXBoard
- Visualizing filters of ConvNets
- Visualizing ConvNet codes as embeddings
This library is licensed under the Apache 2.0 License.