AI Training - Tutorial - Use tensorboard inside a job

Wissensdatenbanken

KB0048551

AI Training - Tutorial - Use tensorboard inside a job


Icons/System/eye-open Created with Sketch. 219 Ansichten 01.09.2022 AI Training

Objective

The purpose of this tutorial is to show how it is possible to launch a TensorBoard with AI Training.

TensorBoard is a tool made by TensorFlow, for providing the measurements and visualizations needed during the machine learning workflow. It enables tracking experiment metrics like loss and accuracy, visualizing the model graph, projecting embeddings to a lower dimensional space, and much more.

TensorBoard provides a visual interface:

image

The tutorial presents a simple example of launching TensorBoard in a job.

Requirements

Instructions

Have an object store container where your metric logs are saved

First of all, you must have trained your model and saved your results in an object store container (exemple: my_tf_metrics located in Gravelines GRA).

Alternatively, you can have a job already RUNNING that is connected to that object store container and is writing metric logs inside it (example: my_tf_metrics@GRA:/runs:RW:cache). In that last case, don't forget the cache parameter indicating that the volume is cached and sharable among jobs. More information about volumes configuration in jobs can be found here, information about volume caching can be found here.

If you want to see an example of how to use TensorBoard to train a model, please refer to this notebook on GitHub.

Launch TensorBoard in a job

To launch TensorBoard in a job, you need to access the ovhai CLI and run this command:

ovhai job run tensorflow/tensorflow \
    --cpu 1 \
    --default-http-port 6006 \
    --volume my_tf_metrics@GRA:/runs:RO:cache \
    -- tensorboard --logdir=/runs --bind_all

First, set the number of CPUs. For this type of job you don't necessarily need a lot of resources.

--cpu 1 indicates that you request 1 CPU for that job.

The default port for TensorBoard is 6006.

--default-http-port 6006 indicates that the port to reach on the job url is the 6006.

Connect the volume containing your tensorboard metric logs.

--volume my_tf_metrics@GRA:/runs:RO:cache indicates that you are connecting the container my_tf_metrics from Gravelines (GRA) Object Store into the /runs directory of your job. The read only RO permission is enough because TensorBoard does not need access to write. The container my_tf_metrics@GRA should contain your tensorflow metrics.

Specify the tensorboard launch command.

tensorboard --logdir=/runs --bind_all indicates that we want tensoboard to be watching over the /runs directory. Don't forget the --bind_all parameter or you won't be able to access your tensorboard from the public network.

Consider adding the --unsecure-http attribute if you want your application to be reachable without any authentication.

Once the job is running you can access your TensorBoard directly from the job's url.

Go further

You can compare AI models based on resource consumption, accuracy and training time. Refer to this tutorial.

If you need training or technical assistance to implement our solutions, contact your sales representative or click on this link to get a quote and ask our Professional Services experts for a custom analysis of your project.

Feedback

Please send us your questions, feedback and suggestions to improve the service: