Using TensorBoard

Using TensorBoard

To use TensorBoard with the Cerebras software container, follow these steps:

  1. Open a new Terminal window.

  2. Log in to any compute node in the CS system cluster.

    ssh [username]@worker1

  3. Run Cerebras Singularity container in interactive mode.

    singularity shell /data/shared/software/singularity/cbcore_latest.sif

  4. Run TensorBoard command in the Singularity shell.

    tensorboard --bind_all --logdir=[PATH_TO_MODEL_DIR]

  5. Open the TensorBoard-generated link in your local browser. Depending on how your internal environment is set up, you may need to use port forwarding to view the link locally.

Saving summaries

To save summaries, set save_summary_steps in CSRunConfig. See an example in the CSRunConfig section.

Writing summary data

Use summary ops to write the summary data as you would in TensorFlow, by using the tf.summary module in your model function.

In addition, to enable the summaries, you must wrap your Estimator call with the following:

from cerebras.tf.summary import cs1_enable_summaries
with cs1_enable_summaries():
   # estimator code