The training component contains a TensorBoard visualization (TensorBoard viewer), which makes is comfortable to open the TensorBoard session for training jobs.
The logs of the training are uploaded to a Google Cloud Storage Bucket. TensorBoard automatically references this log location and displays the corresponding data.
To open TensorBoard click on the training component in your experiment run. Located on the ride side is the artifact windows which shows a very handy button called (Open TensorBoard).
In order to use his visualizations, your pipeline component must write a JSON file. Kubeflow provides a good documenation on how visualizations are working and what types are available.
# write out TensorBoard viewer
metadata = {
'outputs' : [{
'type': 'tensorboard',
'source': args.input_job_dir,
}]
}
with open('/mlpipeline-ui-metadata.json', 'w') as f:
json.dump(metadata, f)