Tensorboard

Introduction

Tensorboard is a Tensorflow extension to conveniently compare the training performance of different neural networks, with respect to their loss and metrics. You can specify a metrics to measure the performance in config_model_parameters.yml, then use the instructions below to create the training logs and browse them.

On GPU:

python steer_analysis.py

It will create logs/{train date}_{train time} directory automatically.

To see the outputs from TB:

tensorboard --logdir logs

This might be slow if one uses ssh connection.

In this case, one can create tar.gz file of the directory by tar -zcvf command and download it to your personal laptop.

On your PC:

After downloading:

tar -zxvf {file_name}
tensorboard --logdir {directory_name}

Copy and paste to your web browser the TensorBoard link provided by the last command. GUI will appear and one can play with it.

Running TensorBoard directly from remote server over SSH:

On your PC open a new terminal and type:

ssh -N -f -L localhost:{local_port_number}:localhost:{remote_port_number} {user@remote_machine}

In a separate terminal session connect to aliceml. cd to TPCwithDNN/tpcwithdnn and type:

tensorboard --logdir logs --port {remote_port_number}

Finally, on your PC open a web browser and type the following address in the search bar:

http://localhost:local_port_number/