CORE - Using the Stats

How to Use

During the training session, some optimizations analysis are performed and stored as a readable file for Tensorboard. This file is automatically filled at each Network optimization step in the stats repository of the current working directory.

Tensorboard will be automatically launched for the current training session in a web browser. To open a previous training session analysis, launch Tensorboard with:

$ cd <working_session>/stats
$ tensorboard --logdir .
>> Serving TensorBoard on localhost; to expose to the network, use a proxy or pass --bind_all
   TensorBoard x.x.x at http://localhost:6006/ (Press CTRL+C to quit)

Then copy-paste the provided link into a web browser.

Adding fields

By default, six data fields will be stored:

• The value of the loss function at each batch / epoch

• The mean of the loss function at each batch / epoch over the last 50 elements

• The variance of the loss function at each batch / epoch

Custom data fields can be set in Tensorboard as well. After an optimization step, the Optimization class will by default return a dictionary with a single item named loss containing the current loss value. Then, the Trainer will add every item contained in this dictionary to the StatsManager. To fill this dictionary with custom fields, a custom Optimization class must be implemented where the transform_loss method will fill that dictionary (see dedicated section). This new field will have the same name as the item and take each value given in this item at each optimization step.

See following example:

# Import BaseOptimization
from DeepPhysX_Core.Network.BaseOptimization import BaseOptimization

# Create a new Optimization class
class MyOptimization(BaseOptimization):

    # Override the transform_loss method
    def transform_loss(self, data):
        return {'loss': self.loss_value     # Default loss value
                'new_field': ...}           # Additional field