modeltrack 1.0.2

PyTorch model config and result tracker

Tool to track all PyTorch model config and architecture and prepare training summary reports

Usage Guide

The following will explain how to install the tool to your environment, and use the tool within a Jupyter Notebook

Installation

To install the tool simply run the following command in your terminal:

pip install --upgrade modeltest

Quick Start

To begin using the tool you must import the package into your notebook by adding the following:

import modeltrack.experiment as exp

After importing the experiment module you will be able to create a new tracker for your model:

tracker = exp.ModelTracker('model-name', config={"max_epochs":100})

When instantiating the new ModelTracker, you must pass in the name of the experiment being run as well as the model configuration. Optionally, you can pass a directory to root_dir to specify where the tracking output should be stored. The following configuration variables must be set, along with any other model-specific configuration:

config = {
   "batch_size": [INT],
   "learning_rate": [FLOAT],
   "max_epochs": [INT],
   "overwrite": [BOOL] - if set to True, most recent experiment with that name will be overwritten
}

The ModelTracker object has four functions that are useful during training:

• tracker.start_training():

  Signal to the ModelTracker that a new training session has begun. Re-initializes parameter watchers

• tracker.save_epoch_stats(train_loss, test_loss, train_acc, test_acc):

  Store the epoch statistics to be displayed and analyzed in output, and automatically log values

  train_loss:

  training loss of single epoch

  test_loss:

  training accuracy of single

  train_acc:

  testing/validation loss of single epoch

  test_acc:

  testing/validation accuracy of single epoch

• tracker.save_model(model, epoch, optimizer, loss):

  Save the state of the model in a checkpoint file

  model:

  nn.Module object

  epoch:

  current epoch count

  optimizer:

  torch optimizer

  loss:

  current validation loss

• tracker.finish_training(model=None):

  Save the training parameters for review and produce training report

  model:

  [Default=None] current nn.Module model being used at end of training

Examples

Please see demo_model.ipynb to see how the tool is used in a Jupyer Notebook or sample.py to see the tool used in a python script