Advanced Usage

Advanced Parameters during Training

Note that you can change how long the model trains for by tweaking either the num_epochs or the max_iterations parameters during training, passed into turicreate.drawing_classifier.create. Here are a few optional parameters that you can pass to .create that can change how you train models:

• batch size is the number of training examples in one forward/backward pass. The larger the batch size, the more memory you would need during training.
• num_epochs is the number of forward/backward passes of all the training examples.
• max_iterations is the number of passes, each pass using a batch_size number of examples.

If you specify both max_iterations and num_epochs, max_iterations would take precedence and the model would train for that number of iterations and the provided value of num_epochs would be ignored.

When training a model it may be useful to look at the progress table in the output to see if the model has converged. This will help determine whether max_iterations needs to be increased to allow model to converge to an optimal solution.

Warm Start

To boost the accuracy of the Drawing Classifier models you train, and to help those models converge faster, we provide the option of loading in a pretrained model for a warm start.

A cold start to training would be when the weights are randomly initialized. A warm start is when the weights in the neural network are loaded from those of an already trained model. This improves the initial values of the weights in the network to something better than random, thereby improving accuracy and also helping the model converge faster.

We have published a pre-trained model that automatically gets downloaded when you pass in auto to the warm_start parameter. This model is trained on millions of drawings from the “Quick,Draw!” dataset.