SmartML is Sense's own model training backend. SmartML enables a user to train a computer vision model using their labeled datasets without writing a single line of code. It takes a Sense labeled dataset and configuration as input and outputs model files, Tensorboard logs, and evaluation metrics. The model output and metrics from the training are available to download after the training process completes.
SmartML is used when you train a model in Sense. To get started with training a model from a labeled dataset, start here:
SmartML will produce a zip file containing the following:
model.pth A Detectron2 weights file.
model_config.yam A Detectron2 model config file.
config.json The SmartML configuration options used for training. See below for details on options.
metadata.json A JSON dictionary containing additional data about the model relevant to running in inference mode, including:
“classes”: A list of the model classes, in the order used by Detectron2.
“inference_resize_desc”: The inference resizing strategy used during evaluation, which should also be used when doing inference in production. See configuration options below for details.
“preprocessor”: The preprocessing augmenter used during training and evaluation, which should also be used when doing inference in production. This should be applied before resizing. See configuration options below for details on the augmenter description format.
“score_threshold”: The score threshold used for evaluation.
Tensorboard logs, potentially spread across multiple files with
tfevents in the name.
In order to get predictions from your model, you will need to use the SmartML Inference Service to return visualizations or detections.