Introduction to alwaysAI Model Training

For information on model training in general, please see our introductory article.

The alwaysAI model catalog has a wide range of models, and great diversity. We offer multiple model types, e.g. image classification, object detection, etc. as well as a mix of frameworks, architectures, and networks, e.g. caffe, tensorflow, YOLO models, SqueezeNet, GoogLeNet, and a large list of labels from various datasets, e.g. COCO, ImageNet, etc. However, if you want a model that can detect your specific cat, or your child’s favorite toy, we don’t have that… yet. This guide provides an overview of model training with alwaysAI.

Overview

With alwaysAI’s Model Training Toolkit, you can train an object detection model to recognize any object you choose. All you need to do is generate data, annotate that data, and train that data using the model training tools built into the alwaysAI CLI. Once your model is trained, you can test it by deploying it locally or on an edge device, or upload it to the alwaysAI model catalog. If you feel it needs more training or data, you can either add to your dataset, continue training from where you left off, or both.

Specifications

Our initial model training tool uses TensorFlow 1.14 as the framework for training an object detection model. We train by transfer-learning from a MobileNet-SSD that has been trained on the COCO Dataset. We offer both CPU and GPU versions. See the set up guide to configure your system for model training.

Data Generation

To support data generation we have created a data generation starter app that allows you to capture video directly from your edge device. This ensures that your dataset will come from the same hardware you will use for computer vision tasks. For additional information on data collection, please see this guide.

Data Annotation

We integrate CVAT into the CLI, which streamlines the annotation process. For more information on data annotation, please see this guide.

Model Training

The model training tool is integrated into the CLI, and provides the ability to train an object detection model using tensorflow 1.14. To start training immediately, read our quickstart guide. You can reference the following documentation for more information on model training:

Exporting a Model

The training tool exports a model that is immediately compatible with the alwaysAI model catalog and which can be used in an alwaysAI starter app in order to gauge the performance in real-world settings. For more information on using models trained using the alwaysAI model training tool, please see here.