SmartML: Train Vision AI Models with One Click

Training models is one component of the computer vision process that typically requires investing considerable time and resources into complicated work. Our user-friendly model training framework changes that, reducing much of this hard work and even enabling one-click training. 

What Is SmartML?

SmartML is Plainsight’s proprietary model training toolset. It is at the heart of our promise to make vision AI’s impressive capabilities more accessible to users of all experience levels and enterprises of all types. Thanks to SmartML, non-technical platform users can use their labeled datasets to train vision AI models with built-in intelligence without the need for a single line of code. For more technical users, SmartML also offers customizable settings for advanced configuration.  

SmartML for Plainsight On-Demand and Plainsight Enterprise

Plainsight customers can take advantage of SmartML at whatever level of scale and customization they need. In addition to signing up for Plainsight On-Demand to independently train and deploy models, enterprises can collaborate directly with Plainsight’s team to address their challenges. Armed with SmartML, our experts can quickly introduce high-impact vision AI models and transform operations for customers in a range of industries.

    • One-Click Training: Plainsight On-Demand customers can train custom models with a single click and little hands-on coding experience.
    • Resource Efficiency: Features for automated model training and performance tracking make it easy to optimize resource allocation and avoid unnecessary effort or spending.
    • Pre-Built Models: We offer a range of pre-built vision AI models to help kickstart computer vision projects intended to improve key processes and help businesses overcome persistent challenges.
    • Custom Models: On-Demand users can build custom models quickly with free, unlimited labels and powerful tools like AutoLabel. Enterprise customers can trust our team to develop and deploy customized models.

SmartML Hyperparameters 

For experienced machine learning “power users with clear goals in mind, our platform provides a number of optional configurations for model performance. These hyperparameters directly impact how models behave so it’s important to understand them well and adjust them with care.  

Split Dataset

It’s a common best practice to train deep learning and computer vision models by splitting datasets into three smaller groups: training, validation, and test datasets. Training datasets include the data a model learns from, validation datasets help users assess how well their model is performing, and test datasets are used last to verify that models will work with data they’ve never seen. The platform offers users the ability to customize the percentage breakdown of their three sub-datasets and select whether or not to randomize their data. 


The SmartML toolset offers three different backbone options. They are, from smallest to largest, R50 FPN 3X, R101 FPN 3X, and X101 FPN 3X. Selecting the larger option typically means more accuracy, but will also lead to longer inference times. X101 FPN 3X, for example, takes more than twice as long for inference as the default backbone. Differences in accuracy are more difficult to quantify. 

Learning Rate

Learning rate establishes the step size for each iteration during the training process. It effectively refers to the speed at which models learn. SmartML’s default learning rate is 0.0001.

Batch Size 

Batch size refers to the number of training examples (in this case, pieces of visual data) used in each iteration. SmartML’s default batch size setting is 16 and the recommended batch size is between four and 16. Batch size cannot exceed the number of images included in a dataset. 


The tiling process divides an image into a number of smaller “tiles,” inferences on each tile, and then reassembles each image. Users can customize tiling hyperparameters related to factors like tile size and tile stride (the amount of overlap between tiles). 

Start Training Models Today

On-Demand users can get started by signing up now. In addition to free, unlimited labeling, everyone who creates an On-Demand account receives a $100 credit to leverage pay-as-you-go features for use in training models with SmartML and deploying them to support your business. For enterprises, please schedule a demo with our team to discuss your needs and the ways we can help you transform operations by developing use case-specific applications. We’ll provide the vision AI services and resources you need for creating, training, and deploying custom vision AI models to see value in days, not months.

More Plainsight Blog Posts:

5 Ways Agribusinesses Can Prevent Recalls, Shutdowns, and Delays with Vision AI

5 Ways Agribusinesses Can Prevent Recalls, Shutdowns, and Delays with Vision AI

Investments in computer vision technology can help agribusinesses and food manufacturers of all types spot signs of trouble early and stop costly, potentially deadly recalls before they happen. Deployed across the production and manufacturing cycles, these models can detect hazards ranging from contaminants and foreign objects to defective equipment and non-compliant behavior. Organizations capture hundreds of thousands of hours of visual data in the form of video footage and imagery every day, and computer vision allows these businesses to put this data to work for process transformation.

How Can Vision AI Predict and Prevent Supply Chain Disruptions?

How Can Vision AI Predict and Prevent Supply Chain Disruptions?

Introducing computer vision across the supply chain can help manufacturers predict and prevent the types of conditions that lead to these kinds of disruptions. From potential contaminants and foreign objects to defective products and packaging to unsafe or unsanitary behavior, computer vision is the key to recognizing supply chain obstacles early and stopping shortages in their tracks. AI solutions could prove especially useful in volatile, high-production periods where errors and disruptions are both especially likely and especially costly.