Train Your Model with SmartML™
Train a custom model with your dataset using SmartML

Overview

Start training computer vision models right from Plainsight using your labeled datasets. With our SmartML model training backend, you can easily set up a training run with a few clicks of your mouse, no coding required.
SmartML uses your labeled datasets to train an ML model

Configure a SmartML Training Run

To begin training a model, you will need a locked version of your dataset. See Versioning for more information.
After you've created and locked a dataset version, you'll be able to train a model using your labeled data.
From your dataset:
  1. 1.
    From the "Versions" tab of your dataset, click the "Train" link under Actions for the dataset version you wish to train
Train from the "Versions" tab
From elsewhere:
  1. 1.
    Navigate to Models from the side navigation, or click "Create New" and select "Model" from the menu
  2. 2.
    Click "Add Model" at the top right, or select an existing model and click "New Version"
  3. 3.
    If this is a new model, enter a name for your model.
  4. 4.
    Configure the dataset and model training options.
Create New Model to configure training

Model Training Options

SmartML requires some basic options to be set in order to begin your training run.

Dataset Name and Version

The dataset and version to be used as input for training your model. Only locked versions can be used. If you've entered the SmartML configuration from the "Versions" tab of your dataset, this will be pre-populated.
Select your dataset and version for your training run

Dataset Label for Model Output

Model Output Options - Select your desired model output. Your options will depend on the labels in your dataset. Options include: Bounding box, Polygon, Semantic Segmentation, Classification, Regression.
Labels for this Model Version Output - Select the labels you want to include in your training. A model can be trained on only one label type. If no labels are found for the selected label type, ensure that you have locked a version of your dataset that contains those labels.
Only Bounding box (rectangle labels), Polygon (instance or semantic segmentation), Classification, and Regression are currently supported in SmartML training. Other label types (Feature Points, Text, and MultiClassification) can be exported and trained outside of Plainsight.

Training Length

Training length, between 1-24 hrs (defaults to 2 hrs). This is the desired amount of time allocated to training your model. Larger, more complex datasets require longer training times, especially if tiling is used.
The desired training length will depend on the dataset. If you are unsure about how much training time is needed we suggest starting with 1 hour per a 1000 image dataset. SmartML will automatically end training when the model has stopped learning.
Read more about Training Time in the Model Training Results section.
Set the appropriate training length for your model

SmartML Hyperparameters

Optional parameters for advanced settings. Do not enable this section unless you are an advanced user and understand the appropriate settings to configure.
See our documentation on hyperparameters:

Starting a Training Run

Once you've completed selecting your options, scroll down and click "Save and Start Training"
Click to start your training
You will be taken to the Model Version Details where you can monitor the progress of your model.

Training States

When a model enters the training process, it goes through several states:
  1. 1.
    Starting - Training job is waiting to begin
  2. 2.
    Running - Training is actively running
  3. 3.
    Succeeded - Training is complete. This model can be deployed for inferencing
  4. 4.
    Failed - Training encountered an error and was stopped prematurely by the system
  5. 5.
    Canceled - Training run was stopped prematurely by the user
Your model will automatically stop training when it stops measurably improving or when the training reaches the budgeted time.
Your model version will be created and displayed at the top of your model's version list.

Canceling a Training Run

You can stop a training run of an actively training model. This option is only available when a model is in a "Starting" or "Running" state.
Click the "Stop Training" button from the Model Version Details screen:
or from the Model Versions list:

Training Results

When training has completed, you will see "Succeeded" as the status on the model version details.
A successfully trained object detection model.
Depending on your model type, you may see a "Loss Over Time" graph or a confusion matrix depicting your model performance.
Occasionally your training run may return with a "Failed" status and display a training error. Read more about training errors and how to troubleshoot them:

Model Training Time

The model training time is constrained to the selected training length that is specified in the Training Length model options.
When the model finishes training, the Training Time is listed in 2 parts:
Used time - The total amount of time the training process takes from beginning to end, including pre and post-processing (initializing a container). This time involves container setup and generating predictions on the dataset images. Unpredictable processing times may cause slight overages especially for larger datasets (10-15%).
Budgeted Time - this is the amount of time allocated for the actual training portion, as specified in your model training options. SmartML will automatically stop training when your model's validation loss stops decreasing, saving unnecessary training hours. Note: Training a model using Tiling can add up to 50% pre-processing time.
Actual time used may be a few minutes higher as there is some setup that occurs before the training process starts.

Predictions API

A successfully trained model can be used for inference. Start your Image API to begin running an Live API endpoint and use the Predictions API to run inference. See our guide to using the Predictions API:
You can set up your pub/sub credentials here to receive output from your model inference.
See our Google Pub/Sub guide for more information on how to configure
After training your model, you can view the detections on your dataset images and compare it to the labeled annotations. At the bottom of the model version details, you can preview the splits in Test Set, Validation Set, and Train Set.
  1. 1.
    Select the tab for test Set, Validation Set, or Train Set.
  2. 2.
    Click on an image thumbnail to open the image.
  3. 3.
    Under Display, select "Model Detection" to see the labels that the model detected. Select "Labeled Annotations" to see what was actually labeled.
Successfully trained models can be deployed for testing and inferencing. Read about how to deploy your model: