Skip to main content

Training Your Own Models

Whether you're detecting objects, classifying scenes, or identifying unique patterns in your videos, Vidsy.ai lets you train your own models from scratch or on top of pre-trained ones.

This guide walks you through the entire process — from setting up an index to reviewing training results.

1. Create a New Project​

  1. Launch Vidsy.ai.
  2. Open the Home tab and click on New to create a new project
  3. Open the Engine tab from the top menu.
  4. Click the Setup button.
  5. In the left column, click "Add new..." to create a new index.
tip

An index is a database that stores the data and configurations for a single model. You can create multiple indexes.

  1. Fill in the required fields:
    • Name the index.
    • Set it as primary.
    • Add a description.
tip

The primary model is used as the entry point by the analysis engine. There can only be 1 entry point per project.

screenshot of the Engine tab and Setup screen

2. Configure the Model Pipeline​

Next, set up the model’s purpose and behavior.

  1. Choose the model purpose:

    • Locate: Find object positions.
    • Classify: Label the entire image.
    • Identify: Recognize specific instances of objects.

  2. Adjust the slider to balance between:

    • Performance (speed, size)
    • Accuracy

  3. Enable detection options:

    • Bounding boxes
    • Keypoints

  4. Choose a starting model:

    • Pretrained (COCO 2017)
    • Blank model

  5. If using classification:

    • Choose between single-label or multi-object detection.

  6. Click Save to finish setting up the index and model.

screenshot of model pipeline configuration

3. Import Your Dataset​

Vidsy supports the following dataset formats:

  • COCO (JSON)
  • Labeled folders (one folder per class)
  • Previous Vidsy exports

To import:

  • Open your index
  • Click Import Dataset or drag and drop your dataset

screenshot of importing a dataset

4. Open the Trainer View​

  1. Click the Chart button chart-icon in the interface.
  2. This switches the view from index content to Trainer View — a dedicated space for visualizing training metrics.
info

In Trainer View, you’ll see charts for accuracy, loss, learning rate, and more.

screenshot of the Trainer View with charts

5. Configure and Start Training​

  1. Click the Play button ▶️ to open training options.

  2. Choose a training strategy:

    • Fine-tune the full model
    • Train only the top layer

  3. Training parameters (prefilled with defaults):

    • Batch size
    • Minimum epochs
    • Optimizer, learning rate, momentum, decay

  4. Prepare dataset (first time only):

    • Vidsy will organize training and validation sets.

  5. Optional:

    • Clip and align keypoints (if available)
    • Enable data augmentations:
      • Rotate, scale, flip, brightness, contrast, saturation, hue, translateX/Y
      • You can also set how many combinations per image are allowed

  6. Click Start to begin training.

screenshot of training configuration screen

6. Monitor Training Progress​

During training, you'll see:

  • Accuracy/loss over time
  • Validation results (data not seen during training)
  • Live feedback in charts

⏱️ A dataset of ~600 images over 50 epochs typically trains in a few minutes on an NVIDIA 3070 GPU.

screenshot of training charts

7. Save and Use Your Model​

Once training finishes:

  • Choose Save to keep the model.
  • Or discard it if you’re not satisfied.

Saved models are automatically available inside the Vidsy.ai application for:

  • Analyzing new videos
  • Real-time camera feed processing
  • Future fine-tuning

Models stay within Vidsy — there’s no need to export them.

Tips​

  • Use the Trainer View to understand training performance.
  • You can retrain the model later with better data or more epochs.
  • If you annotate incorrect predictions during use, Vidsy can incorporate them into retraining.