Train a model

Menu: AI /  Train a model

In this section, you will use the previously created samples to train your model.

1. Selecting Your Sample Folder
   Begin by selecting the folder where your samples, saved in .PLY format, are stored. Click [...] to browse and choose the folder. Then, enter a name for your new model.

2. Creating a New Model
   You can either:

   • Start from scratch – Define the features that will be used for training, such as geometry, intensity, or RGB.

   • Use an existing model – Load and refine a pre-trained model.

   Next, you must select a Profile for your data. There are four default profiles available, plus a Custom option if your data differs from the predefined profiles.

   Each profile includes specific parameters:

   • Subsampling Distance – The minimum feature size the model will recognize.

   • Neighborhood Radius – The size of the neighborhood sphere used as input for the model.

   If you choose the Custom profile, you must manually define these two parameters.

3. Splitting Your Data
   Your dataset must be divided into:

   • Training Set – Used to train the model.

   • Validation Set – Used to evaluate model accuracy after each epoch.

   ⚠️ Recommendation: For optimal results, the Validation Set should be at least one-third the size of the Training Set.
   Key Considerations:
   ✅ Ensure that the Validation Set is representative of the entire dataset (i.e., it should contain samples from all relevant categories and variations).
   ✅ Randomly shuffle data before splitting to prevent biases.
   ✅ For highly imbalanced datasets, consider stratified sampling to ensure each class is adequately represented in both sets.

4. Configuring Training Parameters
   Before starting the training process, adjust the following settings:

   • Maximum Number of Epochs – Defines how many times the algorithm will iterate over the sample set. More epochs improve learning but increase processing time.

   • Enable GPU Acceleration – If checked, the application will utilize an NVIDIA GPU instead of the CPU.

   • GPU ID – If your system has multiple GPUs, select the ID of the one you want to use.

   • Performance Mode – Determines the number of batches and workers based on GPU memory:

     • Minimal (GPUs with less than 12GB)

     • Balanced (GPUs with 12–16GB)

     • High (GPUs with 24GB or more)

5. Starting the Training
   Once everything is set, click [Create] to begin training. The trained model will be saved in the install folder and will be ready for Automatic Classification.

6. Monitoring Training Progress
   Once training starts, a progress window will appear, displaying:

   • A graph showing training mIoU (blue) and validation mIoU (orange).

   • A progress bar tracking the training process.

   Key Insights:

   • A well-trained model shows a smooth increase in both training and validation mIoUs.

   • A large gap between the two curves indicates overfitting, meaning more training and validation data is needed.

   • If validation mIoU does not improve for 25 consecutive epochs, training will stop automatically, and the best previous model will be saved.

   • Sharp fluctuations in the graph suggest insufficient data, affecting model stability.

By following these steps, you can efficiently train your AI model for LiDAR classification. 🚀