- 21 Nov 2023
- 3 Minutes to read
- Updated on 21 Nov 2023
- 3 Minutes to read
LandingLens offers two methods for training models:
- Fast Training: This is the default option. LandingLens will automatically fine-tune its settings to optimize model training speed.
- Custom Training: Also called "Advanced Training". This is an advanced option recommended for users who are familiar with machine learning. Custom Training allows these users to have more granular control over pre-processing transforms and manual hyperparameter tuning. For more information, go to Custom Training.
Images Are Resized During Model Training
When you train your model, LandingLens resizes all images to make training faster and more effective. This resize does not affect the original images. This means that the original high-resolution images are always available in the Data Browser.
When you use Fast Training (the default Train option), all images are auto-resized according to your Project Type.
- Classification: Resized to 512x512px
- Object Detection: Rescaled with padding to 640x640px
- Segmentation: Resized to 800x800px
- Visual Prompting: Resized to 800x800px
- Anomaly Detection: Resized to 512x512px
To train a model:
- Open the project that has the images you want to train on.
- Click Train.(If you want to train a model based on a specific snapshot, learn how to do that here.)
This starts the Fast Training process. (If you're familiar with machine learning and would like to configure advanced settings, go to Custom Training.)
- LandingLens runs the model training process and saves the model.
Model Training Process
LandingLens performs a series of steps when it trains a model. This section describes those steps at a high level.
- Preparing the snapshot of your data: LandingLens saves all the images, labels, and data currently in your dataset.
- Provisioning GPU. LandingLens warms up a virtual computer in the cloud that includes a graphics processing unit (GPU). This virtual computer is where the model training occurs.
- Training & Learning. The platform sends the model's data and details to the virtual computer. This information tells the model how to train itself. Then, the model starts training, and a learning curve displays. This graph shows how well the model is training. If the line drops when it trains, the model is making fewer errors and is learning to be more accurate. The model will test its accuracy by comparing its Predictions to your Ground Truth.
- Calculating Model Performance. LandingLens calculates the metrics for the model, which show how well the model performed. For more information about these metrics, go to Analyze Models.
Default Model Name
When you train a model, LandingLens now automatically saves and names it. This ensures that you can access all versions of your data in the snapshots dashboard. Automatically saving the models also prevents models from being accidentally overwritten.
The default model naming convention is Model-mm-dd-yyyy_n, where n is the number of models in the project trained that day. For example, let’s say you train six models in a project on December 10, 2023. The default name of the sixth model will be Model-12-10-2023_6.
You can rename a model at any time.
You can end model training before it's completed by clicking End Training Now. For example, if you click Train and then realize you need to annotate more images, you can end the training process.
If you end model training early, LandingLens completes its current round (Epoch) of training and evaluates the images in your dataset with the model it had generated up to that point. The model may not be as accurate as it would be if training had completed.
To end model training:
- Click End Training Now.
- A confirmation pop-up appears. Click End Training to confirm.
- The Models panel shows this message: Manually Ended Early. The results for the model generated up to that point display.