Train
Training teaches your machine learning model to predict the correct labels from your examples.
What is training?
What is training?
Training is how your model learns to predict the correct labels from your examples. You can think of your examples as a collection of flashcards. During training, your model will continually look through the flashcards and try to find similar patterns that help it guess the right answers.
Read more machine learning basics
Read more about labeling your examples
How do I start training?
How do I start training?
Lobe automatically starts training when your examples meet the minimum requirements. To start training, you need:
- Imported images to label as examples
- At least two labels
- At least five images per label
Lobe will also follow best practices to continue training when you make changes to your examples. If you make large changes or add/remove labels, Lobe will reset training and start training a new model.
After automatic training has completed, you can manually optimize your model to train for longer for better real-world performance (File > Optimize Model).
How long will training take?
How long will training take?
Training time is quite variable and depends on how hard it is to distinguish between the labels in your examples and how many examples you have. It can take anywhere from minutes to hours, and sometimes days for very large problems.
You can hover over the training progress to see a time estimate. This training time estimate is updated every few seconds based on your progress and computer’s current processing speed. You may see it fluctuate if you are performing other tasks on your computer as available CPU and memory change.
How do I train for longer?
How do I train for longer?
When training has completed, you can optimize your model by selecting File > Optimize Model. Optimizing a model performs additional training and can take much longer to complete, but will generally help find a better version of your model.
While optimizing, Lobe will keep training for as long as your model is improving and does not have a set end-time.
Can I change my model architecture?
Can I change my model architecture?
Lobe automatically picks the best model architecture for your problem; no setup or configuration needed. Project Settings (File > Project Settings) lets you switch your project between two modes:
- Accuracy: a larger model that will generally achieve higher accuracy on harder problems, but will have longer prediction times and use more memory.
- Speed: a smaller model that will have faster prediction speed and smaller memory usage, but may have lower accuracy. This model can also be optimized for edge devices such as mobile phones or the Raspberry Pi.
Changing your project will reset any training completed so far and automatically train a new model.
What models are used?
What models are used?
Lobe uses two popular models for image classification:
- Accuracy: ResNet-50 V2
- Speed: MobileNetV2
Both models utilize transfer learning with pretrained weights from the ImageNet dataset. Transfer learning lets you train better models with less data and gives a better starting point for training on larger data.
Is Lobe performing data augmentation?
Is Lobe performing data augmentation?
When loading images for training, Lobe will crop to the center square and scale your image. Lobe will automatically create small variations of your images to reflect the noisiness of real-world data. During training, Lobe will make five variations of your images with randomly varied:
- Brightness
- Contrast
- Saturation
- Hue
- Rotation
- Zoom
- JPEG encoding noise