After sending data to the engine, you need to set up the training parameters to start a training experiment. Each training experiment will produce a specific AI model based on the chosen parameters. This means that a project can have multiple models of the same dataset, depending on the number of experiments.
Start a new experiment.
On the New Experiment tab:
Step 1: Select the Model Type
- FAST will reduce model accuracy in exchange for inference speed
- ACCURATE is the balance between inference speed and accuracy (standard option)
- MOST ACCURATE will reduce inference speed in exchange for model accuracy
- CUSTOM will allow users to add more model types for their training. (in development)
Choosing the suitable model type will depend on your application's requirements, but the accurate and most accurate type will take longer for the training process unless you have a robust GPU.
Step 2: Select other parameters
- Batch size: This will divide the training dataset into smaller groups to reduce GPU load. The more powerful GPU, the higher batch size you can choose. ODHUB will auto-recommend batch size based on your GPU power.
- Number of steps: During each step, an amount of batch size is processed. The more number of steps, the better the trained model is. The general guideline is to use 2000 steps for each object class. For example, if we have two categories, such as cats and dogs, the total number of steps would be 4000. However, the ODHUB training algorithm includes an early stop mechanism. This means that the algorithm will stop running if it doesn't find any further improvement. Therefore, it's acceptable to set a large value for the number of steps, as the algorithm may stop before reaching the maximum number of steps.
Step 3: Select the learning rate
The default value is recommended.
Model Training
Once the parameter is set, click Start Training to begin the model training process.
- The loss value will be updated constantly on the training graph.
- The mAP value will be updated on the test graph after each 1000 trained steps. mAP stands for mean Average Precision, ranging from 0% to 100%. It's an indicator that defines whether the trained model is good enough; the closer to 100%, the better the model is.
Review Previous Experiments
Once any experiment completes its training process, its training log will be stored in the Previous Experiments tab.
Use the drop-down experiments list to select your experiment (if you have many). From the selected experiment, you can:
- Review its training history, including the model type, trained steps, batch size, current loss and mAP; OR
- Continue to train that experiment to further improve the loss and accuracy value (you can only modify the learning rate or add extra steps of a trained experiment. To modify other parameter, you need to create a new experiment)
Next Action
Once you satisfy with the training result, click Move On to Evaluate & Export to go to next phase.