The purpose of reference model training is to generate a model that can be used for auto labelling for a specific application. DeepView Enterprise provides YoloV8 for training the reference model.
To train the following steps are required:
Prepare Dataset for Training
Training YoloV8 required a Training Set, a Validation set and Optionally a Test set. These datasets can be independently imported. To generate these from a single dataset, the split dataset option can also be used.
- To split a dataset into Train, Validate and Test datasets create two datasets (or three of test is required). Name them appropriately, for example, Train, Val and Test.
- Add an annotation set in teach one of them by the name (say) GT, for ground truth.
- Click on the three dots on the Dataset card (which you want to split) and select Copy dataset.
- In the Copy Dataset dialog select the source Project, source dataset and source annotation set.
- Then click on the Destination Dataset (+) to add another destination dataset (Val).
- Repeat Destination Dataset (+) to add another destination dataset Test (Test).
- Now should have three datasets under destination datasets.
- Select Train Dataset for the first one and select GT as the annotation set.
- Select Val Dataset for the first one and select GT as the annotation set.
- Select Test Dataset for the first one and select GT as the annotation set.
- Select the percentage for each destination datasets - The percentages may add up to less than 100 but not above 100.
- Then click "apply" to populate datasets.
Create Training Set / Cache
Training set is a cache of datasets in the format that can be used by the training to train the model.
To create Training Set, go to the training section from the App Selection Icon (top left screen).
The top left icon is for creating the Training set and the second icon for creating an experiment and its sessions.
Click on the + New Training Set.
Enter the name and description.
Select the Training, Validation and Test datasets and the annotation to use for each dataset.
Now create the Training Cache.
On the Training Set Card and click on + besides Data Set Cache and select Darknet.
This will generate the Darknet Cache.
Generate Experiment and Sessions
Click the second Icon on the training screen.
Click on +Create New Experiment.
Enter name and description to create a new experiment.
Now one or more training sessions can be created in the experiment.
Click on three dots to add a session.
Click on the show Training Graphs to load standard Tensor Board graphs.