Nilvana Vision Studio - Model Training
Please follow the instructions to get started model training with nilvana vision studio.
Preparation Before Model Training
Before starting model training, confirm that you have completed data annotation and generated the dataset version. To a certain extent, the quality of the model is related to the quality of the dataset; if it is not clear how to generate a specific dataset version, refer to the Nilvana Vision Studio - Dataset Versioning".
Start Model Training
To make effective use of the workstation resources, all training work will be included in the training work schedule. You can set up your own training by pressing the Training button on the specific dataset version or by going to the Training Plan List page.
The "Create Training Plan" dialog box has two options that affect the quality of training: the method of training and the ratio of data split. It is recommended to use the preset [small size/low accuracy] method for your first time. Models with high accuracy usually require hours to days of work, while models with low accuracy obtained quickly may be sufficient for identification. Model training is a form of supervised learning. The system will divide the data into the training set (previous exams) and testing set (simulated exams). The proportion of said split can be determined according to your personal preference. However, the training set obtained by the same split ratio will not be exactly the same every time as this is a process of random sampling. The ideal split ratio does not have too many testing sets. A good ratio would be about 10% to 20%, but if you have hundreds of thousands of data in your dataset version, you can lower the testing set ratio even further.
Check Model Training Status
The generation of a model requires hours to days of GPU computing time. Depending on the work schedule that we have carefully prepared for you, you can monitor the current training status of other items in the system, or you can rest assured that you can check the training results later. All the training work is properly arranged, and the system will automatically queue up for execution.
Based on the statistics of the dataset version, we preset the ideal maximum number of training terations for you. The training can be set without considerable parameter adjustment. However, if you find that the curve converges at the desired point during the training, you can interrupt the training at any time, without waiting for all iterations to be completed.
You can find models that have completed training in the history of the training plan list or in the model overview. The history of the training plan list can be seen by scrolling up the mouse wheel.
In the detailed information of the model, you can see the curve of parameters established by the model and the training process. You can quickly evaluate the effect of the model by simply uploading an image. The quality of the supervised learning model is closely related to the quality of the dataset. If you feel that the effect is not satisfactory, you can try to add meaningful training materials first and then experiment with different methods of training and split.
Once complete, the trained model is available via the Download button. You can serve this model on the Deep Detector suite on nilvana edge device or as a base model for machine annotations to assist you in further augmentation of your own dataset.
If you choose not to export the model, you can also choose to serve using this model directly in Vision Studio’s model serving area.