Summary

The BML intelligent vision module provides common image model training solutions. Major scenarios include image classification and object detection.

Create an Image Classification Model

Create an experiment

Click the “Create” button on the intelligent vision list page, enter the experiment name in the pop-up window.

After creation, you can view the creation experiment in the page of intelligent vision experiment list. Click "Experiment Name" or "View" to enter the page of experiment details:

The experiment process has been provided on the interface, and default parameters are available for the component. Usually, training can be completed only by changing data sets.

Modeling process

During the experiment, each component is equipped with the features of "Start Execution Here", "Execute Here" and "Execute the Node" to facilitate debugging.

1.Data selection: It supports selection of all data sets or filtration by multipart.
2.Data segmentation: It supports customization of segment percent of the training set and verification set and customization of the test set source.
3.Statistical node: It is mainly used for statistics of classified data size of the training set/test set/verification set which are used for training.
4.Image classification training: It supports adjustment of algorithm parameters and selection of basic models. For image classification training, only support the basic models of ResNet50, InceptionV4 and DPN131 at present.

Right click “Image classification training” to view the training process of real-time index monitoring model, for example:

5.Model evaluation: Right click “View an evaluation report” to check indexes such as the accuracy of outputting models during model training. Click “View a prediction result” to check the specific performance of each model on the test set, for example:

6.Filtration of prediction results: Filter the prediction results by confidence level, etc.:

7.Model selection: For multiple models generated by the training node, select the optimal model for subsequent use through the model selection component. The selection methods include: Automatic selection (by specifying the highest index in the test set or verification set) or customization. Model name needs to be manually filled in during customization.

8.Model conversion: It is used to convert the checkpoint generated during the modeling process into a model that can be used for prediction. You do not have to define parameters in this step.

9.Start and stop online prediction service: It is used to test whether the model can start the prediction service normally.

Create an Object Detection Model

Create an experiment

Click the “Create” button on the intelligent vision list page, enter the experiment name in the pop-up window, and select “Object detection” for the production line.

After creation, you can view the creation experiment in the page of intelligent vision experiment list. Click "Experiment Name" or "View" to enter the page of experiment details:

The experiment process has been provided on the interface, and default parameters are available for the component. Usually, training can be completed only by changing data sets.

Modeling process

The modeling process is similar to that of image classification, in which:

1. Data cleaning node: It supports to filter out untagged data during training.
2. Object detection training node: It supports adjustment of training parameters and automatic adjustment of batch_size and lr. Only FasterRCNN, RetinaNet and YOLOV3 are supported for the detection algorithm at present.
3. Model evaluation node: It supports to view the mAP index in the detection model.

Release the Model to the Model Warehouse

For experiments that have run successfully and with which data are not cleaned, it supports to release the model to the model warehouse. After release the model, further enable the online prediction service for use.