The purpose of Model Training is to conduct neural network Model Training using the generated training data, and iteratively evaluate the model according to Set Parameters to obtain the available neural network model.

It has built-in open source frameworks such as Pytorch and PaddlePaddle, and trains Machine Learning or deep learning models with different Dataset categories. The overall Training Procedure obtains the network model with better Training Result through multiple epochs, and uses the pre-training model based on large-scale basic training data in the training model to reduce the training time. Improve Model Training efficiency and accuracy through hyperparameter tuning (such as Batchsize).

Function entrance

• Function Entry: Toolbox-> Machine Learning-> Video Analysis-> Model Training.

Parameter Description

• Training Data Path: Select the folder where the training data generated by Manage Picture Sample is located.
• Model Algorithm: Select the appropriate Model Algorithm according to the purpose of the model. Available: YOLOv5x, YOLOv5l, YOLOv5m, YOLOv5s, YOLOv5m, YOLOv7-E6E, YOLOv7-D6, YOLOv7-E6, YOLOv7-W6, YOLOv7-X, YOLOv7, YOLOv7-Tiny.
• Training Times: The number of Model Training times (epoch). The default value is 10. With the increase of Training Times, the fitting degree of the model is greater. Too many times may lead to over-fitting. Users can choose according to their needs. Training Times is proportional to the running time.
• Training Picture Size: Set the Picture Size for Model Training, 640 by default. The larger the picture is, the higher the detection precision is, but the recall will decrease. Therefore, the training Picture Size needs to be given according to the actual situation to improve the detection performance, and the Picture Size must be a multiple of 32.
• Single Step Operation Amount: The number of pictures in a single step operation in one training (Batch Size). The default value is 0. Within a reasonable range, the number of single-step operations is directly proportional to the memory (video memory) occupation and inversely proportional to the training time. When the Model Algorithm is YOLOv7, you need to specify the amount of single-step operation, and the calculation will not be automatically performed.
• Load Pretrained Model: If this option is checked, the training log path of the previous training can be selected as the path of the pre-training model, and the training will be performed on the basis of the previous training. Note: The size, category, and purpose of the training data are the same for the two training sessions.
• Model storage path: Set the storage path of the training result model.
• Model name: The name of the Set Model.
• Datasource: Sets the Datasource used to store property sheets and images.
• Property Sheet: Output the property sheet of the Model Training Result. In Model Training, a.csv file will be generated to record the model accuracy and other information of each training. After the training, the information will be converted into an attribute table and stored in Datasource. The generated.csv is shown in the following figure:



• Image: Output Model Training image. The Model Training Result will be represented in the form of a chart. The following figure shows the image information of the Output Result.

Related content

Video Analysis Environment Configuration

Model Convert

Time Recognition