Skip to content

YOLO Model Training

This section describes how to train the digital meter model on the COCO digital meter datasets. The implementations of yolo digital meter detection model is based on the Swfit-YOLO and power by mmyolo

Prepare Datasets

SSCMA uses Digital Meter Datasets by default to train the Swfit-YOLO model, please refer to the following steps to complete the preparation of datasets.

  1. Download digital meter datasets with COCO datasets mode

  2. Remember its folder path (e.g. datasets\digital_meter) of the unpacked datasets, you may need to use this folder path later.

Choose a Configuration

We will choose a appropriate configuration file depending on the type of training task we need to perform, which we have already introduced in Config, for a brief description of the functions, structure, and principles of the configuration file.

For the Swfit-YOLO model example, we use as the configuration file, which is located in the folder under the SSCMA root directory configs/swift_yolo and its additionally inherits the configuration file.

For beginners, we recommend to pay attention to the data_root and epochs parameters in this configuration file at first.
_base_ = ["./"]

anchors = [
    [(10, 13), (16, 30), (33, 23)],  # P3/8
    [(30, 61), (62, 45), (59, 119)],  # P4/16
    [(116, 90), (156, 198), (373, 326)]  # P5/32
num_classes = 11
deepen_factor = 0.33
widen_factor = 0.15

strides = [8, 16, 32]

model = dict(
            in_channels=[256, 512, 1024],

Training Model

Training the model requires using our previously configured SSCMA working environment, if you follow our Installation guide using Conda to install SSCMA in a virtual environment named sscma, please first make sure that you are currently in the virtual environment.

Then, in the SSCMA project root directory, we execute the following command to train a Swfit-YOLO digital meter detection model.

python3 tools/ \
    configs/swift_yolo/ \
    --cfg-options \
        data_root='datasets/digital_meter' \

During training, the model weights and related log information are saved to the path work_dirs/swift_yolo_tiny_1xb16_300e_coco by default, and you can use tools such as TensorBoard to monitor for training.

tensorboard --logdir work_dirs/swift_yolo_tiny_1xb16_300e_coco

After the training is completed, the path of the latest Swfit-YOLO model weights file is saved in the work_dirs/swift_yolo_tiny_1xb16_300e_coco/last_checkpoint file. Please take care of the path of the weight file, as it is needed when converting the model to other formats.


If you have a virtual environment configured but not activated, you can activate it with the following command.

conda activate sscma

Testing and Evaluation


After have finished training the Swfit-YOLO model, you can specify specific weights and test the model using the following command.

python3 tools/ \
    configs/swift_yolo/ \
    "$(cat work_dirs/swift_yolo_tiny_1xb16_300e_coco/last_checkpoint)" \
    --show \
    --cfg-options \


If you want a real-time preview while testing, you can append a parameter --show to the test command to show the predicted results. For more optional parameters, please refer to the source code tools/


In order to further test and evaluate the model on a realistic edge computing device, you need to export the model. In the process of exporting the model, SSCMA will do some optimization on the model, such as model pruning, distillation, etc. You can refer to the Export section to learn more about how to export models.


After exporting the model, you can deploy the model to the edge computing device for testing and evaluation. You can refer to the Deploy section to learn more about how to deploy models.

Released under the Apache 2.0 License