Object Detection (YOLOv4)

YOLOv4 Training Parameters and Value Ranges

Table 1 describes the names, types, value ranges, default values, and descriptions of the YOLOv4 training parameters.

**Table 1** YOLOv4 training parameters (**yolov4_mindspore**/**model_train.py**)
Parameter	Type	Value Range	Default Value	Description
--train_dataset_path	String	-	None	Path of the training dataset.
--pretrained_ckpt_path	String	-	"pre_trained_ckpt"	Path of the pre-trained model.
--train_output_path	String	-	"train_output_path"	Output path of the training result.
--device_id	Integer	[0,7]	0	ID of the NPU used for training.
--device_num	Integer	1,2,4,8	1	Number of NPUs used for training.
--rank_table_file	String	-	None	Used for multi-device training.
--input_width	Integer	A multiple of 32 within [192,1920]	1472	Width of the model input image.
--input_height	Integer	A multiple of 32 within [192,1920]	1472	Height of the model input image.
--split_ratio	Float	(0,1)	0.8	Ratio of the training dataset to the validation dataset.
--per_batch_size	Integer	[1, 3]	2	Number of image batches for training.
--lr	Float	(0,1)	0.0005	Training learning rate.
--t_max	Integer	[1,10000]	100	T-max learning rate parameter.
--max_epoch	Integer	[1,10000]	100	Number of training epochs.
--run_eval	Bool	True or False	True	Whether to perform evaluation during training.
--eval_start_epoch	Integer	[1,10000]	40	Sequence number of epochs from which evaluation is performed during training.
--auto_set_eval_interval	Bool	True or False	False	Whether to enable automatic evaluation. If this parameter is enabled, --eval_start_epoch and --eval_interval are invalid and do not need to be configured.
--eval_detail_info	Bool	True or False	False	Whether to perform detailed evaluation.
--eval_interval	Integer	[1,10000]	1	Evaluation interval. Retain the default value.
--min_score	Float	(0,1)	0.2	Minimum score for evaluation.
--nms_threshold	Float	(0,1)	0.3	NMS threshold for evaluation.
--max_box_num	Integer	[1,1000]	90	Maximum number of evaluation bounding boxes.
--multi_scale_input	Bool	True or False	False	Whether to enable multi-scale training.
--mosaic	Bool	True or False	False	Whether to enable data augmentation.
--self_fit	Bool	True or False	False	Adaptively adjust the model width and height based on the dataset analysis result.
--self_fit_prior_choice	String	'enlarge_image', 'shrink_image', 'enlarge_aspect_ratio', or 'shrink_aspect_ratio'	""	Priority policy for adaptively adjusting the model width and height. Only one of the four policies is supported and takes effect only when self_fit detects multiple scenarios.
--dump_preprocessed_image	Bool	True or False	False	Save the preprocessed images to MindInsight.
--reg_loss	String	"ciou" or "giou"	"ciou"	Loss calculation method.
--enable_modelarts	Bool	True or False	False	Whether to use ModelArts for training.
--data_url	String	-	None	Dataset path for the ModelArts training.
--train_output_url	String	-	None	Output path for the ModelArts training.
--checkpoint_url	String	-	None	Path of the pre-trained model for the ModelArts training.
--split_image	Bool	True or False	False	Manually enable the image split function for small object detection. This parameter is valid only when self_fit is set to False.
--split_resize_ratio	String	Range of the two numbers in a value: [0.1,10].	"1,1"	Scaling ratio of the original image. The two numbers in the value indicate the width and height.
--split_block_size	String	Range of the two numbers in a value: [192,1920]	"1472,1472"	Size of image blocks. The two numbers in the value indicate the width and height. The width and height of an image block must be less than the width and height of the original image multiplied by the zoom-in ratio.
--split_overlap_size	String	Range of the two numbers in a value: [96, minimum value of split_block_size - 96]	"256,256"	Size of the overlapping area of image blocks. The two numbers in the value indicate the width and height.
--force_fp16	Bool	True or False	False	Whether to use the fp16 precision mode when exporting the OM model.
Note: The original SSD small object detection parameter split_auto will be discarded in later versions.

Training Command Reference for YOLOv4

Run the following command to start the YOLOv4 model training:

python3 model_train.py --train_dataset_path=/path_to/images --pretrained_ckpt_path=/path_to/ckpt --train_output_path=./outputs/ --device_num=1 --device_id=0 --per_batch_size=2 --input_width=1472 --input_height=1472 --max_epoch=20 --lr=0.0005 --min_score=0.01 --nms_threshold=0.5 --eval_start_epoch=20 --run_eval=True --split_ratio=0.8

The model training process is random. Use the evaluated accuracy. When the preceding parameters are used, you can view the training accuracy results of 10 images and 20 epochs in Figure 1.

Figure 1 Accuracy results of the YOLOv4 model training

Figure 2 shows the log information after the model training is complete.

Figure 2 Log information of the YOLOv4 model training

After the training is complete, the .ckpt, .a310.om, .a310p.om, and .air model files are generated in the output directory specified by the --train_output_path parameter.

YOLOv4 Evaluation Parameters and Value Ranges

The following table describes the names, types, value ranges, default values, and descriptions of the YOLOv4 evaluation parameters.

**Table 2** YOLOv4 model evaluation parameters (**yolov4_mindspore**/**model_eval.py**)
Parameter	Type	Value Range	Default Value	Description
--eval_dataset_path	String	-	None	Path of the evaluation dataset.
--eval_ckpt_path	String	-	"pre_trained_ckpt"	Path of the trained model.
--eval_output_path	String	-	"eval_output_path"	Output path of the evaluation results.
--device_id	Integer	[0,7]	0	ID of the NPU used for evaluation.
--per_batch_size	Integer	The value can only be 1.	1	Number of image batches for evaluation.
--input_width	Integer	[192,1920]	Consistent with the value listed in the training parameter table	Width of the model input image.
--input_height	Integer	[192,1920]	Consistent with the value listed in the training parameter table	Height of the model input image.
--min_score	Float	(0,1) or -1	0.2	Minimum score for evaluation.
--nms_threshold	Float	(0,1)	0.6	NMS threshold for evaluation.
--max_box_num	Integer	[0,1000]	90	Maximum number of evaluation bounding boxes.
--eval_detail_info	Bool	True or False	True	Whether to perform detailed evaluation.
--enable_modelarts	Bool	True or False	False	Whether to use ModelArts for training.
--data_url	String	-	None	Dataset path for the ModelArts training.
--eval_output_url	String	-	None	Output path for the ModelArts training.
--checkpoint_url	String	-	None	Path of the pre-trained model for the ModelArts training.

Evaluation Command Reference for YOLOv4

The command for evaluating the YOLOv4 model is as follows:

python3 model_eval.py --eval_dataset_path=path_to/images --eval_ckpt_path=path_to/your_ckpt --eval_output_path=./outputs_eval/ --min_score=0.001 --nms_threshold=0.5

Use the checkpoint generated during the training to evaluate the model accuracy, as shown in Figure 3.

Figure 3 YOLOv4 model evaluation results

The folders and files shown in Figure 4 are generated in the evaluation directory. The ok_images, ng_fp_images, and ng_fn_images folders store the image evaluation results, and the statistics.csv file stores the corresponding accuracy results.

Figure 4 YOLOv4 model evaluation directory

New feature in version 3.0: Image-level accuracy evaluation metrics are added to the statistics.csv file.

Figure 5 Image-level accuracy evaluation metrics

Parent topic: Training Component