[Testing] Comparison of Accuracy Performance (mAP) between Default Parameters and Scaled Parameters

[jaskiratsingh2000]

The yolov5s.yaml configuration file inside the ultralytics/yolov5 repo contains various scaling parameters. Hence I scaled the depth_multiple and width_multiple values while experimented the performance evaluation.

Default Values:

depth_multiple: 0.33
width_multiople: 0.50

Scaled Values:

depth_multiple: 0.01
width_multiople: 0.02

[jaskiratsingh2000]

I trained the Mask Wearing Dataset with the pre-trained --weights 'yolov5s.pt' --batch-size 10 --epochs 50

The time took to train the dataset with default parameters is 4.256 hours and to train the dataset with scaled parameters is 0.396 hours

[jaskiratsingh2000]

Performance Accuracy results with Default Parameters:

Fusing layers... 
Model Summary: 224 layers, 7266973 parameters, 0 gradients
val: Scanning 'valid/labels.cache' images and labels... 29 found, 0 missing
val: Scanning 'valid/labels.cache' images and labels... 29 found, 0 missing
               Class     Images     Labels          P          R     mAP@.5
                 all         29          0          0          0          0          0
Speed: 15.9ms pre-process, 3107.6ms inference, 6.8ms NMS per image at shape (32, 3, 640, 640)
Results saved to runs/test/exp

Performance Accuracy results with Scaled Parameters:

Fusing layers... 
Model Summary: 224 layers, 7266973 parameters, 0 gradients
val: Scanning 'valid/labels.cache' images and labels... 29 found, 0 missing
val: Scanning 'valid/labels.cache' images and labels... 29 found, 0 missing
val: Scanning 'valid/labels.cache' images and labels... 29 found, 0 missing
               Class     Images     Labels          P          R     mAP@.5
                 all         29          0          0          0          0          0
Speed: 15.3ms pre-process, 2286.9ms inference, 4.7ms NMS per image at shape (32, 3, 640, 640)
Results saved to runs/test/exp

[jaskiratsingh2000]

Didn't come out to be fruitful as mAP value is 0