search

AlexeyAB / darknet

YOLOv4 / Scaled-YOLOv4 / YOLO - Neural Networks for Object Detection (Windows and Linux version of Darknet )
http://pjreddie.com/darknet/
Other
21.8k stars 7.97k forks source link

"csresnext50-panet-spp-original-optimal" model FPS #5122

Open mnbv7581 opened 4 years ago

mnbv7581 commented 4 years ago

We combined our custom dataset with the COCO dataset to learn. Target models used the two versions of [csresnext50-spp-original-optimal] and [yolo-spp] and unified the scale to 608x608. The problem here is that [csresnext50-spanet-spp-original-optimal] models show faster and more accurate performance than [yolo-spp] models, but the mAP shows better performance, but the speed is rather slow and confusing. Is there a possibility that I didn't learn well?

Specipication OS - WIN10 64bit GPU - 2080ti CPU - i9-9900k RAM - 32GB

yolo-spp [net] batch=1 subdivisions=1 width=608 height=608 channels=3 momentum=0.9 decay=0.0005 angle=0 saturation = 1.5 exposure = 1.5 hue=.1 learning_rate=0.00025 burn_in=1000 max_batches = 800200 policy=steps steps=400000,500000,600000,700000 scales=.1,.1,.1,.1 [yolo] mask = 0,1,2 anchors = 15, 29, 35, 70, 46,162, 100,102, 82,265, 209,199, 139,385, 262,440, 469,485 classes=5 num=9 jitter=.3 ignore_thresh = .7 truth_thresh = 1 random=1 iou_normalizer=1.25 iou_loss=giou

yolo-spp

csresnext50-panet-spp-original-optimal [net] batch=1 subdivisions=1 width=608 height=608 channels=3 momentum=0.949 decay=0.0005 angle=0 saturation = 1.5 exposure = 1.5 hue=.1 learning_rate = 0.0006525 burn_in=4000 max_batches = 2002000 policy=steps steps=400000,450000 scales=.1,.1 [yolo] mask = 6,7,8 anchors = 15, 29, 35, 70, 46,162, 100,102, 82,265, 209,199, 139,385, 262,440, 469,485 classes=5 num=9 jitter=.3 ignore_thresh = .7 truth_thresh = 1 random=1 scale_x_y = 1.05 iou_thresh=0.213 cls_normalizer=1.0 iou_normalizer=0.07 uc_normalizer=0.07 iou_loss=ciou nms_kind=greedynms beta_nms=0.6 beta1=0.6

csresnext50-panet-spp-original-optimal_coco

WongKinYiu commented 4 years ago

it seems group convolution is not efficient in modern GPU, i suggest you to use CSPDarknet53-PANet-SPP instead of CSPResNeXt50-PANet-SPP in this case.

#

Detector FPS on GeForce RTX 2070 (Tensor Cores):

FPS - measured using the command: ./darknet detector demo cfg/coco.data ... -benchmark CUDNN_HALF=1 (Mixed-precision is forced for Tensor Cores (if groups==1))

  1. 512x512: yolov3-spp - 52.0 FPS - (--ms ) csresnext50-panet-spp - 36.5 FPS - (--ms )

  2. 608x608: yolov3-spp - 38.0 FPS - (--ms ) csresnext50-panet-spp - 33.9 FPS (--ms )

CUDNN_HALF=0

  1. 512x512: yolov3-spp - 41.4 FPS - (--ms) csresnext50-panet-spp - 34.5 FPS - (--ms)

  2. 608x608: yolov3-spp - 26.1 FPS - (--ms) csresnext50-panet-spp - 30.0 FPS (--ms)

more details