search

deep-diver / Soccer-Ball-Detection-YOLOv2

YOLOv2 trained against custom dataset
116 stars 43 forks source link

no prediction with yolov2-tiny.weight #25

Open aimxu opened 4 years ago

aimxu commented 4 years ago

Thanks for the effort in collecting data and annotation @deep-diver

My os is windows 10, and the GPU just 2G ram

but when I train soccer data with yolov2-tiny.weight,(my GPU just 2G ram, can't train with yolov2.weight) there is no prediction after 40 epochs

I use yolov2-tiny.cfg and yolov2-tiny.wights in darknet official website, and change cfg' class to 1 and filter to 30

image

and the command here

python ./flow --labels labels_soccer.txt --model cfg/yolov2-tiny-soccer.cfg --load bin/yolov2-tiny.weights --train --annotation "E:/dataset/soccer_ball_data/Annotations" --dataset "E:/dataset/soccer_ball_data/images" --batch 4 --epoch 40 --save 20000 --backup soccer --savepb --gpu 0.5

The loss is far below 1, and it can generate .pb and .meta file in ./built_graph

but when I test image, the output image (using 10 images from original data)

python flow --pbLoad built_graph/yolov2-tiny-soccer.pb --metaLoad built_graph/yolov2-tiny-soccer.meta --imgdir sample_img/soccer --gpu 0.5

the is no boundingbox

Did somebody train this data with yolov2-tiny.weight? thanks in advance

yangx789 commented 2 years ago

I also used yolov2-tiny.weights and got empty array for the outputs. Did you find a solution for it? @aimxu