search

pjreddie / darknet

Convolutional Neural Networks
http://pjreddie.com/darknet/
Other
25.73k stars 21.33k forks source link

GPU and the prediction images have many labels #1380

Open xiayq1 opened 5 years ago

xiayq1 commented 5 years ago

I change the Makefile into: GPU=1

But I use: Nvidia-smi to see the usage of GPU is very low (0~2%)

I use the voc data to train the model. But the Avg IOU(is very low:0.2) I test the dog.jpg: ./darknet detect cfg/yolo3-voc.cfg yolo3-voc_100.weights data/dog.jpg

but there are full of labels . Ecen can't the image.

Then I run: ./darknet detect cfg/yolo3.cfg yolo3.weights data/dog.jpg

I found the auther only use 0.02s But I need 24s.

zh7117 commented 5 years ago

you need install cuda and cudnn if you wan to use nvidia gpu to accelerate calculation. and set CUDNN=1 in makefile. and yolo3-voc_100.weights means you only trained 100 batches. That' too small. I think you may need to train at least 10000 steps to make it have a certain degree of accuracy.

xiayq1 commented 5 years ago

Hi, I have changed GPU=1 and cudnn=1 in makefile. I also make it again. But it didn't work.

zh7117 commented 5 years ago

maybe you didn't install cuda correctly or your gpu didn't support cuda for accelerating calculation.

xiayq1 commented 5 years ago

But I can use GPU to accelerating calculation for other project.

xiayq1 commented 5 years ago

I have solved this problem. Pay attention to Makefile. Use GPU=1 , Not GPU=1 #0. The same rule to CUDNN=1.