search

daijifeng001 / MNC

Instance-aware Semantic Segmentation via Multi-task Network Cascades
Other
489 stars 182 forks source link

speed #4

Open kaishijeng opened 8 years ago

kaishijeng commented 8 years ago

It is quite slow to process one frame, like 300-400ms/fame on a titanX card. R-FCN is faster from the paper, but I am not able to make it working because of lacking Matlab. Just wondering how much speed will improve if I use Alexnet or Squeezenet instead of VGG16

Thanks,

HaozhiQi commented 8 years ago

We didn't test the speed of Alexnet. But you may reach ~160ms/frame on a TitanX card if you use ZF net.

kaishijeng commented 8 years ago

Do you have network arch, test.prototxt and pre-trained model of ZF net available so that I can try it?

Thanks

On Mon, Jul 11, 2016 at 1:50 AM, Haozhi Qi notifications@github.com wrote:

We didn't test the speed of Alexnet. But you may reach ~160ms/frame on a TitanX card if you use ZF net.

— You are receiving this because you authored the thread. Reply to this email directly, view it on GitHub https://github.com/daijifeng001/MNC/issues/4#issuecomment-231676753, or mute the thread https://github.com/notifications/unsubscribe/AMGg3n6YZ1RbG_v3EZKMQuPD-yTQSgSVks5qUgPpgaJpZM4JJCWH .

qinhaifangpku commented 8 years ago

@Oh233 hi~ i am train on my own data. about 12s/iter on K80, is it something wrong happend?

HaozhiQi commented 8 years ago

@qinhaifangpku

I use iter_size = 8 in the solver, in order to simulate the 8gpu training described in paper. Thus the time should be 1.5s/iter in your case. Maybe that is a reasonable time for K80 because I didn't have such GPU to test.

If you are interested, you may set a timer to see which part cause the slow speed. (e.g. preparing data in CPU or the network forward)

qinhaifangpku commented 8 years ago

@Oh233 thank you ! i set a timer in mnc_data_layer to see which part to slow speed but in show that get_next_minibatch did not take too long. i wonder if i can see excute time in every layer? my data only has too much mask which different from yours.