-
Notifications
You must be signed in to change notification settings - Fork 1.7k
New issue
Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.
By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.
Already on GitHub? Sign in to your account
SSD300 Detection Time #123
Comments
Did you use the ssd_pascal_speed.py? |
Thanks for the quick response. No, I used the ssd_pascal.py instead. Can you briefly talk about the difference? |
use |
When I use the model we trained to detect ,we can only get the speed of 26fps. I remember that you have get 58fps in your paper.I just want to know why you are so fast? We use the pic of size 300x300, and the GPU we use is K40. The cpp file we use is ssd_detect.cpp from the file we download from your github address! |
The author suggested in another issue that the Titan X (card used by author) is about twice as powerful as the K40 which would explain your 26fps. |
https://github.com/ShaoqingRen/faster_rcnn has the speed for using K40 and Titan X. I haven't tested the speed of Faster R-CNN myself. |
Thanks for that. I find it very odd that using the titan x with faster rcnn produces very little improvement. Thats a question for the faster rcnn folks though. |
I have uesed the same as you, but only get 26fps. Can you tell me about how to be faster? I use the ssd_detect.cpp @weiliu89 |
@Li1991 You could try ssd_pascal_speed.py and you can see how long it takes to process the whole VOC07 test dataset. And from your previous comment, you said you are using K40, which is slower than Titan X. |
Ok, Thank you very much! I will have a try ! @weiliu89 |
Hi @weiliu89 , I was benchmarking SSD on voc07, using Titan X and cudnn. SSD 500 is about 13 fps, and SSD 300 is about 23 fps (similar to @CHUNYUWANG ). Since both are two times slower than what you reported, I guess I might did something wrong. Any insights of the reasons are much appreciated! |
@jzwang1 Did you use ssd_pascal_speed.py to see how much time it takes to process 4952 image? Besides I would also suggest set |
Thank you for your suggestions, @weiliu89 . Much appreciated! I was using the following caffe interface to benchmark the time taken by one forward pass: ./build/tools/caffe time --model models/VGGNet/VOC0712/SSD_300x300/deploy.prototxt --gpu 0 This above command gives: I0920 17:00:35.315847 47696 caffe.cpp:401] Average time per layer: Where at the last three lines, we see average forward pass takes 43ms for SSD 300. I guess this time is irrelevant from the batch size? I am also running this on a server, and no one else is using the same GPU. I also tested using ssd_pascal_speed.py just now. It takes 101s to finish 4952 images (see below). So roughly 50fps, which is two times faster than what reported by the above caffe interface. I am not quite sure which one measures the running time more appropriately. Or is there anything special in ssd_pascal_speed.py that causes such difference? I0921 09:09:09.878713 48790 net.cpp:693] Ignoring source layer mbox_loss |
Here is what I get when I run: ./build/tools/caffe time --model models/VGGNet/VOC0712/SSD_300x300/deploy.prototxt --gpu 0 I0921 12:44:57.308749 93668 caffe.cpp:369] *** Benchmark begins *** |
Thanks a lot! @weiliu89 |
The time above is using cuDNN v5. Using v4 has the following timing: I0921 12:51:19.585435 101116 caffe.cpp:412] Average Forward pass: 19.9982 ms. ssd_pascal_speed.py process images with batch size of 8, which is slightly faster. |
I see. Thank you for the statistics. I will investigate more. |
I get backward faster than forward, what am I doing wrong I1014 18:31:08.736440 31775 caffe.cpp:404] mbox_priorbox forward: 0.100224 ms. |
It turns out that I was not using cuDNN with v5 I get your numbers. I1014 19:37:53.356597 19125 caffe.cpp:412] Average Forward pass: 17.2102 ms. But it's weird that backward pass in faster than forward pass without cuDNN |
Hi @weiliu89 ,
here's the time test without cudnn5:
Here's my time test with cudnn5:
My relu type is PRelu, so it spend some time. And when I use nvidia-smi -lms, I find that the gpu doesn't run in most of time. It usually waits a lot of time and than hit some level of usage at one time, and wait again! |
@CHUNYUWANG Hi, I use ssd_detect.py and get 16 fps on GTX 1080ti, so which graphics card do you use? I have submitted a new issue about the speed problem #832 |
@mxmxlwlw hi,did you solve it? i have the same problem now,and i don't know why it is toooo slow,about 6 fps on TITAN V...thanks!! |
I was testing SSD 300 on a new dataset consisting of 30 object classes. Testing on an image (cropped to 300*300) takes about 0.04 seconds (25fps). This is about two times slower than 58fps. I am using cudnn v4, Titan X, on windows server 2012. Do you know what factors can cause the difference?
The text was updated successfully, but these errors were encountered: