Why using GPU become less efficient?

[xiaoxinyi]

Test time for detect_face:

start_time = time.time()
# run detector
results = detector.detect_face(img)
end_time = time.time()
diff = end_time - start_time
print datetime.timedelta(seconds=diff)

• ctx=mx.gpu(0), num_work=1 takes 1.807s

• ctx=mx.gpu(0), num_work=4 takes 2.307s

• ctx=mx.cpu(0), num_work=4 takes0.601s

ps : GTX1070

[Li1991]

When I use GPU, I encounter this problem:

[11:16:43] /home/njfh/alex/mxnet/dmlc-core/include/dmlc/logging.h:300: [11:16:43] src/c_api/c_api_ndarray.cc:390: Operator _zeros cannot be run; requires at least one of FCompute, NDArrayFunction, FCreateOperator be registered

Stack trace returned 10 entries: [bt] (0) /usr/local/lib/python2.7/dist-packages/mxnet-0.9.5-py2.7.egg/mxnet/libmxnet.so(_ZN4dmlc15LogMessageFatalD1Ev+0x3c) [0x7f34bf503a5c] [bt] (1) /usr/local/lib/python2.7/dist-packages/mxnet-0.9.5-py2.7.egg/mxnet/libmxnet.so(MXImperativeInvoke+0x66d) [0x7f34bfd9811d] [bt] (2) /usr/lib/x86_64-linux-gnu/libffi.so.6(ffi_call_unix64+0x4c) [0x7f34c2976adc] [bt] (3) /usr/lib/x86_64-linux-gnu/libffi.so.6(ffi_call+0x1fc) [0x7f34c297640c] [bt] (4) /usr/lib/python2.7/lib-dynload/_ctypes.x86_64-linux-gnu.so(_ctypes_callproc+0x48e) [0x7f34c2b8d5fe] [bt] (5) /usr/lib/python2.7/lib-dynload/_ctypes.x86_64-linux-gnu.so(+0x15f9e) [0x7f34c2b8ef9e] [bt] (6) python(PyEval_EvalFrameEx+0x98d) [0x5244dd] [bt] (7) python(PyEval_EvalCodeEx+0x2b1) [0x555551] [bt] (8) python(PyEval_EvalFrameEx+0x1a10) [0x525560] [bt] (9) python(PyEval_EvalCodeEx+0x2b1) [0x555551]

Can you tell me how to solve this problem? Thank you very much! @xiaoxinyi

[xiaoxinyi]

@Li1991 Try mxnet==0.9.2.

[Li1991]

Thank you for answering me! I have found the version of 0.9.2 for a day,but can not find it. Would you please send it to me ?(891935370@qq.com) Thank you very much! @xiaoxinyi

[xiaoxinyi]

@Li1991 Alternatives:

git clone --recursive https://github.com/dmlc/mxnet.git

git checkout -b v0.9.2

[Li1991]

Thank you very much @xiaoxinyi

[YYuanAnyVision]

@xiaoxinyi mxnet spends more time on GPU memory allocation at the first time, it should be ok when the initialization is done. you can detect the same image twice, it should be faster the second time.

[Li1991]

The paper author said in the paper that he can get the speed of 99fps. But I can only get 17fps using GTX 1080. Do you know why, and how to solve this problem? Thank you ! @xiaoxinyi @pangyupo

[xiaoxinyi]

I can't even get 17fps both in Mxnet and tensorflow.

[Li1991]

Have you tried this method in caffe and test the speed? @xiaoxinyi

[xiaoxinyi]

@Li1991 No.

[flankechen]

same issue, I guess time counting is not right here, and GPU needs time for data dumping.

[tcye]

the same problem, and i can't event get 16 fps on cpu @3.1G

[hujuan940506]

@xiaoxinyi @Li1991 @tcye How big are your pictures？

[lynnw123]

since minL = org_L(12/minisize)factor^(n), the size of your input image, the minisize and factor you set will affect the total number of calculations and thus the speed

[chen849157649]

Test time for detect_face:

start_time = time.time()
# run detector
results = detector.detect_face(img)
end_time = time.time()
diff = end_time - start_time
print datetime.timedelta(seconds=diff)

• ctx=mx.gpu(0), num_work=1 takes 1.807s
• ctx=mx.gpu(0), num_work=4 takes 2.307s
• ctx=mx.cpu(0), num_work=4 takes0.601s

ps : GT

Test time for detect_face:

start_time = time.time()
# run detector
results = detector.detect_face(img)
end_time = time.time()
diff = end_time - start_time
print datetime.timedelta(seconds=diff)

• ctx=mx.gpu(0), num_work=1 takes 1.807s
• ctx=mx.gpu(0), num_work=4 takes 2.307s
• ctx=mx.cpu(0), num_work=4 takes0.601s

ps : GTX1070

hi , Are you sure to use the GPU? The detection process uses the numpy data format, and does not use nd.ndarray. If I want to detect multiple images at the same time, detect it on two gpus. How to use the GPU? thank you very much!