Open psyhtest opened 7 years ago
@Rhymmor Many thanks for the contribution! Please include the licensing terms next time :))
Just for reference (not expecting to do anything at this point), both VGG models fail on the laptop M1000M GPU with:
I0405 16:28:47.284858 19674 net.cpp:435] Input 0 -> data
I0405 16:28:47.285311 19674 layer_factory.hpp:76] Creating layer conv1_1
I0405 16:28:47.285326 19674 net.cpp:110] Creating Layer conv1_1
I0405 16:28:47.285331 19674 net.cpp:477] conv1_1 <- data
I0405 16:28:47.285336 19674 net.cpp:433] conv1_1 -> conv1_1
F0405 16:28:47.286236 19674 cudnn_conv_layer.cpp:138] Check failed: status == CUDNN_STATUS_SUCCESS (6 vs. 0) CUDNN_STATUS_ARCH_MISMATCH
*** Check failure stack trace: ***
@ 0x7f24121c53e6 google::LogMessage::Fail()
@ 0x7f24121c532a google::LogMessage::SendToLog()
@ 0x7f24121c4c66 google::LogMessage::Flush()
@ 0x7f24121c82ed google::LogMessageFatal::~LogMessageFatal()
@ 0x7f24121075c8 caffe::CuDNNConvolutionLayer<>::Reshape()
@ 0x7f241215adab caffe::Net<>::Init()
@ 0x7f241215cc7b caffe::Net<>::Net()
@ 0x4096d7 time()
@ 0x4064c0 main
@ 0x7f2410e4d830 __libc_start_main
@ 0x406c69 _start
@ (nil) (unknown)
when using the package:lib-caffe-nvidia-fp16-cudnn
variant. (package:lib-caffe-nvidia-fp16-cuda
works just fine.)
In fact, I see the same failure with GoogleNet... Will try other targets...
The Visual Geometry Group at the University of Oxford designed several highly accurate DNNs.
Add CK-Caffe support for:
Include a copy of the license terms.
Note that only the deployment topology is available. I'll provide an example of how to deal with that shortly.