search

naibaf7 / caffe

Caffe: a fast open framework for deep learning. With OpenCL and CUDA support.
http://caffe.berkeleyvision.org/
Other
85 stars 20 forks source link

CPU-related NetTest/0.TestSharedWeightsUpdate fails #18

Closed pkainz closed 8 years ago

pkainz commented 8 years ago

Hi,

I just compiled (without Greentea and cuDNN) and ran 

make runtest

All tests run fine, apart from one, which I wanted to report:

src/caffe/test/test_net.cpp:1164: Failure
Value of: shared_params.cpu_diff()[i]
  Actual: 8.2940083
Expected: ip1_weights->cpu_diff()[i] + ip2_weights->cpu_diff()[i]
Which is: 8.2939987
src/caffe/test/test_net.cpp:1164: Failure
Value of: shared_params.cpu_diff()[i]
  Actual: 10.278276
Expected: ip1_weights->cpu_diff()[i] + ip2_weights->cpu_diff()[i]
Which is: 10.278267
src/caffe/test/test_net.cpp:1164: Failure
Value of: shared_params.cpu_diff()[i]
  Actual: 10.855959
Expected: ip1_weights->cpu_diff()[i] + ip2_weights->cpu_diff()[i]
Which is: 10.855949
src/caffe/test/test_net.cpp:1164: Failure
Value of: shared_params.cpu_diff()[i]
  Actual: 9.1899281
Expected: ip1_weights->cpu_diff()[i] + ip2_weights->cpu_diff()[i]
Which is: 9.1899185
src/caffe/test/test_net.cpp:1164: Failure
Value of: shared_params.cpu_diff()[i]
  Actual: 12.916765
Expected: ip1_weights->cpu_diff()[i] + ip2_weights->cpu_diff()[i]
Which is: 12.916756
src/caffe/test/test_net.cpp:1164: Failure
Value of: shared_params.cpu_diff()[i]
  Actual: 7.1070118
Expected: ip1_weights->cpu_diff()[i] + ip2_weights->cpu_diff()[i]
Which is: 7.1070023
src/caffe/test/test_net.cpp:1164: Failure
Value of: shared_params.cpu_diff()[i]
  Actual: 13.106203
Expected: ip1_weights->cpu_diff()[i] + ip2_weights->cpu_diff()[i]
Which is: 13.106194
src/caffe/test/test_net.cpp:1164: Failure
Value of: shared_params.cpu_diff()[i]
  Actual: 9.9245319
Expected: ip1_weights->cpu_diff()[i] + ip2_weights->cpu_diff()[i]
Which is: 9.9245224
[  FAILED  ] NetTest/0.TestSharedWeightsUpdate, where TypeParam = caffe::CPUDevice<float> (1 ms)

I assume that this is a CPU-only issue and does not affect computations on the GPU? Thanks! Cheers, Phil

naibaf7 commented 8 years ago

@pkainz This is weird and should not happen (and I can't reproduce it), but the difference between actual value and expected value is small enough that it can be ignored. It also just affects CPU. It can be ignored.