jyegerlehner

jyegerlehner

jyegerlehner

CIFAR10 Benchmark
../build/tools/caffe  time -model examples/cifar10/cifar10_quick_train_test.prototxt -gpu=0


PR2610 Hawaii R9-290X, ViennaCL
---------------------------------
I0711 08:01:09.819653 13747 caffe.cpp:277] Average time per layer: 
I0711 08:01:09.819666 13747 caffe.cpp:280]      cifar	forward: 0.68638 ms.
I0711 08:01:09.819691 13747 caffe.cpp:283]      cifar	backward: 0.0209 ms.
I0711 08:01:09.819710 13747 caffe.cpp:280]      conv1	forward: 4.59994 ms.

jyegerlehner

[  FAILED  ] 25 tests, listed below:
[  FAILED  ] SGDSolverTest/2.TestLeastSquaresUpdate, where TypeParam = caffe::FloatGPU
[  FAILED  ] SGDSolverTest/2.TestLeastSquaresUpdateLROneTenth, where TypeParam = caffe::FloatGPU
[  FAILED  ] SGDSolverTest/2.TestLeastSquaresUpdateWithWeightDecay, where TypeParam = caffe::FloatGPU
[  FAILED  ] SGDSolverTest/2.TestLeastSquaresUpdateWithMomentum, where TypeParam = caffe::FloatGPU
[  FAILED  ] SGDSolverTest/2.TestLeastSquaresUpdateWithMomentumMultiIter, where TypeParam = caffe::FloatGPU
[  FAILED  ] SGDSolverTest/2.TestLeastSquaresUpdateWithEverything, where TypeParam = caffe::FloatGPU
[  FAILED  ] AdaGradSolverTest/2.TestAdaGradLeastSquaresUpdate, where TypeParam = caffe::FloatGPU
[  FAILED  ] AdaGradSolverTest/2.TestAdaGradLeastSquaresUpdateLROneTenth, where TypeParam = caffe::FloatGPU
[  FAILED  ] AdaGradSolverTest/2.TestAdaGradLeastSquaresUpdateWithWeightDecay, where TypeParam = caffe::FloatGPU

jyegerlehner

1) #include <tuple> requires c++11, added compiler flag not specified in CMakeLists.txt.
2) -DUSE_GREENTEA -DVIENNACL_WITH_OPENCL not defined in CMakelists.txt
3) libviennacl required (though ViennaCL claims to be header-only). Get viennacl from github, 
build to create the library, fix cmake files because it was hopelessly unable to find the
OpenCL installation, copy in libviennacl.so and .h to /usr/local/*. 
Modify caffe cmake files to find the library and header.
4) undefined reference Caffe::GetDeviceCOntext because CPU_ONLY is defined, 
though not in the cmake file. Due to:

  if(NOT HAVE_CUDA)

jyegerlehner

name: "net name"
input: "data"    <---------- Name of the input blob to the network. The blob is provided as argument to Net::Forward()
input_dim: 1     <---------- Number of images in the blob
input_dim: 3     <---------- Number of color channels in the input (e.g. RGB=3 for images)
input_dim: 32    <---------- width and height of the input image.
input_dim: 32    <-------/

layer {
  name: "encode1_conv"
  type: "Convolution"

jyegerlehner

# reduce learning rate after 120 epochs (60000 iters) by factor 0f 10
# then another factor of 10 after 10 more epochs (5000 iters)

# The train/test net protocol buffer definition
net: "examples/cifar10/cifar10_full_train_test_prelu.prototxt"
# test_iter specifies how many forward passes the test should carry out.
# In the case of CIFAR10, we have test batch size 100 and 100 test iterations,
# covering the full 10,000 testing images.
test_iter: 100
# Carry out testing every 1000 training iterations.

jyegerlehner

name: "Inception_3a"

input: "maxpool_2"
input_dim: 1
input_dim: 192
input_dim: 28
input_dim: 28

layers {
  name: "1x1_conv"