GoogLeNet128Residual

I0706 03:08:35.628417 11608 solver.cpp:338] Iteration 320000, Testing net (#0)
I0706 03:09:23.798849 11608 solver.cpp:406] Test net output #0: acc = 0.63414
I0706 03:09:23.799021 11608 solver.cpp:406] Test net output #1: loss1/loss1 = 2.14683 (* 0.3 = 0.644049 loss)
I0706 03:09:23.799041 11608 solver.cpp:406] Test net output #2: loss1/top-1 = 0.506081
I0706 03:09:23.799049 11608 solver.cpp:406] Test net output #3: loss1/top-5 = 0.758601
I0706 03:09:23.799070 11608 solver.cpp:406] Test net output #4: loss2/loss1 = 1.93225 (* 0.3 = 0.579675 loss)
I0706 03:09:23.799082 11608 solver.cpp:406] Test net output #5: loss2/top-1 = 0.544339
I0706 03:09:23.799093 11608 solver.cpp:406] Test net output #6: loss2/top-5 = 0.791
I0706 03:09:23.799108 11608 solver.cpp:406] Test net output #7: loss3/loss3 = 1.56532 (* 1 = 1.56532 loss)
I0706 03:09:23.799116 11608 solver.cpp:406] Test net output #8: loss3/top-5 = 0.84674
I0706 03:09:23.799125 11608 solver.cpp:323] Optimization Done.
I0706 03:09:23.799134 11608 caffe.cpp:216] Optimization Done.

vgg16_128_all_tricks

VGGNet16_128_All 0.682 1.47 ELU (a=0.5. a=1 leads to divergence :( ), avg+max pool, color conversion, linear lr_policy

caffenet128_ELU_alpha=0.5

ELU 0.485 2.29 alpha=0.5

vgg16_128

Here is pre-trained vgg-16 on image size 128 px.

I0507 14:04:05.104858 9864 solver.cpp:317] Iteration 320000, loss = 0.903925
I0507 14:04:05.104883 9864 solver.cpp:337] Iteration 320000, Testing net (#0)
I0507 14:06:25.156137 9864 solver.cpp:404] Test net output #0: accuracy = 0.651261
I0507 14:06:25.156198 9864 solver.cpp:404] Test net output #1: loss = 1.46029 (* 1 = 1.46029 loss)

ResNet-50ELU-2xThinner

Weights of the trained model ResNet-50ELU-2xThinner