2,215 views
この記事は最終更新から 961日 が経過しています。
(22) cuda-convnet2はやってみれない (T_T) では、買ったばかりの GTX760のスペック不足で cuda-convnet2が動かせないことが判明し、とってもがっかり…
当時は高額で購入をあきらめていた GTX780だが、最近は値下がりが進んでいることもあり、思い切って購入！
早速 cuda-convnet2にリベンジしてみた。
結果
cuda-convnet2が動いた！
cuda-convnet2に付属の CIFAR-10を 10epochs指定で動かしてみた。
10epochs で 68.5% の正解率だった。
========================= Running on CUDA device(s) 0 Current time: Thu Sep 11 22:46:45 2014 Saving checkpoints to /home/user/cuda/cuda-convnet2/save/cifar-10/ConvNet__2014-09-11_22.46.42 ========================= 1.1 (0.00%)... logprob: 2.120814, 0.784000, 0.784000 (0.887 sec) 1.2 (2.00%)... logprob: 1.835924, 0.675100, 0.675100 (0.836 sec) 1.3 (4.00%)... logprob: 1.682182, 0.615700, 0.615700 (0.836 sec) 1.4 (6.00%)... logprob: 1.637781, 0.596200, 0.596200 (0.837 sec) 1.5 (8.00%)... logprob: 1.574231, 0.575300, 0.575300 ======================Test output====================== logprob: 1.446012, 0.519900, 0.519900 ----------------------Averages------------------------- logprob: 1.446012, 0.519900, 0.519900 ------------------------------------------------------- Layer 'conv1' weights[0]: 7.474001e-03 [5.467787e-05] [7.315743e-03] Layer 'conv1' biases: 1.841275e-03 [6.501494e-06] Layer 'conv2' weights[0]: 8.184642e-03 [1.752141e-05] [2.140767e-03] Layer 'conv2' biases: 2.113849e-03 [7.906822e-06] Layer 'local3' weights[0]: 3.137425e-02 [2.088059e-06] [6.655327e-05] Layer 'local3' biases: 2.683400e-04 [9.420025e-07] Layer 'local4' weights[0]: 3.142803e-02 [2.171371e-06] [6.909025e-05] Layer 'local4' biases: 6.031835e-04 [1.956280e-06] Layer 'fc10' weights[0]: 9.292197e-03 [3.430814e-05] [3.692145e-03] Layer 'fc10' biases: 3.025318e-02 [7.367652e-05] ------------------------------------------------------- Saved checkpoint to /home/user/cuda/cuda-convnet2/save/cifar-10/ConvNet__2014-09-11_22.46.42 ======================================================= (1.218 sec) 2.1 (10.00%)... logprob: 1.506601, 0.543700, 0.543700 (0.843 sec) 2.2 (12.00%)... logprob: 1.475231, 0.522700, 0.522700 (0.842 sec) 2.3 (14.00%)... logprob: 1.398977, 0.501300, 0.501300 (0.843 sec) 2.4 (16.00%)... logprob: 1.388852, 0.493800, 0.493800 (0.840 sec) 2.5 (18.00%)... logprob: 1.347464, 0.480800, 0.480800 ======================Test output====================== logprob: 1.258990, 0.445000, 0.445000 ----------------------Averages------------------------- logprob: 1.258990, 0.445000, 0.445000 ------------------------------------------------------- Layer 'conv1' weights[0]: 1.066526e-02 [7.180368e-05] [6.732486e-03] Layer 'conv1' biases: 3.529629e-03 [4.447127e-06] Layer 'conv2' weights[0]: 8.477319e-03 [2.198427e-05] [2.593305e-03] Layer 'conv2' biases: 3.669117e-03 [5.606374e-06] Layer 'local3' weights[0]: 3.088271e-02 [2.739488e-06] [8.870618e-05] Layer 'local3' biases: 4.430772e-04 [9.733770e-07] Layer 'local4' weights[0]: 3.093651e-02 [2.908529e-06] [9.401605e-05] Layer 'local4' biases: 1.001289e-03 [2.039535e-06] Layer 'fc10' weights[0]: 1.015366e-02 [4.788852e-05] [4.716380e-03] Layer 'fc10' biases: 4.879263e-02 [9.227377e-05] ------------------------------------------------------- Saved checkpoint to /home/user/cuda/cuda-convnet2/save/cifar-10/ConvNet__2014-09-11_22.46.42 ======================================================= (1.204 sec) 3.1 (20.00%)... logprob: 1.320204, 0.466300, 0.466300 (0.849 sec) 3.2 (22.00%)... logprob: 1.322136, 0.469200, 0.469200 (0.849 sec) 3.3 (24.00%)... logprob: 1.254244, 0.445500, 0.445500 (0.844 sec) 3.4 (26.00%)... logprob: 1.265054, 0.444200, 0.444200 (0.852 sec) 3.5 (28.00%)... logprob: 1.239677, 0.438100, 0.438100 ======================Test output====================== logprob: 1.172489, 0.419600, 0.419600 ----------------------Averages------------------------- logprob: 1.172489, 0.419600, 0.419600 ------------------------------------------------------- Layer 'conv1' weights[0]: 1.303541e-02 [7.437402e-05] [5.705536e-03] Layer 'conv1' biases: 4.526725e-03 [5.591116e-06] Layer 'conv2' weights[0]: 8.756692e-03 [2.101911e-05] [2.400348e-03] Layer 'conv2' biases: 4.467094e-03 [5.962670e-06] Layer 'local3' weights[0]: 3.039985e-02 [2.865414e-06] [9.425750e-05] Layer 'local3' biases: 5.414516e-04 [1.031628e-06] Layer 'local4' weights[0]: 3.045136e-02 [3.177895e-06] [1.043597e-04] Layer 'local4' biases: 1.232660e-03 [2.390751e-06] Layer 'fc10' weights[0]: 1.063668e-02 [4.992195e-05] [4.693376e-03] Layer 'fc10' biases: 5.920979e-02 [9.777919e-05] ------------------------------------------------------- Saved checkpoint to /home/user/cuda/cuda-convnet2/save/cifar-10/ConvNet__2014-09-11_22.46.42 ======================================================= (1.224 sec) 4.1 (30.00%)... logprob: 1.194791, 0.418200, 0.418200 (0.852 sec) 4.2 (32.00%)... logprob: 1.224260, 0.438500, 0.438500 (0.851 sec) 4.3 (34.00%)... logprob: 1.163682, 0.408000, 0.408000 (0.852 sec) 4.4 (36.00%)... logprob: 1.175036, 0.417000, 0.417000 (0.852 sec) 4.5 (38.00%)... logprob: 1.155333, 0.406400, 0.406400 ======================Test output====================== logprob: 1.062875, 0.378300, 0.378300 ----------------------Averages------------------------- logprob: 1.062875, 0.378300, 0.378300 ------------------------------------------------------- Layer 'conv1' weights[0]: 1.509653e-02 [6.724540e-05] [4.454362e-03] Layer 'conv1' biases: 5.414162e-03 [6.911395e-06] Layer 'conv2' weights[0]: 9.012531e-03 [2.557474e-05] [2.837686e-03] Layer 'conv2' biases: 5.005291e-03 [6.856871e-06] Layer 'local3' weights[0]: 2.992433e-02 [3.229656e-06] [1.079274e-04] Layer 'local3' biases: 6.130654e-04 [1.145043e-06] Layer 'local4' weights[0]: 2.997499e-02 [3.427017e-06] [1.143292e-04] Layer 'local4' biases: 1.398424e-03 [2.556241e-06] Layer 'fc10' weights[0]: 1.108736e-02 [4.734496e-05] [4.270172e-03] Layer 'fc10' biases: 6.623243e-02 [9.970758e-05] ------------------------------------------------------- Saved checkpoint to /home/user/cuda/cuda-convnet2/save/cifar-10/ConvNet__2014-09-11_22.46.42 ======================================================= (1.210 sec) 5.1 (40.00%)... logprob: 1.112436, 0.392300, 0.392300 (0.854 sec) 5.2 (42.00%)... logprob: 1.132824, 0.396600, 0.396600 (0.853 sec) 5.3 (44.00%)... logprob: 1.092543, 0.382400, 0.382400 (0.854 sec) 5.4 (46.00%)... logprob: 1.111692, 0.393300, 0.393300 (0.854 sec) 5.5 (48.00%)... logprob: 1.080376, 0.383600, 0.383600 ======================Test output====================== logprob: 1.014830, 0.362300, 0.362300 ----------------------Averages------------------------- logprob: 1.014830, 0.362300, 0.362300 ------------------------------------------------------- Layer 'conv1' weights[0]: 1.681544e-02 [5.656535e-05] [3.363894e-03] Layer 'conv1' biases: 6.239711e-03 [6.276071e-06] Layer 'conv2' weights[0]: 9.254504e-03 [2.295077e-05] [2.479956e-03] Layer 'conv2' biases: 5.374766e-03 [5.774366e-06] Layer 'local3' weights[0]: 2.945584e-02 [3.233688e-06] [1.097809e-04] Layer 'local3' biases: 6.639807e-04 [9.701089e-07] Layer 'local4' weights[0]: 2.950781e-02 [3.495634e-06] [1.184647e-04] Layer 'local4' biases: 1.516043e-03 [2.256748e-06] Layer 'fc10' weights[0]: 1.149275e-02 [4.630167e-05] [4.028772e-03] Layer 'fc10' biases: 7.100429e-02 [9.157424e-05] ------------------------------------------------------- Saved checkpoint to /home/user/cuda/cuda-convnet2/save/cifar-10/ConvNet__2014-09-11_22.46.42 ======================================================= (1.213 sec) 6.1 (50.00%)... logprob: 1.060685, 0.376300, 0.376300 (0.850 sec) 6.2 (52.00%)... logprob: 1.069578, 0.375200, 0.375200 (0.853 sec) 6.3 (54.00%)... logprob: 1.032226, 0.359800, 0.359800 (0.851 sec) 6.4 (56.00%)... logprob: 1.064357, 0.373500, 0.373500 (0.854 sec) 6.5 (58.00%)... logprob: 1.045632, 0.365300, 0.365300 ======================Test output====================== logprob: 0.934916, 0.329700, 0.329700 ----------------------Averages------------------------- logprob: 0.934916, 0.329700, 0.329700 ------------------------------------------------------- Layer 'conv1' weights[0]: 1.835800e-02 [7.618176e-05] [4.149785e-03] Layer 'conv1' biases: 7.070946e-03 [4.932727e-06] Layer 'conv2' weights[0]: 9.482100e-03 [2.152068e-05] [2.269611e-03] Layer 'conv2' biases: 5.701233e-03 [4.869142e-06] Layer 'local3' weights[0]: 2.899688e-02 [3.212075e-06] [1.107731e-04] Layer 'local3' biases: 7.021243e-04 [9.119989e-07] Layer 'local4' weights[0]: 2.904793e-02 [3.381710e-06] [1.164183e-04] Layer 'local4' biases: 1.607356e-03 [2.076248e-06] Layer 'fc10' weights[0]: 1.183851e-02 [3.967304e-05] [3.351186e-03] Layer 'fc10' biases: 7.488154e-02 [7.453511e-05] ------------------------------------------------------- Saved checkpoint to /home/user/cuda/cuda-convnet2/save/cifar-10/ConvNet__2014-09-11_22.46.42 ======================================================= (1.209 sec) 7.1 (60.00%)... logprob: 0.999895, 0.350200, 0.350200 (0.855 sec) 7.2 (62.00%)... logprob: 1.039784, 0.365700, 0.365700 (0.856 sec) 7.3 (64.00%)... logprob: 0.990986, 0.348300, 0.348300 (0.851 sec) 7.4 (66.00%)... logprob: 0.995326, 0.342200, 0.342200 (0.854 sec) 7.5 (68.00%)... logprob: 0.969341, 0.339700, 0.339700 ======================Test output====================== logprob: 1.005710, 0.351300, 0.351300 ----------------------Averages------------------------- logprob: 1.005710, 0.351300, 0.351300 ------------------------------------------------------- Layer 'conv1' weights[0]: 1.968702e-02 [7.124672e-05] [3.618968e-03] Layer 'conv1' biases: 7.793765e-03 [5.323519e-06] Layer 'conv2' weights[0]: 9.687451e-03 [2.075292e-05] [2.142248e-03] Layer 'conv2' biases: 5.936751e-03 [4.238163e-06] Layer 'local3' weights[0]: 2.854378e-02 [3.352751e-06] [1.174600e-04] Layer 'local3' biases: 7.324723e-04 [9.184912e-07] Layer 'local4' weights[0]: 2.859599e-02 [3.617547e-06] [1.265054e-04] Layer 'local4' biases: 1.686072e-03 [2.135282e-06] Layer 'fc10' weights[0]: 1.215715e-02 [4.211784e-05] [3.464451e-03] Layer 'fc10' biases: 7.772458e-02 [7.265426e-05] ------------------------------------------------------- Saved checkpoint to /home/user/cuda/cuda-convnet2/save/cifar-10/ConvNet__2014-09-11_22.46.42 ======================================================= (1.214 sec) 8.1 (70.00%)... logprob: 0.970245, 0.340100, 0.340100 (0.858 sec) 8.2 (72.00%)... logprob: 0.988053, 0.348600, 0.348600 (0.853 sec) 8.3 (74.00%)... logprob: 0.949711, 0.330300, 0.330300 (0.861 sec) 8.4 (76.00%)... logprob: 0.964428, 0.331700, 0.331700 (0.853 sec) 8.5 (78.00%)... logprob: 0.951977, 0.332200, 0.332200 ======================Test output====================== logprob: 0.902606, 0.315900, 0.315900 ----------------------Averages------------------------- logprob: 0.902606, 0.315900, 0.315900 ------------------------------------------------------- Layer 'conv1' weights[0]: 2.086830e-02 [7.139416e-05] [3.421177e-03] Layer 'conv1' biases: 8.485496e-03 [4.683458e-06] Layer 'conv2' weights[0]: 9.879518e-03 [1.769133e-05] [1.790708e-03] Layer 'conv2' biases: 6.132266e-03 [3.429028e-06] Layer 'local3' weights[0]: 2.809841e-02 [3.082417e-06] [1.097008e-04] Layer 'local3' biases: 7.562819e-04 [7.674955e-07] Layer 'local4' weights[0]: 2.815062e-02 [3.344105e-06] [1.187933e-04] Layer 'local4' biases: 1.745867e-03 [1.828604e-06] Layer 'fc10' weights[0]: 1.240342e-02 [3.713923e-05] [2.994273e-03] Layer 'fc10' biases: 8.005437e-02 [5.533533e-05] ------------------------------------------------------- Saved checkpoint to /home/user/cuda/cuda-convnet2/save/cifar-10/ConvNet__2014-09-11_22.46.42 ======================================================= (1.215 sec) 9.1 (80.00%)... logprob: 0.941026, 0.327300, 0.327300 (0.852 sec) 9.2 (82.00%)... logprob: 0.953385, 0.335800, 0.335800 (0.856 sec) 9.3 (84.00%)... logprob: 0.931941, 0.324100, 0.324100 (0.853 sec) 9.4 (86.00%)... logprob: 0.933768, 0.324900, 0.324900 (0.856 sec) 9.5 (88.00%)... logprob: 0.928577, 0.320800, 0.320800 ======================Test output====================== logprob: 0.883051, 0.309200, 0.309200 ----------------------Averages------------------------- logprob: 0.883051, 0.309200, 0.309200 ------------------------------------------------------- Layer 'conv1' weights[0]: 2.197285e-02 [6.753075e-05] [3.073373e-03] Layer 'conv1' biases: 9.153372e-03 [4.402948e-06] Layer 'conv2' weights[0]: 1.006375e-02 [1.881227e-05] [1.869310e-03] Layer 'conv2' biases: 6.313587e-03 [3.937244e-06] Layer 'local3' weights[0]: 2.766130e-02 [3.224194e-06] [1.165598e-04] Layer 'local3' biases: 7.823407e-04 [7.771447e-07] Layer 'local4' weights[0]: 2.771336e-02 [3.435646e-06] [1.239707e-04] Layer 'local4' biases: 1.799375e-03 [1.811628e-06] Layer 'fc10' weights[0]: 1.262947e-02 [3.785512e-05] [2.997363e-03] Layer 'fc10' biases: 8.223019e-02 [5.913675e-05] ------------------------------------------------------- Saved checkpoint to /home/user/cuda/cuda-convnet2/save/cifar-10/ConvNet__2014-09-11_22.46.42 ======================================================= (1.217 sec) 10.1 (90.00%)... logprob: 0.899000, 0.315300, 0.315300 (0.853 sec) 10.2 (92.00%)... logprob: 0.920165, 0.318600, 0.318600 (0.856 sec) 10.3 (94.00%)... logprob: 0.907986, 0.316500, 0.316500 (0.854 sec) 10.4 (96.00%)... logprob: 0.912464, 0.320900, 0.320900 (0.855 sec) 10.5 (98.00%)... logprob: 0.886616, 0.314900, 0.314900 ======================Test output====================== logprob: 0.887153, 0.315100, 0.315100 ----------------------Averages------------------------- logprob: 0.887153, 0.315100, 0.315100 ------------------------------------------------------- Layer 'conv1' weights[0]: 2.302468e-02 [6.561136e-05] [2.849610e-03] Layer 'conv1' biases: 9.744839e-03 [4.328744e-06] Layer 'conv2' weights[0]: 1.024172e-02 [1.659464e-05] [1.620298e-03] Layer 'conv2' biases: 6.463769e-03 [3.200243e-06] Layer 'local3' weights[0]: 2.723007e-02 [3.112101e-06] [1.142891e-04] Layer 'local3' biases: 8.001459e-04 [6.947578e-07] Layer 'local4' weights[0]: 2.728442e-02 [3.306802e-06] [1.211974e-04] Layer 'local4' biases: 1.843459e-03 [1.568608e-06] Layer 'fc10' weights[0]: 1.286868e-02 [3.347505e-05] [2.601281e-03] Layer 'fc10' biases: 8.373742e-02 [2.511169e-05] ------------------------------------------------------- Saved checkpoint to /home/user/cuda/cuda-convnet2/save/cifar-10/ConvNet__2014-09-11_22.46.42 ======================================================= (1.207 sec) 11.1 (100.00%)... logprob: 0.880133, 0.312300, 0.312300 (0.855 sec)