◯ Chainer 5.0.0 + CuPy 5.0.0 + CUDA 10.0 + cuDnn 7.3.1 + python 3.6.7 の組み合わせで実験を行いました。
◯ Tesla V100 : FP32
$ time python ../imagenet/train_imagenet.py -a alex -g 0 -E 50 train.txt test.txt
epoch iteration main/loss validation/main/loss main/accuracy validation/main/accuracy lr
4 1000 3.39625 0.277813 0.01
9 2000 1.96937 0.51525 0.01
13 3000 1.31864 0.653625 0.01
18 4000 0.858497 0.761844 0.01
23 5000 0.594805 0.832781 0.01
27 6000 0.411216 0.879719 0.01
32 7000 0.315708 0.909844 0.01
37 8000 0.262002 0.926344 0.01
41 9000 0.20459 0.942563 0.01
46 10000 0.189098 0.945 0.01
real 5m43.091s
user 53m31.981s
sys 1m12.694s
◯ Tesla V100 : FP16
$python ../imagenet/train_imagenet.py -a alex_fp16 -g 0 -E 50 train.txt test.txt
epoch iteration main/loss validation/main/loss main/accuracy validation/main/accuracy lr
4 1000 3.402 0.2715 0.01
9 2000 2.032 0.506 0.01
13 3000 1.297 0.6415 0.01
18 4000 0.8445 0.79 0.01
23 5000 0.5955 0.878 0.01
27 6000 0.42675 0.9275 0.01
32 7000 0.30725 0.956 0.01
37 8000 0.249625 0.9705 0.01
41 9000 0.19925 0.9815 0.01
46 10000 0.189375 0.985 0.01
real 5m23.050s
user 53m37.481s
sys 1m9.477s
[
◯ GeForce GTX 1080 Ti : FP32
$ time python ../imagenet/train_imagenet.py -a alex -g 1 -E 50 train.txt test.txt
epoch iteration main/loss validation/main/loss main/accuracy validation/main/accuracy lr
4 1000 3.3708 0.282844 0.01
9 2000 1.97459 0.518375 0.01
13 3000 1.28717 0.661 0.01
18 4000 0.860205 0.762469 0.01
23 5000 0.582611 0.832125 0.01
27 6000 0.410189 0.881156 0.01
32 7000 0.319799 0.909125 0.01
37 8000 0.259161 0.924344 0.01
41 9000 0.209578 0.940156 0.01
46 10000 0.172087 0.951656 0.01
real 6m49.150s
user 62m17.459s
sys 1m46.578s
◯ GeForce GTX 1080 Ti : FP16
$ time python ../imagenet/train_imagenet.py -a alex_fp16 -g 1 -E 50 train.txt test.txt
epoch iteration main/loss validation/main/loss main/accuracy validation/main/accuracy lr
4 1000 3.318 0.2815 0.01
9 2000 2.038 0.5025 0.01
13 3000 1.319 0.6405 0.01
18 4000 0.877 0.786 0.01
23 5000 0.6085 0.86 0.01
27 6000 0.425 0.9255 0.01
32 7000 0.3225 0.955 0.01
37 8000 0.25675 0.968 0.01
41 9000 0.235375 0.972 0.01
46 10000 0.189625 0.986 0.01
real 6m7.448s
user 60m7.559s
sys 1m26.810s
◯ Tesla V100 : FP32
$ time python ../imagenet/train_imagenet.py -a alex -g 0 -E 50 train.txt test.txt
epoch iteration main/loss validation/main/loss main/accuracy validation/main/accuracy lr
4 1000 3.39625 0.277813 0.01
9 2000 1.96937 0.51525 0.01
13 3000 1.31864 0.653625 0.01
18 4000 0.858497 0.761844 0.01
23 5000 0.594805 0.832781 0.01
27 6000 0.411216 0.879719 0.01
32 7000 0.315708 0.909844 0.01
37 8000 0.262002 0.926344 0.01
41 9000 0.20459 0.942563 0.01
46 10000 0.189098 0.945 0.01
real 5m43.091s
user 53m31.981s
sys 1m12.694s
◯ Tesla V100 : FP16
$python ../imagenet/train_imagenet.py -a alex_fp16 -g 0 -E 50 train.txt test.txt
epoch iteration main/loss validation/main/loss main/accuracy validation/main/accuracy lr
4 1000 3.402 0.2715 0.01
9 2000 2.032 0.506 0.01
13 3000 1.297 0.6415 0.01
18 4000 0.8445 0.79 0.01
23 5000 0.5955 0.878 0.01
27 6000 0.42675 0.9275 0.01
32 7000 0.30725 0.956 0.01
37 8000 0.249625 0.9705 0.01
41 9000 0.19925 0.9815 0.01
46 10000 0.189375 0.985 0.01
real 5m23.050s
user 53m37.481s
sys 1m9.477s
[
◯ GeForce GTX 1080 Ti : FP32
$ time python ../imagenet/train_imagenet.py -a alex -g 1 -E 50 train.txt test.txt
epoch iteration main/loss validation/main/loss main/accuracy validation/main/accuracy lr
4 1000 3.3708 0.282844 0.01
9 2000 1.97459 0.518375 0.01
13 3000 1.28717 0.661 0.01
18 4000 0.860205 0.762469 0.01
23 5000 0.582611 0.832125 0.01
27 6000 0.410189 0.881156 0.01
32 7000 0.319799 0.909125 0.01
37 8000 0.259161 0.924344 0.01
41 9000 0.209578 0.940156 0.01
46 10000 0.172087 0.951656 0.01
real 6m49.150s
user 62m17.459s
sys 1m46.578s
◯ GeForce GTX 1080 Ti : FP16
$ time python ../imagenet/train_imagenet.py -a alex_fp16 -g 1 -E 50 train.txt test.txt
epoch iteration main/loss validation/main/loss main/accuracy validation/main/accuracy lr
4 1000 3.318 0.2815 0.01
9 2000 2.038 0.5025 0.01
13 3000 1.319 0.6405 0.01
18 4000 0.877 0.786 0.01
23 5000 0.6085 0.86 0.01
27 6000 0.425 0.9255 0.01
32 7000 0.3225 0.955 0.01
37 8000 0.25675 0.968 0.01
41 9000 0.235375 0.972 0.01
46 10000 0.189625 0.986 0.01
real 6m7.448s
user 60m7.559s
sys 1m26.810s