Chainer 4.4.0 と imagenet

以下の環境で imagenet の実験を行ってみました。
◯ Chainer 4.4.0 + CuPy 4.4.1 + cuDnn 7.3

◯ 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 3.42292 0.275406 0.01
9 2000 2.02364 0.504469 0.01
13 3000 1.33507 0.649781 0.01
18 4000 0.879725 0.759781 0.01
23 5000 0.596242 0.829 0.01
27 6000 0.429185 0.876781 0.01
32 7000 0.324969 0.907 0.01
37 8000 0.263767 0.92475 0.01
41 9000 0.218982 0.937219 0.01
46 10000 0.169761 0.952187 0.01

real 5m24.372s
user 52m0.814s
sys 1m56.314s

◯ Tesla V100 : FP16
$ time 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 3.322 0.278 0.01
9 2000 2.015 0.50275 0.01
13 3000 1.297 0.643 0.01
18 4000 0.878 0.773 0.01
23 5000 0.6015 0.872 0.01
27 6000 0.4385 0.9215 0.01
32 7000 0.2955 0.961 0.01
37 8000 0.253125 0.972 0.01
41 9000 0.198375 0.978 0.01
46 10000 0.19975 0.9815 0.01

real 4m43.798s
user 47m21.842s
sys 1m36.036s

◯ 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 3.44381 0.270656 0.01
9 2000 2.01306 0.508563 0.01
13 3000 1.3339 0.650937 0.01
18 4000 0.872331 0.758812 0.01
23 5000 0.600323 0.829594 0.01
27 6000 0.427101 0.876281 0.01
32 7000 0.318739 0.90825 0.01
37 8000 0.254063 0.927406 0.01
41 9000 0.217029 0.939 0.01
46 10000 0.195928 0.945063 0.01

real 6m41.923s
user 61m42.514s
sys 2m36.189s

◯ 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 3.406 0.26825 0.01
9 2000 2.046 0.50325 0.01
13 3000 1.329 0.6425 0.01
18 4000 0.88 0.774 0.01
23 5000 0.598 0.8725 0.01
27 6000 0.43225 0.919 0.01
32 7000 0.353 0.947 0.01
37 8000 0.25725 0.969 0.01
41 9000 0.220375 0.978 0.01
46 10000 0.1905 0.9805 0.01

real 6m2.500s
user 60m16.126s
sys 2m10.853s


参考：Chainer 4.3.1 + CuPy 4.3.0 + cuDnn 7.1.3
◯ 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 3.42937 0.275687 0.01
9 2000 2.0102 0.50725 0.01
13 3000 1.31432 0.652125 0.01
18 4000 0.85658 0.763563 0.01
23 5000 0.585906 0.831094 0.01
27 6000 0.422727 0.879281 0.01
32 7000 0.310402 0.910875 0.01
37 8000 0.244246 0.930906 0.01
41 9000 0.209276 0.94025 0.01
46 10000 0.171013 0.951531 0.01

real 5m33.551s
user 54m17.491s
sys 1m47.513s

◯ Tesla V100 : FP16
$ time 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 3.334 0.2735 0.01
9 2000 2.036 0.505 0.01
13 3000 1.331 0.649 0.01
18 4000 0.8945 0.7775 0.01
23 5000 0.6175 0.86 0.01
27 6000 0.43175 0.923 0.01
32 7000 0.32125 0.953 0.01
37 8000 0.2715 0.967 0.01
41 9000 0.212625 0.977 0.01
46 10000 0.173 0.984 0.01

real 4m47.544s
user 48m53.431s
sys 1m36.040s


◯ 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 3.37873 0.281375 0.01
9 2000 1.97629 0.511594 0.01
13 3000 1.29579 0.658469 0.01
18 4000 0.867154 0.760156 0.01
23 5000 0.596059 0.830938 0.01
27 6000 0.40342 0.882469 0.01
32 7000 0.314912 0.910125 0.01
37 8000 0.243203 0.929281 0.01
41 9000 0.22406 0.93675 0.01
46 10000 0.17793 0.949656 0.01

real 6m42.212s
user 63m2.643s
sys 1m51.877s

◯ 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 3.35 0.26975 0.01
9 2000 2.045 0.501 0.01
13 3000 1.306 0.637 0.01
18 4000 0.8805 0.778 0.01
23 5000 0.594 0.875 0.01
27 6000 0.4405 0.9165 0.01
32 7000 0.337 0.951 0.01
37 8000 0.255125 0.97 0.01
41 9000 0.2175 0.9775 0.01
46 10000 0.196 0.98 0.01

real 6m2.340s
user 60m37.003s
sys 1m43.699s