How many hours on ImageNet of training?

[betterhalfwzm]

How many hours on ImageNet of training?

[chenxin061]

It took about 72 hours with 8 Nvidia Tesla V100 GPUs for a single run.

[betterhalfwzm]

@chenxin061Thank you for your answer，and How to visualize the search net?

[chenxin061]

@betterhalfwzm You can refer to visualize.py, which is the same as the one in the original DARTS code.

[betterhalfwzm]

@chenxin061
05/31 10:34:22 AM Train Step: 000 Objs: 3.854694e-02 Acc: 100.000000
05/31 10:34:41 AM Train Step: 050 Objs: 1.033505e-01 Acc: 97.763480
05/31 10:35:00 AM Train Step: 100 Objs: 9.169728e-02 Acc: 98.174505
05/31 10:35:20 AM Train Step: 150 Objs: 8.977429e-02 Acc: 98.199503
05/31 10:35:38 AM Train Step: 200 Objs: 8.777784e-02 Acc: 98.235386
05/31 10:35:58 AM Train Step: 250 Objs: 8.747335e-02 Acc: 98.188496
05/31 10:36:18 AM Train Step: 300 Objs: 8.840194e-02 Acc: 98.146802
05/31 10:36:38 AM Train Step: 350 Objs: 8.770526e-02 Acc: 98.188212
05/31 10:36:58 AM Train Step: 400 Objs: 8.683270e-02 Acc: 98.203709
05/31 10:37:17 AM Train Step: 450 Objs: 8.628383e-02 Acc: 98.215771
05/31 10:37:37 AM Train Step: 500 Objs: 8.557520e-02 Acc: 98.222305
05/31 10:37:56 AM Train Step: 550 Objs: 8.519212e-02 Acc: 98.247505
05/31 10:38:16 AM Train Step: 600 Objs: 8.487809e-02 Acc: 98.278910
05/31 10:38:35 AM Train Step: 650 Objs: 8.480616e-02 Acc: 98.300691
05/31 10:38:55 AM Train Step: 700 Objs: 8.507484e-02 Acc: 98.312678
05/31 10:39:15 AM Train Step: 750 Objs: 8.438019e-02 Acc: 98.333472
05/31 10:39:28 AM Train_acc: 98.346000
05/31 10:39:28 AM Valid Step: 000 Objs: 6.910484e-02 Acc: 96.875000
05/31 10:39:32 AM Valid Step: 050 Objs: 9.949709e-02 Acc: 97.334559
05/31 10:39:35 AM Valid Step: 100 Objs: 9.746562e-02 Acc: 97.571163
05/31 10:39:38 AM Valid Step: 150 Objs: 1.013715e-01 Acc: 97.506209
05/31 10:39:39 AM Valid_acc: 97.520000
Epoch time: 317s.
05/31 10:39:39 AM Epoch: 558 lr 3.010405e-04
05/31 10:39:39 AM Train Step: 000 Objs: 1.087680e-01 Acc: 98.437500
05/31 10:40:00 AM Train Step: 050 Objs: 9.098258e-02 Acc: 98.437500
05/31 10:40:19 AM Train Step: 100 Objs: 8.530718e-02 Acc: 98.391089
this is result of retrain，acc is high，is right？

[chenxin061]

Yes, it is a quite impressive acc and it is the expected one.

[thinkInJava33]

i took me 3 hours per epoch with 2 1080ti

[hoangtuanvu]

It took about 72 hours with 8 Nvidia Tesla V100 GPUs for a single run.

@chenxin061 Which input size do you use?

[chenxin061]

It took about 72 hours with 8 Nvidia Tesla V100 GPUs for a single run.

@chenxin061 Which input size do you use?

224 x 224.

[hoangtuanvu]

@chenxin061 Could you give me some suggestion about which NAS model can use to achieve best accuracy on ImageNet? Good trade-off accuracy and speed.

[chenxin061]

@chenxin061 Could you give me some suggestion about which NAS model can use to achieve best accuracy on ImageNet? Good trade-off accuracy and speed.

You can refer to Circumventing Outliers of AutoAugment with Knowledge Distillation for state-of-the-art performance on ImageNet.

[hoangtuanvu]

Thank you!