Line 144 of code

[mengxiangming]

Line 144 of code

[mengxiangming]

Hi Sajad, Thanks so much for your kindness. I see it in your repo the results are quite good. I will have a try with it though actrually I hope to train Bi-Rreal Net from scratch:) Thanks! Best regards, Xiangming Sajad Darabi <[email protected]> 于 2019年9月18日周三 04:11写道：

…

Hi Xiang Ming, Yes, they it should converge properly if you use the correct settings. I've added pretrained weights (which should probably save you a week..), and accuracy curves in the readme. The top-1 accuracies I get for the fp model at epoch 120 for bireal18 and bireal34 are 67.928% and 70.246% respectively. As for the code, I currently don't have access to my laptop. I'll try to push it end of this week or early next week. Hope this helps :) Sajad — You are receiving this because you authored the thread. Reply to this email directly, view it on GitHub <#1?email_source=notifications&email_token=AD6RNT3I2PUGXDNMP47DHQLQKETW5A5CNFSM4IXJKZBKYY3PNVWWK3TUL52HS4DFVREXG43VMVBW63LNMVXHJKTDN5WW2ZLOORPWSZGOD65S4OQ#issuecomment-532360762>, or mute the thread <https://github.com/notifications/unsubscribe-auth/AD6RNTZQZ6A4QVN6RBEJRLLQKETW5ANCNFSM4IXJKZBA> .

[sajaddarabi]

Hi,

self.tanh1 is really a placeholder that could be used as an activation after the first bn1, though I believe in the original birealnet they don't do so.

I'll try to push the complete training code this weekend or the following week once I have some headroom :)

[mengxiangming]

Hi Sajad: Thanks so much for your reply. I noted that in your code, the activation seems to be added in the self.layer1 where the self._make_layer uses tanh as the first operation. So I think they may be the same as the original one. Recently I only achieved Top1 accuracy 13% in 7 epochs using your network, which is quite slow compared to full precision case (10-20 in 1 epoch). What is your Top1 acc vs. epochs? Is it also such slow? And what is your final Top1 acc? Looking forward to your complete code since I am doing some comparisons. Is it possible to send me first by email since I am in a hurry. Thanks very very much. Xiangming

…

在 2019年9月17日，下午3:40，Sajad Darabi ***@***.***> 写道： Hi, self.tanh1 is really a placeholder that could be used as an activation after the first bn1, though I believe in the original birealnet they don't do so. I'll try to push the complete training code this weekend or the following week once I have some headroom :) — You are receiving this because you authored the thread. Reply to this email directly, view it on GitHub <#1?email_source=notifications&email_token=AD6RNTYKO7FSTCL26VKN5LTQKB3ULA5CNFSM4IXJKZBKYY3PNVWWK3TUL52HS4DFVREXG43VMVBW63LNMVXHJKTDN5WW2ZLOORPWSZGOD63PDRY#issuecomment-532083143>, or mute the thread <https://github.com/notifications/unsubscribe-auth/AD6RNT5Z36M5LTW5ZNONSRLQKB3ULANCNFSM4IXJKZBA>.

[sajaddarabi]

Hi Xiang Ming,

Yes, it should converge properly if you use the correct settings.

I've added pretrained weights (which should probably save you a week..), and accuracy curves in README.md.

The top-1 accuracies I get for the fp model at epoch 120 for bireal18 and bireal34 are 67.928% and 70.246% respectively.

As for the code, I currently don't have access to my laptop. I'll try to push it end of this week or early next week.

Hope this helps :)
Sajad