Accuracy drop using pretrained NIN model.

[analog75]

Hi. Thank you for your uploading. I have downloaded NIN cifar10 pretrained model and load it. Then, I run the inference evaluation. At this time. the accuracy is just only 59.36%. In this evaluation, there is no modification and the model and data are used according to Readme.
But, when I run the training using pretrained model, the accuracy is very close to the best accuracy. I think that there seems to be any dependency in train and test in your code. To solve this problem, is there any trick or solution? Thanks.

[Lanweichao]

@analog75 Hello, have you solved your problem? I also obtained a lower accuracy using NIN. Thank you

[analog75]

Hi. I solve this. This code adopts random, which should be modified int pseudo-random. In addition, batch normalization can make your accuracy lowered, which is very natural. From: Lanweichao <notifications@github.com> Sent: Monday, April 6, 2020 6:25 PM To: jiecaoyu/XNOR-Net-PyTorch <XNOR-Net-PyTorch@noreply.github.com> Cc: analog75 <hyunjin2.kim@gmail.com>; Mention <mention@noreply.github.com> Subject: Re: [jiecaoyu/XNOR-Net-PyTorch] Accuracy drop using pretrained NIN model. (#77) @analog75 <https://github.com/analog75> Hello, have you solved your problem? I also obtained a lower accuracy using NIN. Thank you — You are receiving this because you were mentioned. Reply to this email directly, view it on GitHub <#77 (comment)> , or unsubscribe <https://github.com/notifications/unsubscribe-auth/ANJ6W3O2OBNVFD7HQS2YSODRLGNWTANCNFSM4I2XALYQ> . <https://github.com/notifications/beacon/ANJ6W3OICKRPTDEP6WV7RB3RLGNWTA5CNFSM4I2XALY2YY3PNVWWK3TUL52HS4DFVREXG43VMVBW63LNMVXHJKTDN5WW2ZLOORPWSZGOERLPOCQ.gif>

[ngocqn]

Hi. Thank you for your uploading. I have downloaded NIN cifar10 pretrained model and load it. Then, I run the inference evaluation. At this time. the accuracy is just only 59.36%. In this evaluation, there is no modification and the model and data are used according to Readme.
But, when I run the training using pretrained model, the accuracy is very close to the best accuracy. I think that there seems to be any dependency in train and test in your code. To solve this problem, is there any trick or solution? Thanks.

I had the same issue with you.
Test set: Average loss: 9.3183, Accuracy: 1003/10000 (10.03%)
Best Accuracy: 86.28%

How to solve this issue?