Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

about inference image shape #7

Open
wenhe-jia opened this issue Jul 27, 2017 · 3 comments
Open

about inference image shape #7

wenhe-jia opened this issue Jul 27, 2017 · 3 comments

Comments

@wenhe-jia
Copy link

wenhe-jia commented Jul 27, 2017

In scope.py, the shape of input image is [3, 320, 320] during inference step, i want to know how did you do it?
I saw you use ImageRecorderIter to preprocess the input images, does it just resize the image into [3, 320, 320], or some other operations?
Thx for your help!~~

@hcwang95
Copy link

I use im2rec in MXNet to generate images based according to #4 , then choosing choose resize=320 will yield default image size as 320*320. Also you could set resize=224 when doing .rec file generation. But still I cannot reproduce the results mentioned in the README and the Top-1 error is usually lower by around 1% and Top-5 errors is lower by 0.2%, I try to boost the accuracy by center croping but I don't know how to reproduce the similar result. @cypw Could you give us more hints?

@cypw
Copy link
Owner

cypw commented Jul 29, 2017

@whcacademy

Please set the unchanged option to 1 while generating the .rec file using im2rec as I have already indicated in run_val.sh.

Note that: By setting unchanged=1, im2rec will directly save the raw image data without even decode it, see: im2rec.cc. So you won't lose any valuable information in this step, which may potentially damage the performance.

@hcwang95
Copy link

Sorry for ignoring import comments in that file.
And new test does reach the performance in the README.

Thanks!

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment
Labels
None yet
Projects
None yet
Development

No branches or pull requests

3 participants