search

shicai / SENet-Caffe

A Caffe Re-Implementation of SENet
BSD 3-Clause "New" or "Revised" License
169 stars 101 forks source link

What kind of data-augmentation? #2

Closed youngwanLEE closed 7 years ago

youngwanLEE commented 7 years ago

I wonder what kind of data-augmentation you used in training.

Thank you :)

shicai commented 7 years ago

scale & aspect ratio augmentation.

408550969 commented 7 years ago

caffe how to implement centre crop? The default is random crop would you mind share train.prototxt? Thanks!

shicai commented 7 years ago

in test phase, center crop is the the default setting.

408550969 commented 7 years ago

Thanks

408550969 commented 7 years ago

May I ask another question? What is the use of scale?I know it can adjust the value of data to 0-4.352

shicai commented 7 years ago

Lots of guys asked me this question. You can find answers here: https://github.com/shicai/MobileNet-Caffe/issues/17 (If you are Chinese, you can also take a look at this link: https://www.zhihu.com/question/63649386/answer/211508137)

408550969 commented 7 years ago

Thanks!

miraclewkf commented 6 years ago

You only use scale and aspect ratio augmentation in training? What is the parameter of aspect ratio you set? You don't use random_mirror in training?

miraclewkf commented 6 years ago

One more question, do you resize the original image to one size (for example, resize shorter side to 256) before other data argumentations during training?