pytorch-deep-image-matting icon indicating copy to clipboard operation
pytorch-deep-image-matting copied to clipboard

Published 20 hours ago • verified publisher icon huochaitiantang

A question on data augmentation part

Open haichaoyu opened this issue 5 years ago • 5 comments

np.random.randint() returns coordinates of (h, w) instead of (w, h). However, in core/data.py, line 42 - 44, the coordinates are regarded as (w, h)? Is it a bug or ... ? Please correct me if there is any problem.

Thanks.

haichaoyu avatar Mar 15 '19 19:03 haichaoyu

Thanks for your review! You are right. The coordinates of h and w is inversed. This mistake will result in that patches cropped from original image may not include any unknown region. I will fix this bug soon and conduct some new experiments later.

huochaitiantang avatar Mar 16 '19 01:03 huochaitiantang

Hi Liang,

May I ask that how many epochs or iterations (batch_size == 1?) of training make the evaluation performance of 72.9 in SAD on val set? Also, what is the test mode (crop, resize, whole) for 72.9?

Thanks.

haichaoyu avatar Mar 16 '19 19:03 haichaoyu

batch_size=1, epochs=22 and test mode is whole.

huochaitiantang avatar Mar 18 '19 01:03 huochaitiantang

Thanks!

haichaoyu avatar Mar 18 '19 01:03 haichaoyu

hi Thanks for your great work! I am confused about why you set the batch-size=1 in training process?

wrrJasmine avatar Jul 20 '19 06:07 wrrJasmine