search

huochaitiantang / pytorch-deep-image-matting

Pytorch implementation of deep image matting
294 stars 71 forks source link

A question on data augmentation part #9

Open haichaoyu opened 5 years ago

haichaoyu commented 5 years ago

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.

huochaitiantang commented 5 years ago

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.

haichaoyu commented 5 years ago

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.

huochaitiantang commented 5 years ago

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

haichaoyu commented 5 years ago

Thanks!

wrrJasmine commented 5 years ago

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