search

yun-liu / DEL

DEL: Deep Embedding Learning for Efficient Image Segmentation
Other
61 stars 15 forks source link

image resize #8

Open ShiAngWang opened 2 years ago

ShiAngWang commented 2 years ago

Hi， while training， have you resized the images into a specific size or used the original image without transformation? If you resized the images, can you show where the code is?

yun-liu commented 2 years ago

We just used the original image size without resizing. To do this, we used a batch size of 1 (code) and set the iter_size in Caffe to 5 (code). In this way, the gradients during training are averaged among 5 images, similar to the effect of the batch size of 5 (we don't use batchnorm, so iter_size=5 euqals to batch size=5 here).