search

pkuCactus / BDCN

The code for the CVPR2019 paper Bi-Directional Cascade Network for Perceptual Edge Detection
MIT License
341 stars 71 forks source link

Why batch_size =1 ? #36

Closed Dreamerswile closed 4 years ago

Dreamerswile commented 4 years ago

I would like to ask why batch_size=1?I believe this may be the best result of tuning parameters, Is it because the model has an image with a gray background?

pkuCactus commented 4 years ago

We do not crop the train data to fixed size when training, the size are different that can't batch to together, so we forward one image each iteration. Here, we adopt the itersize=10 which analogs to batchsize=10, ie, forward 10 times and then update the weights.

Dreamerswile commented 4 years ago

I couldn't agree with you more.But why did you choose this way instead of resizing the image.Cropping the image size also increases the batchsize and reduces training time.Is the method you used the best result of the experiment?Thank you very much for your answer