search

davidtvs / PyTorch-ENet

PyTorch implementation of ENet
MIT License
383 stars 129 forks source link

Training on Custom Dataset where height/width are not powers of 2 #26

Closed erosten closed 5 years ago

erosten commented 5 years ago

Hi, I'm trying to train on a custom dataset of size 300x300, and errors are being thrown in the second DownSamplingBottleneck as when concatenating the size of main and ext, there's a size mismatch.

This is due to the fact that after the first downsampling bottleneck, 300x300 becomes 150x150 for both main and ext.

During the second downsampling bottleneck, main becomes 38 since (75 + 2x1 - 1x(3-1) -1)/2 + 1 = 74/2 + 1 = 38.

However, ext becomes (75-0x1 - 1x(2-1) -1)/2 + 1 = 73/2 + 1 = 37.5, which is floored.

Do you have any suggestions to get around this? I'm relatively new to pytorch.

davidtvs commented 5 years ago

Ya, I think that's a limitation of the original model architecture from the paper.

Check out the any_input_size branch. I changed the architecture slightly to get around this problem. I don't have performance numbers for that branch though, never got around to train the network

erosten commented 5 years ago

I ended up just resizing to 256x256, but will check out that branch at some point. Thanks for the response!