Closed p16i closed 2 years ago
Hi @heytitle thanks for the question! I'll pass it on to the author, @guarin 🙂
I think this is indeed a mistake. BarlowTwins works also for small images.
Note that papers like SwaV use additional smaller crops. Since we already have 32x32 images in cifar10 it makes sense to not use smaller crops there. Most of the papers compare benchmarks on the ImageNet dataset with 224x224 input resolution. So if you have larger input images I'd suggest using a larger resolution as well if possible :)
Should the input_size in [1] be
32
instead of224
? In [2], we useinput_size=32
.[1] https://github.com/lightly-ai/lightly/blob/master/examples/pytorch_lightning/barlowtwins.py#L44 [2] https://github.com/lightly-ai/lightly/blob/master/examples/pytorch/barlowtwins.py#L35