Why Resnet first conv layer has kernel_size=3 and not =7, like the original paper?

[acnazarejr]

https://github.com/grib0ed0v/face_recognition.pytorch/blob/b5c4be96d217619b472630c455155475e6a923b3/model/backbones/resnet.py#L24

[grib0ed0v]

@acnazarejr actually, 7x7 convolution is much heavier than 3x3, that's why we decided to choose 3x3. Since we don't use pre-trained weights from ImageNet, that's should be OK.