Is ResNetV1e better than ResNetV1d?

deepinsight / insightface

State-of-the-art 2D and 3D Face Analysis Project

https://insightface.ai

22.84k stars 5.35k forks source link

Closed huangh12 closed 3 years ago

huangh12 commented 3 years ago

Hi @nttstar , as far as I know, this is the first time I've seen an e-version ResNet. Compared with ResNetV1d, it replaces the first 3x3 maxpool with 2x2. Did you observe any accuracy gain from that? Thank you! https://github.com/deepinsight/insightface/blob/54b97b0c23fdbd2083c91e54526a407037399958/detection/scrfd/mmdet/models/backbones/resnet.py#L702

nttstar commented 3 years ago

nn.MaxPool2d(kernel_size=3, stride=2, padding=1) in ResNet-D may lead to inconsistent behaviour while we convert it to caffemodel.

huangh12 commented 3 years ago

Thank you. Why the 3x3 maxpool will cause inconsistency when converting to caffemodel？

nttstar commented 3 years ago

related to ceil_mode