ResNet-D_PyTorch

PyTorch implementation of ResNet-D from "Bag of Tricks for Image Classification with Convolutional Neural Networks" (He et al., CVPR2019)

paper: https://arxiv.org/pdf/1812.01187

summary (KOR): https://bo-10000.tistory.com/133

Models available

resnetD.py:

• resnet18
• resnet34
• resnet50
• resnet101
• resnet152
• resnet50D
• resnet101D
• resnet152D

resnetD_3d.py:

• resnet3d18
• resnet3d34
• resnet3d50
• resnet3d101
• resnet3d152
• resnet3d50D
• resnet3d101D
• resnet3d152D

How to use

#2D models
from resnetD import *

model = resnet50D()

#3D models
from resnetD_3d import *

model = resnet3d50D()

Model parameters

• in_channels (int, defauult=3): input channel dimension
• num_classes (int, default=1000): number of class labels
• zero_init_residual (bool, default=False): Zero-initialize the last BN in each residual branch, so that the residual branch starts with zeros, and each residual block behaves like an identity.
• groups (int, default=1): number of groups for conv3x3
• width_per_group (int, default=64): to change model width
• replace_stride_with_dilation (list of bools, default=None): each element in the tuple indicates if we should replace the 2x2 stride with a dilated convolution instead
• norm_layer (nn.Module, default=None): norm layer to use