VainF / DeepLabV3Plus-Pytorch

Pretrained DeepLabv3 and DeepLabv3+ for Pascal VOC & Cityscapes

Geek Repo:Geek Repo

Github PK Tool:Github PK Tool

How to change input channels

Shaochen7 opened this issue · comments

I want to train five channels at the same time, defaulting to RGB three channels. So how can I modify the input channels?

Hello @Shaochen7, the input channel is defined in backbones, e.g. ResNet

If you need to train a model that can handle 5-channel images, you can modify the first layer as follows:

Ln44 modeling.py

    backbone = resnet.__dict__[backbone_name](
        pretrained=pretrained_backbone,
        replace_stride_with_dilation=replace_stride_with_dilation)
    backbone.conv1 = nn.Conv2d(5, self.conv1.out_channels, ...) #<= insert this line 
    inplanes = 2048
    low_level_planes = 256