Deeplabv3_mobilenet_large with 4 input channels

[JLinden]

I’m trying to train DeepLabV3 with a MobileNetV3 backbone on a dataset with four channels. What is the preferred way of modifying the backbone to perform this? There does not seem to be a parameter for setting the number of input channels. I can not find any as straight forward approach as when using ResNet as the backbone:

Train Deeplabv3 on 4 channels images

Hi there, i want to train deeplabV3 on my own Dataset with 4 channels images. but i didn’t find any PyTorch implementation of deeplabV3 where i could change parameters and input channels number of the model to fit my (4channels) images . How can i modify deeplabV3 to adapt it to my dataset?

The following solution works, but it does not feel right to change a protected attribute.

from functools import partial
from torch import nn
from torchvision import models
from torchvision.ops.misc import ConvNormActivation

model = models.segmentation.deeplabv3_mobilenet_v3_large(pretrained=False, num_classes=2)
model.backbone._modules['0'] = ConvNormActivation(4,
                                              16,
                                              kernel_size=3,
                                              stride=2,
                                              norm_layer=partial(nn.BatchNorm2d, eps=0.001, momentum=0.01),
                                              activation_layer=nn.Hardswish)

[JLinden]

I realized that model.backbone has a getter for the items in _modules, so I changed the code to:

model.backbone['0'] = ConvNormActivation(4,
                                         16,
                                         kernel_size=3,
                                         stride=2,
                                         norm_layer=partial(nn.BatchNorm2d, eps=0.001, momentum=0.01),
                                         activation_layer=nn.Hardswish)