How to modify the structure to fit more than three channels of input pictures？ about deeplabv3plus-pytorch HOT 6 OPEN

Hi @pyh99715. In Deeplabv3-ResNet101, model.backbone.conv1 is the first conv layer. You can build the model directly and then replace model.backbone.conv1 with a new 4-channel Conv layers, i.e., nn.Conv2d(4, ...).

from deeplabv3plus-pytorch.

I have tried to find the first layer of the network structure before(resnet101 for example), but I can't seem to find the location that defines the first layer, is it in network/backbone/resnet.py line 48?
I am new to this field, and some questions may be relatively stupid, please forgive me :p

from deeplabv3plus-pytorch.

The first layer is defined in line 139 of ResNet.

from deeplabv3plus-pytorch.

I'm very sorry, I found it two minutes after I replied to you just now. I looked for it in the code of _deeplab.py for a long time. I am not familiar with this aspect. 💇‍♂ Then if resnet101 can take this form to change to four-channel input, others such as resnet50, mobilenet, etc. can also modify conv1 in this way to modify the entire network ?

from deeplabv3plus-pytorch.

Yes, you can change the first layer to match your inputs.
But, the operation for mobilenet is more complicated because it does not have a separate conv1. You need to modify the first element (L119) of model.backbone.features .

from deeplabv3plus-pytorch.

Thank you so much for your help, I even emailed you for help. I think this project is not the same as many similar projects. It is more complete and efficient. May I ask if you implemented it yourself?
This is really an admirable job 👍 👍 👍

from deeplabv3plus-pytorch.