Hello! First of all thanks for this amazing work.
I am trying to use fine tuning in your framework. I want to use a reduced size dataset (<1k), from a different domain and grayscale images.
I already create my data loader (similar to your cityscapes) where I create labelids masks with 2 classes (out of roi, roi, class1 and class2). I want to use your framework for classification and segmentation tasks.
My question is regarding the number of channels, how can I change the 3 channels network to 1 channel (because I want to use grayscale images) and use your pretrainned model for fine tuning? Can you give any advice on this?
I am thinking on duplicate the image into 3 channels. I am not sure if is the best approach and the implications in overall network. What your opinion about this approach?
Other option I think is to hardcode all networks to accept 1 channel in the 1st conv layer and to be able to use the pretrainned model I could sum the weights of that 1st layer (3 channels) to get just 1 channel. But I am not sure if is possible for all networks.
Hello! First of all thanks for this amazing work. I am trying to use fine tuning in your framework. I want to use a reduced size dataset (<1k), from a different domain and grayscale images. I already create my data loader (similar to your cityscapes) where I create labelids masks with 2 classes (out of roi, roi, class1 and class2). I want to use your framework for classification and segmentation tasks.