Using other input/output size than the 320x320 used in the multi class segmentation example

[asmagen]

How do I extend the image input/output to anything greater than 320x320 which is currently used in the multi class segmentation tutorial in this package? When I simply changed the input size to 512 the run crashed, so I figured that the network should be configured to a different size before. How do I do that? And how is that related to the pretrained model? If I'm using the pretrained model in the multi class segmentation tutorial then do I have to use 320x320 sized images? What other pretrained models are available here, especially for pathology applications?

Thanks

[JordanMakesMaps]

The readme lists all of the architectures and all of the encoders. None of them are ready for use on histopathology images right out of the box because the pre-trained weights are for ImageNet. You'll need to either fine-tune a model with those weights, or train from scratch.

Changing the input/output to anything greater than 320x320 is shown in the docs:

model = sm.Unet(BACKBONE,  encoder_weights='imagenet', input_shape = (512, 512, 3)
# I personally would chose (None, None, 3) for the input shape

If your computer can't run anything larger than 320x320 because it crashes, that's a memory issue and means you need to get a different GPU that can allocate more memory, or just use what it can handle. If its not a memory issue, then you should show the error here otherwise no one will know what you mean by it 'crashing'.

[asmagen]

So the model definition in the example is not using any pretrained network (model = sm.Unet(BACKBONE, classes=n_classes, activation=activation)) and basically means I'mm training from scratch? And I just need to add encoder_weights='imagenet' to have it train starting with the imagenet-based weights which will further be optimized to my problem during training? I consulted with fellow pathology image analysis groups and they confirmed they also use imagenet because apparently there isn't a pathology-specific pretrained network.

[JordanMakesMaps]

"So the model definition in the example is not using any pretrained network (model = sm.Unet(BACKBONE, classes=n_classes, activation=activation)) and basically means I'mm training from scratch? And I just need to add encoder_weights='imagenet' to have it train starting with the imagenet-based weights which will further be optimized to my problem during training?"

Yes, and if you use one of the efficientnet encoders, look at using noisy-student weights, which come from the paper Self-training with Noisy Student improves ImageNet classification