Larger input size

[muminoff]

Can you show us a quick way how to customize the network to use larger input size?
I am willing to test it with 1024x1024 size.

[sdsubhajitdas]

In the bts package inside dataset.py you can see the TumorDataset class.
Inside TumorDataset's constructor you can see there are some default transformations

self.default_transformation = transforms.Compose([
            transforms.Grayscale(),
            transforms.Resize((512, 512))
        ])

You can change the Resize transformation size

self.default_transformation = transforms.Compose([
            transforms.Grayscale(),
            transforms.Resize((1024, 1024))
        ])

I think this will do the trick. After creating the model you can call the summary function on the model object the input size parameter u want. That will show you the sizes in each step.

Inside model.py DynamicUNet class.

def summary(self, input_size=(1, 512, 512), batch_size=-1, device='cuda'):
        """ Get the summary of the network in a chart like form
        with name of layer size of the inputs and parameters 
        and some extra memory details.
        This method uses the torchsummary package.
        For more information check the link.
        Link :- https://github.com/sksq96/pytorch-summary
        Parameters:
            input_size(tuple): Size of the input for the network in
                                 format (Channel, Width, Height).
                                 Default: (1,512,512)
            batch_size(int): Batch size for the network.
                                Default: -1
            device(str): Device on which the network is loaded.
                            Device can be 'cuda' or 'cpu'.
                            Default: 'cuda'
        Returns:
            A printed output for IPython Notebooks.
            Table with 3 columns for Layer Name, Input Size and Parameters.
            torchsummary.summary() method is used.
        """

[sdsubhajitdas]

Let me know if this works.

[muminoff]

@sdsubhajitdas Thanks for the detailed info and apologizes for late response.

How about training process?
Can you suggest how to do it?
There is no instruction on readme.
AFAIK api.py file only provides prediction operation.

[muminoff]

Never mind. Just discovered it in notebook.