Automatically figure out dimension of linear layer


I have a sequential layer with convolutions and max pooling, followed by a linear layer
The problem is, if I change the settings of convolutions or max pooling, the incoming dimension of the linear layer will change

What would be a good way to automatically check and find the right dimension, so the linear layer always works?

Thanks very much!
P.S. I tried LazyLinear but because the weights are uninitialized, I can’t initialize them with my initialize-weights function. So I’m trying to use a nn.Linear, which requires me to know the input dimension before creating it

1 Like

You could still use the lazy layer and initialize it afterwards once the parameters were created.

Thanks a lot for your help ptrblck! I see this in the error message as well. I think I will have an initialize method like this: (I’m not sure about the call to forward() since it’s supposed to be done by proxy of the nn.Module as a callable)

# inside the class hosting the model
initialize(self, dummy_input):
    self.forward(model) # perhaps I should have a compute() where I have all the computation? And then for my forward() I just call the compute()?
    self.network_1.apply(fn) # I have a bunch of networks that I can't put under the same nn.Sequential because I'm inserting extra calculation steps between their usage

Yes, having and internal initialize method should work, but I assume you want to pass the dummy_input to the model instead of the undefined model variable.
So something like this?

def initialize(self, dummy_input):

You don’t need to call self.forward and can just pass the input to self in the same way you would use model(input) instead of model.forward(input), although it will most likely not matter here.

1 Like

Thank you! I know what to do now!