Can you please explain how the network learns 'm' prototypes (initially)?

[FahadMahdi]

After the training is completed, I understand the process. However, I cannot understand how the prototypes are learned at the beginning? Is it user defined? The paper doesn't say so.

[giangnguyen2412]

I have the same question :D it seems to be very important but how the prototypical parts were learned may be neglected. @cfchen-duke can you give us a high-level intuition? I think this is explained in the paper but not easy to understand and a high-level explanation is very useful. Thank you.

[kretes]

After reading the paper I understood that the prototypes are 'learned', and so the initialization is done randomly.
In the supplement Figure 10: Overview of training algorithm this is described as ∀j: prototype pj ←Uniform([0,1]H1×W1×D);
Looking at the model code I found this https://github.com/cfchen-duke/ProtoPNet/blob/master/model.py#L105 which confirms this hypothesis.

[boileddd]

self.prototype_vectors = nn.Parameter(torch.rand(self.prototype_shape), requires_grad=True)
line 105 in model.py