How to convert torch.norm to cosine distance

vision
1

I want to change norm distance to cosine distance, help me convert this function to cosine distance

def feat_prototype_distance(self, feat):
        N, C, H, W = feat.shape
        feat_proto_distance = -torch.ones((N, self.class_numbers, H, W)).to(feat.device)
        for i in range(self.class_numbers):
            feat_proto_distance[:, i, :, :] = torch.norm(self.objective_vectors[i].reshape(-1,1,1).expand(-1, H, W) - feat, 2, dim=1)
        return feat_proto_distance

This is original function using norm distance with shapes:
self.objective_vectors[i].reshape(-1,1,1).expand(-1, H, W): torch.Size([256, 128, 224])
feat: torch.Size([8, 256, 128, 224]) with 8 is batch_size

2 
def get_cosine_distance(x: Tensor, y: Tensor):
  """
  :param x: [B1, D]
  :param y: [B2, D]
  :return: [B1, B2]
  """
  x_norm = torch.norm(x, dim=1, keepdim=True)
  y_norm = torch.norm(y, dim=1, keepdim=True)
  return torch.mm(x, y.transpose(0, 1)) / x_norm / y_norm.transpose(0, 1)

get_cosine_distance(objective_vectors[i].reshape..., feat)

I think replacing the torch.norm part in your code with get_cosine_distance would resolve your issue.
Is it what you want?

3

I have tried but no success. Because feat_proto_distance[:, i, :, :] is 4 dimensional but it returns 2 dimensional

4

Help me! :frowning: :(((((((((((((((((((((((