Hi,
I’m looking for a variant of CosineEmbeddingLoss in which instead of using the cosine distance I would like to use the squared Euclidean distance.
Is it available in Pytorch?
Thank you 
PS: I tried to implement a very first implementation, but it is not vectorized(hence not computational efficient)
Are you looking for MSELoss?
Not exactly; I would like a version of MSELoss that behaves according to the margin and tensor y of similarities, as described in CosineEmbeddingLoss(i.e. exactly the same behavior described at this link, but using mean squared error instead of cosine distance to compute distances).
Maybe I can do in this way:
import torch
import torch.nn as nn
import torch.nn.functional as F
class MarginMSELoss(nn.Module):
def __init__(self, margin, reduction='mean'):
super(MarginMSELoss, self).__init__()
self.margin = margin
self.reduction = reduction
def forward(self, x1, x2, y):
# x1.shape == x2.shape == (batch_size, num_features)
# y.shape == (batch_size,)
dist = F.mse_loss(x1, x2, reduction='none').mean(dim=1) # dist.shape == (batch_size,)
mask = y == 1
out1 = (1 - dist) * mask # out1.shape == (batch_size,)
mask = y != 1
out2 = (torch.max(torch.zeros_like(dist), dist - self.margin)) * mask
out = out1 + out2
if self.reduction == 'none':
return out
elif self.reduction == 'sum':
return out.sum()
elif self.reduction == 'mean':
return out.mean()
else:
raise Exception('Reduction ' + self.reduction + ' is not supported')
>>> loss = MarginMSELoss(1)
>>> x1 = torch.randn(3, 5, requires_grad=True)
>>> x2 = torch.randn(3, 5, requires_grad=True)
>>> y = torch.randint(0, 2, (3,))
>>> output = loss(x1, x2, y)
>>> output.backward()
Thank you for your patience