Siamese Network training

vision
1

I am looking for a way to train a Correlation based Siamese Network, while training i get an error saying that the target size must be the same as the input size. I wanted to ask that is there a way to set up my labels for the training data in a way that it would match the input size.

2

Could you post some information regarding your model architecture, used loss function, etc.
I’m currently not sure, what your use case exactly is and where the error is being thrown.

3

model1.train()
model2.train()

batch_size = 5
epochs = 2

for epoch in range(epochs):
for i in range(0, len(X1), batch_size):
batch_X1 = X1[i:i+batch_size].view(-1, 1, 300, 300)
batch_X2 = X2[i:i+batch_size].view(-1, 1, 150, 150)
batch_y = y[i:i+batch_size]
#print(“check”)
optimizer1.zero_grad()
optimizer2.zero_grad()
#print(“check”)

    feature1 = model1(batch_X1)
    feature2 = model2(batch_X2)
    
    b, c, h, w = feature1.shape
    print("cc:",feature1.view(1,b*c,h,w).shape)
    
    corr_map = F.conv2d(feature1.view(1,b*c,h,w), feature2, groups=b)
    #print("check")
    #loss1 = loss_function(feature1,feature2)
    loss = loss_function(corr_map, batch_y)
    
    #loss = loss1 + loss2
    #print("check")
    loss.backward()
    optimizer1.step()
    optimizer2.step()
    
ep = epoch+1
print(f"Epoch: {ep}. Loss: {loss}")

this is the training code. In this code the shape of corr_map is [1,5,38,38] and the size of batch_y is [1] and i am not able to calculate the loss between them because the size is not same. I wanted to know if there is a way i could overcome this problem. If there is a way through which the size of the target equals the size of the input plz enlighten me.

5

What loss function are you using and how would you like to calculate the loss between a 4-dimensional tensor and a scalar?