Using SupContrast loss for non-image application

[mossishahi]

Hey

I am about to use the SupConLoss for my specific application: As I am embedding some graphs and would like to make the embeddings of similar graphs to get more similar, taking advantage of this loss might be helpful. but I have faced some problems:

• the input of forward function should have a shape like: bsz, n_views, n_features however, I don't have copies for each embedding (equivalent to "images' crops"). So: my model's output has a shape like: (bsz, d_embedding) and I assumed I can match it to you loss function by unsqueeze(1) which makes it shaped like: (bsz, 1, d_embedding)
• The above mentioned input, results in nan As I have tracked the variables in your loss function:
• Also as I use mask, it has one and zero values. mask(i,j)=1 which means to me that the i'th and j'th embedding in my batch should get close.
  • In case of using 'label' argument, I am about to make those embeddings with same label to get close to each other. (mask is then automatically created)

Are these assumptions right?

appreciate

[meiling-fdu]

hello, have you solved this problem?