Hi: I've successfully created an optimization framework using torch, but I'm not sure how to implement the loss function: I have a tensor containing predicted values (x, yhat), and I have a tensor containing actual values (x, y): I need to effectively join these (on x) and then aggregate across the "rows" as SUM((yhat - y)^2) . It's not obvious to me how to do this efficiently: I've been able to use torch_cat to implement the join, so I now have columns x, y and yhat (although many of the y values are 'nan' due to missing data. (But I had to achieve this by converting both actual and predicted tensors to data frames, then join them and then extract the joined yhat column as a tensor to concatenate onto the predicted tensor, which doesn't seem very efficient). And now I don't know whether I can do better than explicitly looping over every "row" of my "joined" tensor and incrementally accumulating the sum of squares. Surprisingly there seems no built-in or package functions available for this sort of operation, although I'm sure I can't be the first person to want to do it! Any help gratefully appreciated! (Note: I need to implement the RSS via tensor operations so that the autograd works.)