Help
Python API Errors
Gradient Operator Checks
Example error message:
“Gradient operator needs output “gpu_0/tmp2” at version 0, but currently we have version 1”
This is also generally seen as some variation of:
Gradient name X is expected to correspond to version Y of X, but currently we have version Z.
or
Gradient output X is expected to correspond to version Y of X, but currently we have version Z.
or
Gradient input X is expected to correspond to version Y of X, but currently we have version Z.
Check if your model uses in-place blobs and that blob is passed forward to another operator. It cannot keep track of the input needed for the gradient for the first operator. If you attempt a backward pass, it needs to have the original input to create the gradients. This message basically tells you that you should not do an in-place op for “gpu0/tmp2”. Once that is addressed, you should be able to generate the backward pass.
Operator Usage
What’s the Difference between net.Sum() and net.Add()?
net.Sum()allows multiple inputs, but does not allow broadcasting.
net.Add()allows only two inputs, but the second one can be broadcasted to match the shape of the first one.