tabensemb.model.AbstractNN.on_before_backward

method

AbstractNN.on_before_backward(loss: Tensor) → None

Called before loss.backward().

Parameters:

loss – Loss divided by number of batches for gradient accumulation and scaled if using native AMP.