closure

README.md

@@ -19,6 +19,8 @@ Some optimizers require closure, which should look like this:
 def closure(backward = True):
   preds = model(inputs)
   loss = loss_fn(preds, targets)
+
+  # skip this part if you have no gradients and use zeroth order methods, but keep the unused backward argument.
   if backward:
     optimizer.zero_grad()
     loss.backward()