Simplify the gradient clipping code. (#4896)

Author: datumbox

references/classification/train.py

@@ -40,13 +40,13 @@ def train_one_epoch(model, criterion, optimizer, data_loader, device, epoch, arg
             if args.clip_grad_norm is not None:
                 # we should unscale the gradients of optimizer's assigned params if do gradient clipping
                 scaler.unscale_(optimizer)
-                nn.utils.clip_grad_norm_(utils.get_optimizer_params(optimizer), args.clip_grad_norm)
+                nn.utils.clip_grad_norm_(model.parameters(), args.clip_grad_norm)
             scaler.step(optimizer)
             scaler.update()
         else:
             loss.backward()
             if args.clip_grad_norm is not None:
-                nn.utils.clip_grad_norm_(utils.get_optimizer_params(optimizer), args.clip_grad_norm)
+                nn.utils.clip_grad_norm_(model.parameters(), args.clip_grad_norm)
             optimizer.step()
 
         if model_ema and i % args.model_ema_steps == 0:

references/classification/utils.py

@@ -409,11 +409,3 @@ def reduce_across_processes(val):
     dist.barrier()
     dist.all_reduce(t)
     return t
-
-
-def get_optimizer_params(optimizer):
-    """Generator to iterate over all parameters in the optimizer param_groups."""
-
-    for group in optimizer.param_groups:
-        for p in group["params"]:
-            yield p