bug fix for Trainer fp16

fastNLP/core/callback.py

@@ -482,7 +482,10 @@ class GradientClipCallback(Callback):
         if self.step%self.update_every==0:
             if self.trainer.fp16:
                 self.grad_scaler.unscale_(self.optimizer)
-            self.clip_fun(self.parameters, self.clip_value)
+            if self.parameters is not None:
+                self.clip_fun(self.parameters, self.clip_value)
+            else:
+                self.clip_fun(self.model.parameters(), self.clip_value)
 
 
 class EarlyStopCallback(Callback):

fastNLP/core/trainer.py

@@ -526,6 +526,7 @@ class Trainer(object):
 
         # check fp16相关的设置
         self.auto_cast, _grad_scaler = _build_fp16_env(dummy=not fp16)
+        self.grad_scaler = _grad_scaler()
         if self.fp16:
             _can_use_fp16(device=device, model=model, func=self._forward_func)
             grad_scaler = kwargs.get('grad_scaler', None)