fix reduce master grad casting back (#2027)

PaddleCV/image_classification/utils/fp16_utils.py

@@ -103,8 +103,16 @@ def create_master_params_grads(params_grads, main_prog, startup_prog, scale_loss
 
 def master_param_to_train_param(master_params_grads, params_grads, main_prog):
     for idx, m_p_g in enumerate(master_params_grads):
-        train_p, _ = params_grads[idx]
-        if train_p.name.startswith("batch_norm"):
-            continue
         with main_prog._optimized_guard([m_p_g[0], m_p_g[1]]):
+            train_p_name = m_p_g[0].name.replace(".master", "")
+            if train_p_name.startswith("batch_norm"):
+                continue
+            train_p = None
+            # find fp16 param in original params_grads list
+            for p, g in params_grads:
+                if p.name == train_p_name:
+                    train_p = p
+            if not train_p:
+                print("can not find train param for: ", m_p_g[0].name)
+                continue
             cast_fp32_to_fp16(m_p_g[0], train_p, main_prog)