diff --git a/codes/models/SR_model.py b/codes/models/SR_model.py
index 5226715b..bf46ea3f 100644
--- a/codes/models/SR_model.py
+++ b/codes/models/SR_model.py
@@ -8,6 +8,7 @@ import models.networks as networks
 import models.lr_scheduler as lr_scheduler
 from .base_model import BaseModel
 from models.loss import CharbonnierLoss
+from apex import amp
 
 logger = logging.getLogger('base')
 
@@ -23,7 +24,7 @@ class SRModel(BaseModel):
         train_opt = opt['train']
 
         # define network and load pretrained models
-        self.netG = networks.define_G(opt).to(self.device)
+        self.netG = amp.initialize(networks.define_G(opt).to(self.device), opt_level=self.amp_level)
         if opt['dist']:
             self.netG = DistributedDataParallel(self.netG, device_ids=[torch.cuda.current_device()])
         elif opt['gpu_ids'] is not None: