not load optimizer for inference

Author: sergey.vilov

NNC/nn.py

@@ -266,24 +266,25 @@ def collate_fn(data):
 
 optimizer = torch.optim.AdamW(model_params, lr=input_params.learning_rate, weight_decay=input_params.weight_decay) #define optimizer
 
+last_epoch = -1
+
 if input_params.load_weights:
 
     if torch.cuda.is_available():
         #load on gpu
         model.load_state_dict(torch.load(input_params.config_start_base + '_model'))
-        optimizer.load_state_dict(torch.load(input_params.config_start_base + '_optimizer'))
+        if not input_params.inference_mode:
+            optimizer.load_state_dict(torch.load(input_params.config_start_base + '_optimizer'))
     else:
         #load on cpu
         model.load_state_dict(torch.load(input_params.config_start_base + '_model', map_location=torch.device('cpu')))
-        optimizer.load_state_dict(torch.load(input_params.config_start_base + '_optimizer', map_location=torch.device('cpu')))
+        if not input_params.inference_mode:
+            optimizer.load_state_dict(torch.load(input_params.config_start_base + '_optimizer', map_location=torch.device('cpu')))
 
     last_epoch = int(input_params.config_start_base.split('_')[-2]) #infer previous epoch from input_params.config_start_base
 
-else:
-
-    last_epoch = -1
-
-lr_scheduler = torch.optim.lr_scheduler.MultiStepLR(optimizer,
+if not input_params.inference_mode:
+    lr_scheduler = torch.optim.lr_scheduler.MultiStepLR(optimizer,
                                                             milestones=[input_params.lr_sch_milestones],
                                                             gamma=input_params.lr_sch_gamma,
                                                             last_epoch=last_epoch, verbose=False) #define learning rate scheduler