update

core/model.py

@@ -815,7 +815,22 @@ def erase_feature_maps(self, atten_map_normed, feature_maps, threshold, flag=Fal
         return erased_feature_maps
 
     def fix_params(self, is_training=True):
-        pass
+        for p in self.backbone.parameters():
+            p.requires_grad = is_training
+        for p in self.erase.parameters():
+            p.requires_grad = is_training
+        for p in self.fc7.parameters():
+            p.requires_grad = is_training
+        for p in self.cls.parameters():
+            p.requires_grad = is_training
+        for p in self.cls_direction.parameters():
+            p.requires_grad = is_training
+        for p in self.erase_fc7.parameters():
+            p.requires_grad = is_training
+        for p in self.erase_cls.parameters():
+            p.requires_grad = is_training
+        for p in self.erase_cls_direction.parameters():
+            p.requires_grad = is_training
 
     def get_loss(self, logits, labels, direction):
 

finetune_tiger_cnn8.py

@@ -54,7 +54,13 @@ def main():
     feature_size = 1024
 
     net = tiger_cnn8(classes=107)
-    ignore_params = list(map(id, net.fuse_fc7.parameters()))
+    ignore_params = list(map(id, net.fc7.parameters()))
+    ignore_params += list(map(id, net.cls.parameters()))
+    ignore_params += list(map(id, net.cls_direction.parameters()))
+    ignore_params += list(map(id, net.erase_fc7.parameters()))
+    ignore_params += list(map(id, net.erase_cls.parameters()))
+    ignore_params += list(map(id, net.erase_cls_direction.parameters()))
+    ignore_params += list(map(id, net.fuse_fc7.parameters()))
     ignore_params += list(map(id, net.fuse_cls.parameters()))
     ignore_params += list(map(id, net.fuse_cls_direction.parameters()))
     base_params = filter(lambda p: id(p) not in ignore_params, net.parameters())
@@ -67,7 +73,7 @@ def main():
     exp_lr_scheduler = StepLRScheduler(optimizer=optimizer, decay_t=20, decay_rate=0.1, warmup_lr_init=1e-5, warmup_t=3)
 
     net.load_state_dict(torch.load('./model/tiger_cnn3/model.ckpt')['net_state_dict'])
-    net.fix_params(is_training=False)
+    # net.fix_params(is_training=False)
     net = net.cuda()
     if multi_gpus:
         net = nn.DataParallel(net)