modified mnist

Author: wbw520

configs.py

@@ -28,8 +28,8 @@
 parser.add_argument('--start_epoch', default=0, type=int, metavar='N', help='start epoch')
 parser.add_argument('--lr', default=0.0001, type=float)
 parser.add_argument('--batch_size', default=128, type=int)
-parser.add_argument('--epoch', default=80, type=int)
-parser.add_argument('--lr_drop', default=60, type=float, nargs="+",
+parser.add_argument('--epoch', default=100, type=int)
+parser.add_argument('--lr_drop', default=80, type=float, nargs="+",
                     metavar='LRSteps', help='epochs to decay learning rate by 10')
 # ========================= Machine Configs ==========================
 parser.add_argument('--num_workers', default=4, type=int)

demo.py

@@ -0,0 +1 @@
+from flask import Flask

main_retri.py

@@ -19,7 +19,7 @@ def main():
         checkpoint = torch.load(os.path.join(args.output_dir,
             f"{args.dataset}_{args.base_model}_cls{args.num_classes}_cpt_no_slot.pt"), map_location=device)
         model.load_state_dict(checkpoint, strict=False)
-        fix_parameter(model, ["layer1", "layer2", "layer3", "back_bone.conv1", "back_bone.bn1"], mode="fix")
+        fix_parameter(model, ["layer1", "layer2", "back_bone.conv1", "back_bone.bn1"], mode="fix")
         print(colored('trainable parameter name: ', "blue"))
         print_param(model)
         print("load pre-trained model finished, start training")

process.py

@@ -36,5 +36,6 @@ def main():
     args = parser.parse_args()
     args.process = True
     args.pre_train = False
+    # args.batch_size = 1
     main()
 

utils/tools.py

@@ -69,10 +69,22 @@ def predict_hash_code(args, model, data_loader, device):
         data, label = data.to(device), label.to(device)
         if not args.pre_train:
             cpt, pred, att, update = model(data)
+            # if args.process:
+            #     att = att[0]
+            #     record_cpt = []
+            #     for i in range(args.num_cpt):
+            #         current_mean = att[i].mean()
+            #         incidence = att[i] > current_mean
+            #         new_att = att[i][incidence]
+            #         record_cpt.append(new_att.sum().unsqueeze(0)/len(incidence))
+            #     cpt = torch.cat(record_cpt, dim=0)
+            #     cpt = cpt.unsqueeze(0).squeeze(-1)
+
             acc = cal_acc(pred, label, False)
             accs += acc
         else:
             cpt = model(data)
+
         if is_start:
             all_output = cpt.cpu().detach().float()
             all_label = label.unsqueeze(-1).cpu().detach().float()