Update model.py

model.py

@@ -4,7 +4,7 @@ Implementation of YOLOv3 architecture
 
 import torch
 import torch.nn as nn
-import config as cfg
+# import config as cfg
 """ 
 Information about architecture config:
 Tuple is structured by (filters, kernel_size, stride) 
@@ -107,8 +107,8 @@ class YOLOv3(nn.Module):
         self.in_channels = in_channels
         self.layers = self._create_conv_layers()
         self.base_model = None
-        self.distill_feature = cfg.DISTILL
-        self.warp = cfg.WARP
+        # self.distill_feature = cfg.DISTILL
+        # self.warp = cfg.WARP
         self.feature_store = None 
         self.enable_warp_train = False
 
@@ -198,42 +198,42 @@ class YOLOv3(nn.Module):
 
 
 
-if __name__ == "__main__":
-    num_classes = 19
-    IMAGE_SIZE = 416
-    model = YOLOv3(num_classes=num_classes)
-    # print(model)
-    print(model.layers[15].pred[1].conv.weight.shape)
-    print(model.layers[15].pred[1].conv.bias.shape)
-    import torch.optim as optim
-    optimizer = optim.Adam(
-        model.parameters(), lr=cfg.LEARNING_RATE, weight_decay=cfg.WEIGHT_DECAY
-    )
-    from utils import load_checkpoint
-    load_checkpoint(
-            cfg.BASE_CHECK_POINT, model, optimizer, cfg.LEARNING_RATE
-    )
-
-    model.adaptation(layer_id = 15, num_class = 20, in_feature = 1024, old_class = num_classes)
-    model.adaptation(layer_id = 22, num_class = 20, in_feature = 512, old_class = num_classes)
-    model.adaptation(layer_id = 29, num_class = 20, in_feature = 256, old_class = num_classes) 
-    # layer1 = 
-    # model.eval()
-    # with torch.no_grad():
-    #     old_weight = model.layers[15].pred[1].conv.weight
-    #     old_bias = model.layers[15].pred[1].conv.bias
-    #     # print(model.layers[22].pred[1])
-    #     # print(model.layers[29].pred[1])
-    #     # out_dims = cfg.BASE_CLASS + cfg.NEW_CLASS + 5
-    #     model.layers[15].pred[1] = CNNBlock(1024, 25 * 3, bn_act=False, kernel_size=1)
-    #     model.layers[15].pred[1].conv.weight[:72] = old_weight
-    #     model.layers[15].pred[1].conv.bias[:72] = old_bias
-    print(model.layers[15].pred[1].conv.weight.shape)
-    # model.layers[22].pred[1] = CNNBlock(512, out_dims * 3,  kernel_size=1)
-    # model.layers[29].pred[1] = CNNBlock(256, out_dims * 3,  kernel_size=1)
-    x = torch.randn((2, 3, IMAGE_SIZE, IMAGE_SIZE))
-    out = model(x)
-    # assert model(x)[0].shape == (2, 3, IMAGE_SIZE//32, IMAGE_SIZE//32, num_classes + 5)
-    # assert model(x)[1].shape == (2, 3, IMAGE_SIZE//16, IMAGE_SIZE//16, num_classes + 5)
-    # assert model(x)[2].shape == (2, 3, IMAGE_SIZE//8, IMAGE_SIZE//8, num_classes + 5)
-    print("Success!")
+# if __name__ == "__main__":
+#     num_classes = 19
+#     IMAGE_SIZE = 416
+#     model = YOLOv3(num_classes=num_classes)
+#     # print(model)
+#     print(model.layers[15].pred[1].conv.weight.shape)
+#     print(model.layers[15].pred[1].conv.bias.shape)
+#     import torch.optim as optim
+#     optimizer = optim.Adam(
+#         model.parameters(), lr=cfg.LEARNING_RATE, weight_decay=cfg.WEIGHT_DECAY
+#     )
+#     from utils import load_checkpoint
+#     load_checkpoint(
+#             cfg.BASE_CHECK_POINT, model, optimizer, cfg.LEARNING_RATE
+#     )
+
+#     model.adaptation(layer_id = 15, num_class = 20, in_feature = 1024, old_class = num_classes)
+#     model.adaptation(layer_id = 22, num_class = 20, in_feature = 512, old_class = num_classes)
+#     model.adaptation(layer_id = 29, num_class = 20, in_feature = 256, old_class = num_classes) 
+#     # layer1 = 
+#     # model.eval()
+#     # with torch.no_grad():
+#     #     old_weight = model.layers[15].pred[1].conv.weight
+#     #     old_bias = model.layers[15].pred[1].conv.bias
+#     #     # print(model.layers[22].pred[1])
+#     #     # print(model.layers[29].pred[1])
+#     #     # out_dims = cfg.BASE_CLASS + cfg.NEW_CLASS + 5
+#     #     model.layers[15].pred[1] = CNNBlock(1024, 25 * 3, bn_act=False, kernel_size=1)
+#     #     model.layers[15].pred[1].conv.weight[:72] = old_weight
+#     #     model.layers[15].pred[1].conv.bias[:72] = old_bias
+#     print(model.layers[15].pred[1].conv.weight.shape)
+#     # model.layers[22].pred[1] = CNNBlock(512, out_dims * 3,  kernel_size=1)
+#     # model.layers[29].pred[1] = CNNBlock(256, out_dims * 3,  kernel_size=1)
+#     x = torch.randn((2, 3, IMAGE_SIZE, IMAGE_SIZE))
+#     out = model(x)
+#     # assert model(x)[0].shape == (2, 3, IMAGE_SIZE//32, IMAGE_SIZE//32, num_classes + 5)
+#     # assert model(x)[1].shape == (2, 3, IMAGE_SIZE//16, IMAGE_SIZE//16, num_classes + 5)
+#     # assert model(x)[2].shape == (2, 3, IMAGE_SIZE//8, IMAGE_SIZE//8, num_classes + 5)
+#     print("Success!")