Ken

classification/main_train.py

@@ -307,6 +307,8 @@ def main():
         model = DenseNet169(n_inputs, numCls)
     elif args.model == 'DenseNet201':
         model = DenseNet201(n_inputs, numCls)
+    elif args.model == 'DenseNet201':
+        model = DenseNet201(n_inputs, numCls)
     # finetune moco pre-trained model
     elif args.model.startswith("Moco"):
         pt_path = os.path.join(args.pt_dir, f"{args.pt_name}_{args.pt_type}_converted.pth")

classification/models/ResNet.py

@@ -310,8 +310,23 @@ def __init__(self, mocoModel, n_inputs = 12, numCls = 17):
         resnet.conv1 = nn.Conv2d(n_inputs, 64, kernel_size=(7, 7), stride=(2, 2), padding=(3, 3), bias=False)
         resnet.load_state_dict(mocoModel["state_dict"])
 
+<<<<<<< HEAD
+        # (testing) -- match the layer adding from moco's conv1 layer in ResNet
+        self.Conv1x1Block = nn.Sequential(
+            nn.Conv2d(n_inputs, 3, kernel_size=1, stride=1, bias=False),
+            nn.BatchNorm2d(3),
+            nn.ReLU(inplace=True)
+        )
+
+        # self.conv1 = nn.Conv2d(n_inputs, 64, kernel_size=(7, 7), stride=(2, 2), padding=(3, 3), bias=False)
+        self.conv1 = nn.Conv2d(3, 64, kernel_size=(256, 256), stride=(2, 2), padding=(3, 3), bias=False)
+        self.encoder = nn.Sequential(
+            self.Conv1x1Block,
+            self.conv1,
+=======
         self.encoder = nn.Sequential(
             resnet.conv1,
+>>>>>>> origin/master
             resnet.bn1,
             resnet.relu,
             resnet.maxpool,

classification/test.py

@@ -13,7 +13,7 @@
 
 from dataset import SEN12MS, ToTensor, Normalize
 from models.VGG import VGG16, VGG19
-from models.ResNet import ResNet50, ResNet101, ResNet152
+from models.ResNet import ResNet50, ResNet101, ResNet152, Moco
 from models.DenseNet import DenseNet121, DenseNet161, DenseNet169, DenseNet201
 from metrics import MetricTracker, Precision_score, Recall_score, F1_score, \
     F2_score, Hamming_loss, Subset_accuracy, Accuracy_score, One_error, \
@@ -25,6 +25,7 @@
 model_choices = ['VGG16', 'VGG19',
                  'ResNet50','ResNet101','ResNet152',
                  'DenseNet121','DenseNet161','DenseNet169','DenseNet201', 'Moco']
+
 label_choices = ['multi_label', 'single_label']
 
 # ------------------------ define and parse arguments -------------------------
@@ -158,7 +159,12 @@ def main():
     elif config['model'] == 'DenseNet169':
         model = DenseNet169(n_inputs, numCls)
     elif config['model'] == 'DenseNet201':
-        model = DenseNet201(n_inputs, numCls)                            
+        model = DenseNet201(n_inputs, numCls)
+
+    # add moco adjusted model - conv1 layer part
+    elif config['model'] == 'Moco':
+        model = Moco(n_inputs, numCls)
+
     else:
         raise NameError("no model")
 

evaluation_assessment.md

@@ -0,0 +1,19 @@
+1. The difference evaluation options:
+> - a. Land cover classification -- image classificaiton.
+> - b. Semantic Segmentation -- assigning a class label to every pixel of the input image.
+
+2. 
+
+3. tmp
+> a. SEN12MS - Supervised Learning Benchmark - Classification
+
+| Backbone  | Land Type  | Modalitities  | Bactch size  | Epochs | Accuracy (%) | Macro-F1 (%) | Micro-F1 (%) |
+|---|---|---|---|---|---|---|---|
+|DenseNet|single-label|_s1s2|64|100|51.16|50.78|62.90|
+|DenseNet|single-label|_s2|64|100|54.41|52.32|64.74|
+|ResNet50|single-label|_RGB|64|100|45.11|45.16|58.98|
+|ResNet50|single-label|_s1s2|64|100|45.52|53.21|64.66|
+|ResNet50|single-label|_s2|64|100|57.33|53.39|66.35|
+|ResNet50|multi-label|_RGB|64|100|89.86|47.57|66.51|
+|ResNet50|multi-label|_s1s2|64|100|91.22|57.46|71.40|
+|ResNet50|multi-label|_s2|64|100|90.62|56.14|69.88|