diff --git a/RARP_ROI_Mask.ckpt b/RARP_ROI_Mask.ckpt
new file mode 100644
index 0000000..0fd7f76
--- /dev/null
+++ b/RARP_ROI_Mask.ckpt
Binary files differ
diff --git a/defs.py b/defs.py
index 1b1b694..21aa8cf 100644
--- a/defs.py
+++ b/defs.py
@@ -12,7 +12,7 @@
     return torch.from_numpy(np.load(path).astype(float).transpose((2, 0, 1))) / 255.0
 
 def load_Img(path):
-    cv2.imread(str(path), cv2.IMREAD_COLOR)
+    return cv2.imread(str(path), cv2.IMREAD_COLOR)
     
 def clip_gradients(model, clip=2.0):
     """Rescale norm of computed gradients.
diff --git a/imageViewer.py b/imageViewer.py
index aec8bd1..db1c1b3 100644
--- a/imageViewer.py
+++ b/imageViewer.py
@@ -2,6 +2,7 @@
 import numpy as np
 from ultralytics import YOLO as YOLOv8
 import PySimpleGUI as sg
+from pathlib import Path
 
 def RemoveBlackBorder(image, shrink=True):
     image = np.array(image)
@@ -116,7 +117,7 @@
                     cv2.waitKey(0)
                     cv2.destroyAllWindows() 
                 case "-verImg-":
-                    image = np.load(vals["-Path-"]) if vals["-NP_Array_IMG-"] else cv2.imread(vals["-Path-"], cv2.IMREAD_COLOR)
+                    image = np.load(Path(vals["-Path-"])) if vals["-NP_Array_IMG-"] else cv2.imread(vals["-Path-"], cv2.IMREAD_COLOR)
                     cv2.namedWindow("Output", cv2.WINDOW_NORMAL)
                     cv2.resizeWindow("Output", (1024, 720))
                     cv2.imshow("Output", image)
diff --git a/noah.py b/noah.py
new file mode 100644
index 0000000..4adcf25
--- /dev/null
+++ b/noah.py
@@ -0,0 +1,50 @@
+import torch
+import torch.nn as nn
+
+
+class NOAH(nn.Module):
+    def __init__(self, inplanes, outplanes, dropout=0.0, key_ratio=0.5, head_num=1, head_split=True, kv_split=True):
+        super(NOAH, self).__init__()
+        self.kv_split = kv_split
+        self.head_split = head_split
+        self.dropout = nn.Dropout(p=dropout)
+        self.key_ratio = key_ratio
+        self.head_num = head_num
+
+        if kv_split:
+            self.k_channel = int(inplanes * key_ratio)
+            self.v_channel = inplanes - self.k_channel
+        else:
+            self.k_channel = inplanes
+            self.v_channel = inplanes
+
+        assert self.k_channel % head_num == 0
+        assert self.v_channel % head_num == 0
+
+        self.groups = head_num if head_split else 1
+        self.query = nn.Conv2d(self.k_channel, head_num * outplanes, kernel_size=1, groups=self.groups,
+                               stride=1, padding=0)
+        self.value = nn.Conv2d(self.v_channel, head_num * outplanes, kernel_size=1, groups=self.groups,
+                               stride=1, padding=0)
+        #self._init_weight()
+
+    def _init_weight(self):
+        for m in self.modules():
+            if isinstance(m, nn.Conv2d):
+                nn.init.kaiming_normal_(m.weight, mode='fan_out')
+                if m.bias is not None:
+                    nn.init.zeros_(m.bias)
+
+    def forward(self, x):
+        x = torch.flatten(x, 2).unsqueeze(dim=-2)
+        N, C, _, L = x.shape
+        if self.kv_split:
+            a = torch.softmax(self.query(x[:, :self.k_channel]).reshape(N, self.head_num, -1, L), dim=3)
+            v = self.value(x[:, self.k_channel:]).reshape(N, self.head_num, -1, L)
+        else:
+            a = torch.softmax(self.query(x).reshape(N, self.head_num, -1, L), dim=3)
+            v = self.value(x).reshape(N, self.head_num, -1, L)
+        v = self.dropout(v)
+        x = torch.sum(a * v, dim=(1, 3))
+        return x
+