Upload Files

Net_tests/test_10x10_model.py

+import torch
+from torch import autograd, nn, optim
+import torch.nn.functional as F
+from tqdm import tqdm
+
+import matplotlib.pyplot as plt
+import matplotlib.image as mpimg
+import numpy as np
+
+class Net(nn.Module):
+    def __init__(self):
+        super(Net, self).__init__()
+        self.conv1 = nn.Conv2d(2, 32, 3)
+        self.conv2 = nn.Conv2d(32, 64, 3)
+        #self.dropout1 = nn.Dropout2d(0.25)
+        #self.dropout2 = nn.Dropout2d(0.5)
+        self.fc1 = nn.Linear(576, 512)
+        self.fc2 = nn.Linear(512, 100)
+
+    # x represents our data
+    def forward(self, x):
+        # Pass data through conv1
+        x = self.conv1(x)
+        # Use the rectified-linear activation function over x
+        x = F.relu(x)
+
+        x = self.conv2(x)
+        x = F.relu(x)
+
+        # Run max pooling over x
+        x = F.max_pool2d(x, 2)
+        # Pass data through dropout1
+        #x = self.dropout1(x)
+        # Flatten x with start_dim=1
+        x = x.view(-1, self.num_flat_features(x))
+
+        # Pass data through fc1
+        x = self.fc1(x)
+        x = F.relu(x)
+        #x = self.dropout2(x)
+        x = self.fc2(x)
+
+        return x
+
+    def num_flat_features(self, x):
+        size = x.size()[1:]  # all dimensions except the batch dimension
+        num_features = 1
+        for s in size:
+            num_features *= s
+        return num_features
+
+def get_data():
+        f = open("D:\\Studium\\Bachelorarbeit\\Unity Projekte\\A-Stern Test\\A-Stern Test\\Assets\\Resources\\single_example_data.txt", "r")
+        # Using readlines()
+        Lines = f.readlines()
+
+        weightmatrix = []
+        points_return = []
+        target_data_return = []
+
+        count = 0
+        # Strips the newline character
+        for line in Lines:
+            elements = line.split(";")
+
+            #get the weightmatrix
+            x = elements[0].split(", ")
+            x.remove("")
+            for i in range(0, len(x), 1):
+                x[i] = int(x[i])
+            weightmatrix.append(x)
+
+            #get the start and endpoints
+            points = [[0, 0, 0, 0, 0, 0, 0, 0, 0, 0],
+                        [0, 0, 0, 0, 0, 0, 0, 0, 0, 0],
+                        [0, 0, 0, 0, 0, 0, 0, 0, 0, 0],
+                        [0, 0, 0, 0, 0, 0, 0, 0, 0, 0],
+                        [0, 0, 0, 0, 0, 0, 0, 0, 0, 0],
+                        [0, 0, 0, 0, 0, 0, 0, 0, 0, 0],
+                        [0, 0, 0, 0, 0, 0, 0, 0, 0, 0],
+                        [0, 0, 0, 0, 0, 0, 0, 0, 0, 0],
+                        [0, 0, 0, 0, 0, 0, 0, 0, 0, 0],
+                        [0, 0, 0, 0, 0, 0, 0, 0, 0, 0]]
+
+            y = elements[1].split(" ")
+            for i in range(0, len(y), 1):
+                y[i] = y[i].strip('[]')
+                y[i] = y[i].split(",")
+                for j in range(0, len(y[i]), 1):
+                    y[i][j] = int(y[i][j])
+                points[y[i][0]][y[i][1]] = 1
+            points_return.append(points)
+
+            #points.append(torch.tensor(y))
+
+            #get the path and transform it into a matrix
+            z = elements[2].split(", ")
+            z.remove("")
+
+            for i in range(0, len(z), 1):
+                z[i] = z[i].strip('()')
+                z[i] = z[i].split(",")
+                for j in range(0, len(z[i]), 1):
+                    z[i][j] = int(z[i][j])
+
+            target_data = [[0, 0, 0, 0, 0, 0, 0, 0, 0, 0],
+                        [0, 0, 0, 0, 0, 0, 0, 0, 0, 0],
+                        [0, 0, 0, 0, 0, 0, 0, 0, 0, 0],
+                        [0, 0, 0, 0, 0, 0, 0, 0, 0, 0],
+                        [0, 0, 0, 0, 0, 0, 0, 0, 0, 0],
+                        [0, 0, 0, 0, 0, 0, 0, 0, 0, 0],
+                        [0, 0, 0, 0, 0, 0, 0, 0, 0, 0],
+                        [0, 0, 0, 0, 0, 0, 0, 0, 0, 0],
+                        [0, 0, 0, 0, 0, 0, 0, 0, 0, 0],
+                        [0, 0, 0, 0, 0, 0, 0, 0, 0, 0]]
+            for i in range(0, len(z), 1):
+                target_data[z[i][0]][z[i][1]] = 1
+            target_data[y[1][0]][y[1][1]] = 1
+            target_data_return.append(target_data)
+
+            count += 1
+        f.close()
+
+        return weightmatrix, points_return, target_data_return, count
+
+
+if __name__=="__main__":
+    # Remember that you must call model.eval() to set dropout and batch normalization layers to evaluation mode before running inference. Failing to do this will yield inconsistent inference results
+    model = Net()
+    model.load_state_dict(torch.load("D:\\Studium\\Bachelorarbeit\\Machine Learning\\ressources\\models\\model_nr_2.pth"))
+    model.eval()
+
+    weightmatrix, points, target_data, batch_size = get_data()
+    weightmatrix = torch.tensor(weightmatrix, dtype=torch.float32)
+    weightmatrix = weightmatrix.view(batch_size, 1,10,10)
+
+    points = torch.tensor(points, dtype=torch.float32)
+    points = points.view(batch_size, 1,10,10)
+
+    #merge the weightmatrix and the start-/endpoints to a two-channeled input
+    weightmatrix = torch.cat((weightmatrix, points), 1)
+
+    target_data = torch.tensor(target_data, dtype=torch.float32)
+    target_data = target_data.view(batch_size, 100)
+
+    result = model(weightmatrix)
+
+    criterion = nn.MSELoss()
+
+    print("loss: ", criterion(result, target_data))
+    #print ("Result: ", result.view(10,10))
+    #result = result * 7
+    result = torch.round(result)
+    #print(np.reshape(result.detach().numpy(), (10,10)))
+    result_plot = plt.matshow(np.reshape(result.detach().numpy(), (10,10)), cmap=plt.cm.gray)
+
+    result = result.type(torch.int8)
+
+    target_data_plot = plt.matshow(np.reshape(target_data.detach().numpy(), (10,10)), cmap=plt.cm.gray)
+    plt.show()
+
+    target_data = target_data.type(torch.int8)
+    print ("Result: ", result.view(10,10))
+    print ("Target Data: ",  target_data.view(10,10))

Net_trainings/Ansatz2.py

+import torch
+from torch import autograd, nn, optim
+import torch.nn.functional as F
+from tqdm import tqdm
+
+class Net(nn.Module):
+    def __init__(self):
+        super(Net, self).__init__()
+        # self.conv1 = nn.Conv2d(1, 32, 3)
+        # self.conv2 = nn.Conv2d(32, 64, 3)
+        #self.dropout1 = nn.Dropout2d(0.25)
+        #self.dropout2 = nn.Dropout2d(0.5)
+        self.fc1 = nn.Linear(100, 512)
+        self.fc2 = nn.Linear(512, 1028)
+        self.fc3 = nn.Linear(1028, 512)
+        self.fc4 = nn.Linear(512, 100)
+
+    # x represents our data
+    def forward(self, x):
+        # Pass data through conv1
+        # x = self.conv1(x)
+        # # Use the rectified-linear activation function over x
+        # x = F.relu(x)
+        #
+        # x = self.conv2(x)
+        # x = F.relu(x)
+        #
+        # # Run max pooling over x
+        # x = F.max_pool2d(x, 2)
+        # # Pass data through dropout1
+        # #x = self.dropout1(x)
+        # # Flatten x with start_dim=1
+        # x = x.view(-1, self.num_flat_features(x))
+
+        # Pass data through fc1
+        x = self.fc1(x)
+        x = F.relu(x)
+        x = self.fc2(x)
+        x = F.relu(x)
+        x = self.fc3(x)
+        x = F.relu(x)
+        #x = self.dropout2(x)
+        x = self.fc4(x)
+
+        return x
+
+    def num_flat_features(self, x):
+        size = x.size()[1:]  # all dimensions except the batch dimension
+        num_features = 1
+        for s in size:
+            num_features *= s
+        return num_features
+
+
+def get_data():
+        f = open("D:\\Studium\\Bachelorarbeit\\Unity Projekte\\A-Stern Test\\A-Stern Test\\Assets\\Resources\\training_data.txt", "r")
+        # Using readlines()
+        Lines = f.readlines()
+
+        weightmatrix = []
+        points_return = []
+        target_data_return = []
+
+        count = 0
+        # Strips the newline character
+        for line in Lines:
+            elements = line.split(";")
+
+            #get the weightmatrix
+            x = elements[0].split(", ")
+            x.remove("")
+            for i in range(0, len(x), 1):
+                x[i] = int(x[i])
+            weightmatrix.append(x)
+
+            #get the start and endpoints
+            points = [[0, 0, 0, 0, 0, 0, 0, 0, 0, 0],
+                        [0, 0, 0, 0, 0, 0, 0, 0, 0, 0],
+                        [0, 0, 0, 0, 0, 0, 0, 0, 0, 0],
+                        [0, 0, 0, 0, 0, 0, 0, 0, 0, 0],
+                        [0, 0, 0, 0, 0, 0, 0, 0, 0, 0],
+                        [0, 0, 0, 0, 0, 0, 0, 0, 0, 0],
+                        [0, 0, 0, 0, 0, 0, 0, 0, 0, 0],
+                        [0, 0, 0, 0, 0, 0, 0, 0, 0, 0],
+                        [0, 0, 0, 0, 0, 0, 0, 0, 0, 0],
+                        [0, 0, 0, 0, 0, 0, 0, 0, 0, 0]]
+
+            y = elements[1].split(" ")
+            for i in range(0, len(y), 1):
+                y[i] = y[i].strip('[]')
+                y[i] = y[i].split(",")
+                for j in range(0, len(y[i]), 1):
+                    y[i][j] = int(y[i][j])
+                points[y[i][0]][y[i][1]] = 1000
+            points_return.append(points)
+
+            #points.append(torch.tensor(y))
+
+            #get the path and transform it into a matrix
+            z = elements[2].split(", ")
+            z.remove("")
+
+            for i in range(0, len(z), 1):
+                z[i] = z[i].strip('()')
+                z[i] = z[i].split(",")
+                for j in range(0, len(z[i]), 1):
+                    z[i][j] = int(z[i][j])
+
+            target_data = [[0, 0, 0, 0, 0, 0, 0, 0, 0, 0],
+                        [0, 0, 0, 0, 0, 0, 0, 0, 0, 0],
+                        [0, 0, 0, 0, 0, 0, 0, 0, 0, 0],
+                        [0, 0, 0, 0, 0, 0, 0, 0, 0, 0],
+                        [0, 0, 0, 0, 0, 0, 0, 0, 0, 0],
+                        [0, 0, 0, 0, 0, 0, 0, 0, 0, 0],
+                        [0, 0, 0, 0, 0, 0, 0, 0, 0, 0],
+                        [0, 0, 0, 0, 0, 0, 0, 0, 0, 0],
+                        [0, 0, 0, 0, 0, 0, 0, 0, 0, 0],
+                        [0, 0, 0, 0, 0, 0, 0, 0, 0, 0]]
+            for i in range(0, len(z), 1):
+                target_data[z[i][0]][z[i][1]] = 100
+            target_data[y[1][0]][y[1][1]] = 100
+            target_data_return.append(target_data)
+
+            count += 1
+        f.close()
+
+        return weightmatrix, points_return, target_data_return, count
+
+if __name__ == '__main__':
+    weightmatrix, points, target_data, batch_size = get_data()
+    weightmatrix = torch.tensor(weightmatrix, dtype=torch.float32)
+    weightmatrix = weightmatrix.view(batch_size, 1,10,10)
+    #weightmatrix = weightmatrix.type(torch.int8)
+
+    points = torch.tensor(points, dtype=torch.float32)
+    points = points.view(batch_size, 1,10,10)
+
+    #merge the weightmatrix and the start-/endpoints to a two-channeled input
+    #weightmatrix = torch.cat((weightmatrix, points), 1)
+
+    target_data = torch.tensor(target_data, dtype=torch.float32)
+    target_data = target_data.view(batch_size, 100)
+    target_data = target_data + weightmatrix.view(batch_size, 100)
+    target_data = target_data + points.view(batch_size, 100)
+    #target_data.type(torch.int16)
+
+    learning_rate = 0.1
+
+    model_path = 'D:\\Studium\\Bachelorarbeit\\Machine Learning\\ressources\\models'
+
+    nn_pathfinder = Net()
+    opt = optim.SGD(params=nn_pathfinder.parameters(), lr=learning_rate)
+
+    result = nn_pathfinder(weightmatrix.view(batch_size, 100))
+    criterion = nn.MSELoss()
+    #smoothl1 - loss testen
+    #print ("Result: ", result)
+
+    print("before training")
+    # print ("Result: ", result.flatten())
+    # print ("Target Data: ",  target_data.flatten())
+    print("loss: ", criterion(result, target_data))
+
+    for epoch in tqdm(range(7500)):
+        result = nn_pathfinder(weightmatrix.view(batch_size, 100))
+        #loss = F.nll_loss(result, target_data)
+
+        loss = criterion(result, target_data)
+        #if(epoch % 10 == 0):
+        print("loss" , loss)
+
+        nn_pathfinder.zero_grad()
+        loss.backward()
+        opt.step()
+
+
+    print("after")
+    # print ("Result: ", result.flatten())
+    # #result = result.type(torch.int8)
+    # result = torch.round(result)
+    # result = result.type(torch.int8)
+    # target_data = target_data.type(torch.int8)
+    # print ("Result: ", result.view(10,10))
+    # print ("Target Data: ",  target_data.view(10,10))
+    print("loss: ", loss)
+
+    torch.save(nn_pathfinder.state_dict(),
+                model_path + "\\ansatz2_model_nr_" + "1" + ".pth")

Net_trainings/ConvNet.py

+import torch
+from torch import autograd, nn, optim
+import torch.nn.functional as F
+from tqdm import tqdm
+
+class Net(nn.Module):
+    def __init__(self):
+        super(Net, self).__init__()
+        self.conv1 = nn.Conv2d(2, 32, 3)
+        self.conv2 = nn.Conv2d(32, 64, 3)
+        self.fc1 = nn.Linear(576, 512)
+        self.fc2 = nn.Linear(512, 100)
+
+    # x represents our data
+    def forward(self, x):
+        # Pass data through conv1
+        x = self.conv1(x)
+        # Use the rectified-linear activation function over x
+        x = F.relu(x)
+
+        x = self.conv2(x)
+        x = F.relu(x)
+
+        # Run max pooling over x
+        x = F.max_pool2d(x, 2)
+        # Flatten x with start_dim=1
+        x = x.view(-1, self.num_flat_features(x))
+
+        # Pass data through fc1
+        x = self.fc1(x)
+        x = F.relu(x)
+        x = self.fc2(x)
+
+        return x
+
+    def num_flat_features(self, x):
+        size = x.size()[1:]  # all dimensions except the batch dimension
+        num_features = 1
+        for s in size:
+            num_features *= s
+        return num_features
+
+
+def get_data():
+        f = open("D:\\Studium\\Bachelorarbeit\\Unity Projekte\\A-Stern Test\\A-Stern Test\\Assets\\Resources\\training_data.txt", "r")
+        # Using readlines()
+        Lines = f.readlines()
+
+        weightmatrix = []
+        points_return = []
+        target_data_return = []
+
+        count = 0
+        # Strips the newline character
+        for line in Lines:
+            elements = line.split(";")
+
+            #get the weightmatrix
+            x = elements[0].split(", ")
+            x.remove("")
+            for i in range(0, len(x), 1):
+                x[i] = int(x[i])
+            weightmatrix.append(x)
+
+            #get the start and endpoints
+            points = [[0, 0, 0, 0, 0, 0, 0, 0, 0, 0],
+                        [0, 0, 0, 0, 0, 0, 0, 0, 0, 0],
+                        [0, 0, 0, 0, 0, 0, 0, 0, 0, 0],
+                        [0, 0, 0, 0, 0, 0, 0, 0, 0, 0],
+                        [0, 0, 0, 0, 0, 0, 0, 0, 0, 0],
+                        [0, 0, 0, 0, 0, 0, 0, 0, 0, 0],
+                        [0, 0, 0, 0, 0, 0, 0, 0, 0, 0],
+                        [0, 0, 0, 0, 0, 0, 0, 0, 0, 0],
+                        [0, 0, 0, 0, 0, 0, 0, 0, 0, 0],
+                        [0, 0, 0, 0, 0, 0, 0, 0, 0, 0]]
+
+            y = elements[1].split(" ")
+            for i in range(0, len(y), 1):
+                y[i] = y[i].strip('[]')
+                y[i] = y[i].split(",")
+                for j in range(0, len(y[i]), 1):
+                    y[i][j] = int(y[i][j])
+                points[y[i][0]][y[i][1]] = 1
+            points_return.append(points)
+
+            #get the path and transform it into a matrix
+            z = elements[2].split(", ")
+            z.remove("")
+
+            for i in range(0, len(z), 1):
+                z[i] = z[i].strip('()')
+                z[i] = z[i].split(",")
+                for j in range(0, len(z[i]), 1):
+                    z[i][j] = int(z[i][j])
+
+            target_data = [[0, 0, 0, 0, 0, 0, 0, 0, 0, 0],
+                        [0, 0, 0, 0, 0, 0, 0, 0, 0, 0],
+                        [0, 0, 0, 0, 0, 0, 0, 0, 0, 0],
+                        [0, 0, 0, 0, 0, 0, 0, 0, 0, 0],
+                        [0, 0, 0, 0, 0, 0, 0, 0, 0, 0],
+                        [0, 0, 0, 0, 0, 0, 0, 0, 0, 0],
+                        [0, 0, 0, 0, 0, 0, 0, 0, 0, 0],
+                        [0, 0, 0, 0, 0, 0, 0, 0, 0, 0],
+                        [0, 0, 0, 0, 0, 0, 0, 0, 0, 0],
+                        [0, 0, 0, 0, 0, 0, 0, 0, 0, 0]]
+            for i in range(0, len(z), 1):
+                target_data[z[i][0]][z[i][1]] = 1
+            target_data[y[1][0]][y[1][1]] = 1
+            target_data_return.append(target_data)
+
+            count += 1
+        f.close()
+
+        return weightmatrix, points_return, target_data_return, count
+
+if __name__ == '__main__':
+    weightmatrix, points, target_data, batch_size = get_data()
+    weightmatrix = torch.tensor(weightmatrix, dtype=torch.float32)
+    weightmatrix = weightmatrix.view(batch_size, 1,10,10)
+
+    points = torch.tensor(points, dtype=torch.float32)
+    points = points.view(batch_size, 1,10,10)
+
+    #merge the weightmatrix and the start-/endpoints to a two-channeled input
+    weightmatrix = torch.cat((weightmatrix, points), 1)
+
+    target_data = torch.tensor(target_data, dtype=torch.float32)
+    target_data = target_data.view(batch_size, 100)
+
+    learning_rate = 0.1
+
+    model_path = 'D:\\Studium\\Bachelorarbeit\\Machine Learning\\ressources\\models'
+
+    nn_pathfinder = Net()
+    opt = optim.SGD(params=nn_pathfinder.parameters(), lr=learning_rate)
+
+    result = nn_pathfinder(weightmatrix)
+    criterion = nn.MSELoss()
+
+    print("before training")
+    print ("Result: ", result.flatten())
+    print ("Target Data: ",  target_data.flatten())
+    print("loss: ", criterion(result, target_data))
+
+    for epoch in tqdm(range(7500)):
+        result = nn_pathfinder(weightmatrix)
+
+        loss = criterion(result, target_data)
+
+        nn_pathfinder.zero_grad()
+        loss.backward()
+        opt.step()
+
+
+    print("after")
+    # print ("Result: ", result.flatten())
+    # result = torch.round(result)
+    # result = result.type(torch.int8)
+    # target_data = target_data.type(torch.int8)
+    # print ("Result: ", result.view(10,10))
+    # print ("Target Data: ",  target_data.view(10,10))
+    print("loss: ", loss)
+
+    torch.save(nn_pathfinder.state_dict(),
+                model_path + "\\model_nr_" + "2" + ".pth")

Net_trainings/FullyConnected.py

+import torch
+from torch import autograd, nn, optim
+import torch.nn.functional as F
+from tqdm import tqdm
+
+class Net(nn.Module):
+    def __init__(self):
+        super(Net, self).__init__()
+        self.conv1 = nn.Conv2d(1, 32, 3)
+        self.conv2 = nn.Conv2d(32, 64, 3)
+        self.fc1 = nn.Linear(576, 128)
+        self.fc2 = nn.Linear(128, 100)
+
+    # x represents our data
+    def forward(self, x):
+        # Pass data through conv1
+        x = self.conv1(x)
+        # Use the rectified-linear activation function over x
+        x = F.relu(x)
+
+        x = self.conv2(x)
+        x = F.relu(x)
+
+        # Run max pooling over x
+        x = F.max_pool2d(x, 2)
+        # Flatten x with start_dim=1
+        x = x.view(-1, self.num_flat_features(x))
+        # Pass data through fc1
+        x = self.fc1(x)
+        x = F.relu(x)
+
+        x = self.fc2(x)
+        x = F.relu(x)
+
+        return x
+
+    def num_flat_features(self, x):
+        size = x.size()[1:]  # all dimensions except the batch dimension
+        num_features = 1
+        for s in size:
+            num_features *= s
+        return num_features
+
+
+def get_data():
+        f = open("D:\\Studium\\Bachelorarbeit\\Unity Projekte\\A-Stern Test\\A-Stern Test\\Assets\\Resources\\data.txt", "r")
+        # Using readlines()
+        Lines = f.readlines()
+
+        weightmatrix = []
+        points = []
+        target_data_return = []
+
+        count = 0
+        # Strips the newline character
+        for line in Lines:
+            elements = line.split(";")
+
+            #get the weightmatrix
+            x = elements[0].split(", ")
+            x.remove("")
+            for i in range(0, len(x), 1):
+                x[i] = int(x[i])
+            weightmatrix.append(x)
+
+            #get the start and endpoints
+            y = elements[1].split(" ")
+            for i in range(0, len(y), 1):
+                y[i] = y[i].strip('[]')
+                y[i] = y[i].split(",")
+                for j in range(0, len(y[i]), 1):
+                    y[i][j] = int(y[i][j])
+
+            points.append(torch.tensor(y))
+
+            #get the path and transform it into a matrix
+            z = elements[2].split(", ")
+            z.remove("")
+
+            for i in range(0, len(z), 1):
+                z[i] = z[i].strip('()')
+                z[i] = z[i].split(",")
+                for j in range(0, len(z[i]), 1):
+                    z[i][j] = int(z[i][j])
+
+            target_data = [[0, 0, 0, 0, 0, 0, 0, 0, 0, 0],
+                        [0, 0, 0, 0, 0, 0, 0, 0, 0, 0],
+                        [0, 0, 0, 0, 0, 0, 0, 0, 0, 0],
+                        [0, 0, 0, 0, 0, 0, 0, 0, 0, 0],
+                        [0, 0, 0, 0, 0, 0, 0, 0, 0, 0],
+                        [0, 0, 0, 0, 0, 0, 0, 0, 0, 0],
+                        [0, 0, 0, 0, 0, 0, 0, 0, 0, 0],
+                        [0, 0, 0, 0, 0, 0, 0, 0, 0, 0],
+                        [0, 0, 0, 0, 0, 0, 0, 0, 0, 0],
+                        [0, 0, 0, 0, 0, 0, 0, 0, 0, 0]]              #torch.zeros([10, 10])
+            for i in range(0, len(z), 1):
+                target_data[z[i][0]][z[i][1]] = 1
+            target_data[y[1][0]][y[1][1]] = 1
+            target_data_return.append(target_data)
+
+            count += 1
+        f.close()
+
+        return weightmatrix, points, target_data_return, count
+
+
+weightmatrix, points, target_data, batch_size = get_data()
+weightmatrix = torch.tensor(weightmatrix)
+weightmatrix = weightmatrix.view(batch_size, 1,10,10)
+
+learning_rate = 0.1
+
+
+
+
+target_data = torch.tensor(target_data)
+target_data = target_data.view(batch_size, 1,10,10)
+
+
+my_nn = Net()
+opt = optim.SGD(params=my_nn.parameters(), lr=learning_rate)
+
+result = my_nn(weightmatrix)
+criterion = nn.MSELoss()
+
+print("before training")
+print ("Result: ", result.flatten())
+print ("Target Data: ",  target_data.flatten())
+print("loss: ", criterion(result, target_data))
+
+for epoch in tqdm(range(7500)):
+    result = my_nn(weightmatrix)
+
+    loss = criterion(result, target_data)
+
+    my_nn.zero_grad()
+    loss.backward()
+    opt.step()
+
+
+print("after")
+print ("Result: ", result.flatten())
+print ("Target Data: ",  target_data.flatten())
+print("loss: ", loss)

Net_trainings/rnn_net.py

+
+
+import torch
+from torch import nn
+
+import numpy as np
+
+class Model(nn.Module):
+    def __init__(self, input_size, output_size, hidden_dim, n_layers):
+        super(Model, self).__init__()
+
+        # Defining some parameters
+        self.hidden_dim = hidden_dim
+        self.n_layers = n_layers
+
+        #Defining the layers
+        # RNN Layer
+        self.rnn = nn.RNN(input_size, hidden_dim, n_layers, batch_first=True)
+        # Fully connected layer
+        self.fc = nn.Linear(hidden_dim, output_size)
+
+    def forward(self, x):
+
+        batch_size = x.size(0)
+
+        #Initializing hidden state for first input using method defined below
+        hidden = self.init_hidden(batch_size)
+
+        # Passing in the input and hidden state into the model and obtaining outputs
+        out, hidden = self.rnn(x, hidden)
+
+        # Reshaping the outputs such that it can be fit into the fully connected layer
+        out = out.contiguous().view(-1, self.hidden_dim)
+        out = self.fc(out)
+
+        return out, hidden
+
+    def init_hidden(self, batch_size):
+        # This method generates the first hidden state of zeros which we'll use in the forward pass
+        hidden = torch.zeros(self.n_layers, batch_size, self.hidden_dim).to(device)
+         # We'll send the tensor holding the hidden state to the device we specified earlier as well
+        return hidden

Statistics/min_distance_plotting.py

+import numpy as np
+import time
+import os
+from matplotlib import pyplot as plt
+import torch
+from torch import autograd, nn, optim
+import torch.nn.functional as F
+import cv2
+#from scipy.optimize import curve_fit
+
+#erste auslesung der Daten
+f = open("D:\\Studium\\Bachelorarbeit\\Machine Learning\\loss_function\\complete_statistics_changing_gapweight_clusterweight0,5.txt", "r")
+# Using readlines()
+Lines = f.readlines()
+f.close()
+array = []
+
+for line in Lines:
+    if "min_distance" in line:
+        #tmp_array = line.split("number_of_pixels: tensor(")
+        tmp_array = line.split("min_distance: ")
+        tmp_array2 = tmp_array[1].split(",")
+        tmp_array2 = tmp_array2[0].split(".")
+        if "tensor" in tmp_array2[0]:
+            tmp_array2 = tmp_array2[0].split("tensor(")
+            tmp_array2 = tmp_array2[1].split(")")
+        array.append(float(tmp_array2[0]))
+        #print(tmp_array2[0])
+array1 = np.array(array)
+
+#zweite auslesung der Daten
+f = open("D:\\Studium\\Bachelorarbeit\\Machine Learning\\loss_function\\complete_statistics_changing_gapweight_clusterweight1,1.txt", "r")
+Lines = f.readlines()
+f.close()
+array = []
+
+for line in Lines:
+    if "min_distance" in line:
+        #tmp_array = line.split("number_of_pixels: tensor(")
+        tmp_array = line.split("min_distance: ")
+        tmp_array2 = tmp_array[1].split(",")
+        tmp_array2 = tmp_array2[0].split(".")
+        if "tensor" in tmp_array2[0]:
+            tmp_array2 = tmp_array2[0].split("tensor(")
+            tmp_array2 = tmp_array2[1].split(")")
+        array.append(float(tmp_array2[0]))
+        #print(tmp_array2[0])
+array2 = np.array(array)
+
+#dritte auslesung der Daten
+f = open("D:\\Studium\\Bachelorarbeit\\Machine Learning\\loss_function\\complete_statistics_changing_gapweight_clusterweight2.txt", "r")
+Lines = f.readlines()
+f.close()
+array = []
+
+for line in Lines:
+    if "min_distance" in line:
+        #tmp_array = line.split("number_of_pixels: tensor(")
+        tmp_array = line.split("min_distance: ")
+        tmp_array2 = tmp_array[1].split(",")
+        tmp_array2 = tmp_array2[0].split(".")
+        if "tensor" in tmp_array2[0]:
+            tmp_array2 = tmp_array2[0].split("tensor(")
+            tmp_array2 = tmp_array2[1].split(")")
+        array.append(float(tmp_array2[0]))
+        #print(tmp_array2[0])
+array3 = np.array(array)
+
+
+
+
+
+
+
+
+#x = np.arange(0, 10, 0.1)
+#arr = np.array([1, 2, 3, 4, 5])
+
+
+x = np.arange(0, 10, 0.1)
+x2 = np.arange(0, 5000, 50)
+
+#changing clusterzize weight
+# plt.plot(x, array1, label="Minimal cluster distance with Gapweight 1000")
+# plt.plot(x, array2, label="Minimal cluster distance with Gapweight 3000")
+# plt.plot(x, array3, label="Minimal cluster distance with Gapweight 5000")
+
+#changing gap-weight
+#plt.plot(x2, array1, label="Minimal cluster distance with Clusterweight 0.5")
+#plt.plot(x2, array2, label="Minimal cluster distance with Clusterweight 1.1")
+plt.plot(x2, array3, label="Minimal cluster distance with Clusterweight 2")
+
+
+plt.xlabel("Weight of the gap-loss")
+plt.ylabel("Minimal distance(in Pixels)")
+
+plt.legend(loc="upper left")
+plt.grid()
+plt.ylim(-1, 10)
+plt.axhline(color='black', lw=0.75)
+plt.axvline(color='black', lw=0.75)
+plt.savefig("min_distance_changing_gapweight_clusterweight2")
+plt.show()

Statistics/startcluster_size_plotting.py

+import numpy as np
+import time
+import os
+from matplotlib import pyplot as plt
+import torch
+from torch import autograd, nn, optim
+import torch.nn.functional as F
+import cv2
+#from scipy.optimize import curve_fit
+
+def determine_startcluster_size(path):
+    filepath = "D:\\Studium\\Bachelorarbeit\\Machine Learning\\loss_function\\" + path + ".txt"
+    f = open(filepath, "r")
+    # Using readlines()
+    Lines = f.readlines()
+    in_result = 0
+    tmp_array2 = ''
+    tmp_array3 = []
+    # Strips the newline character
+    for line in Lines:
+        if(in_result == 1):
+            if "]" in line:
+                tmp_array = line.split("]")
+                tmp_array2 += tmp_array[0]
+                in_result = 0
+                array = tmp_array2.split(".,")
+                result = []
+                for value in array:
+                    if "\n" in value:
+                        result.append(int(value.split("\n        ")[1]))
+                    elif " " in value:
+                        if "." in value:
+                            result.append(int(value.split(" ")[1].split(".")[0]))
+                        else:
+                            result.append(int(value.split(" ")[1]))
+                    else:
+                        # if "." in value:
+                        #     result.append(int(value.split(".")[0]))
+                        # else:
+                        result.append(int(value))
+                #print(torch.tensor(result).view(10,10))
+                tmp_array2 = ''
+
+
+                #find the start_cluster_size
+                result = torch.tensor(result, dtype=torch.float32).view(10,10)
+                img = result.round()
+                img = img.detach().numpy()
+                img = np.array(img, dtype=np.uint8)
+                num_labels, labels_im = cv2.connectedComponents(img)
+
+                start_value = labels_im[4][6]
+                end_value = labels_im[0][2]
+
+                #remodel the matrix, so that distance_transform can be used
+                start_cluster = []
+
+                if(start_value != 0 and end_value != 0):
+                    column_nr = len(labels_im)
+                    row_nr = len(labels_im[0])
+                    for i in range(0, column_nr):
+                        for j in range(0, row_nr):
+                            if (labels_im[i][j] == end_value):
+                                labels_im[i][j] = 0
+                            elif (labels_im[i][j] == 0):
+                                labels_im[i][j] = end_value
+                            elif (labels_im[i][j] == start_value):
+                                start_cluster.append([i,j])
+
+                            if(end_value == start_value):
+                                if(labels_im[i][j] == 0):
+                                    start_cluster.append([i,j])
+                #print(len(start_cluster))
+                tmp_array3.append(len(start_cluster))
+
+            else:
+                tmp_array2 += line
+        # if "min_distance" in line:
+        #     #tmp_array = line.split("number_of_pixels: tensor(")
+        #     tmp_array = line.split("min_distance: ")
+        #     tmp_array2 = tmp_array[1].split(",")
+        #     tmp_array2 = tmp_array2[0].split(".")
+        #     print(tmp_array2[0])
+        if "result_rounded:" in line:
+            #tmp_array = line.split("number_of_pixels: tensor(")
+            tmp_array = line.split("result_rounded: tensor([")
+            tmp_array2 = tmp_array[1]
+            in_result = 1
+    return tmp_array3
+
+#erste auslesung der Daten
+#array1 = determine_startcluster_size("complete_statistics_changing_gapweight_clusterweight0,5")
+array1 = determine_startcluster_size("complete_statistics_changing_clusterweight_gapweight1000")
+
+#zweite auslesung der Daten
+#array2 = determine_startcluster_size("complete_statistics_changing_gapweight_clusterweight1,1")
+array2 = determine_startcluster_size("complete_statistics_changing_clusterweight_gapweight3000")
+
+#dritte auslesung der Daten
+#array3 = determine_startcluster_size("complete_statistics_changing_gapweight_clusterweight2")
+array3 = determine_startcluster_size("complete_statistics_changing_clusterweight_gapweight5000")
+
+
+x = np.arange(0, 10, 0.1)
+arr = np.array([1, 2, 3, 4, 5])
+
+
+x = np.arange(0, 10, 0.1)
+x2 = np.arange(0, 5000, 50)
+
+#changing gapweight stats
+#plt.plot(x2, array1, label="Size of the cluster around the startpoint with Clusterweight 0.5")
+#plt.plot(x2, array2, label="Size of the cluster around the startpoint with Clusterweight 1.1")
+#plt.plot(x2, array3, label="Size of the cluster around the startpoint with Clusterweight 2")
+
+#changing clusterweight stats
+plt.plot(x, array1, label="Size of the cluster around the startpoint with Gapweight 1000")
+plt.plot(x, array2, label="Size of the cluster around the startpoint with Gapweight 3000")
+plt.plot(x, array3, label="Size of the cluster around the startpoint with Gapweight 5000")
+
+
+plt.xlabel("Weight of the clustersize-loss")
+           # family='serif',
+           # color='r',
+           # weight='normal',
+           # size = 16,
+           # labelpad = 6)
+plt.ylabel("Clustersize(Number of Pixels)")
+plt.legend(loc="upper left")
+plt.grid()
+plt.ylim(-1, 50)
+plt.axhline(color='black', lw=0.75)
+plt.axvline(color='black', lw=0.75)
+plt.savefig("startcluster_size_changing_clusterweight")
+plt.show()

basics/ReadFile.py

+import torch
+
+if __name__=="__main__":
+    f = open("D:\\Studium\\Bachelorarbeit\\Unity Projekte\\A-Stern Test\\A-Stern Test\\Assets\\Resources\\data.txt", "r")
+    # Using readlines()
+    Lines = f.readlines()
+
+    # Strips the newline character
+    for line in Lines:
+        #print("{}".format(line.strip()))
+        #print(line)
+        elements = line.split(";")
+
+        #get the weightmatrix
+        x = elements[0].split(", ")
+        x.remove("")
+        for i in range(0, len(x), 1):
+            x[i] = int(x[i])
+        tensor = torch.tensor(x)
+        weightmatrix = tensor.view(-1, 10)
+
+        #get the start and endpoints
+        y = elements[1].split(" ")
+        for i in range(0, len(y), 1):
+            y[i] = y[i].strip('[]')
+            y[i] = y[i].split(",")
+            for j in range(0, len(y[i]), 1):
+                y[i][j] = int(y[i][j])
+
+        points = torch.tensor(y)
+
+        #get the path and transform it into a matrix
+        z = elements[2].split(", ")
+        z.remove("")
+        print(z)
+
+        for i in range(0, len(z), 1):
+            z[i] = z[i].strip('()')
+            z[i] = z[i].split(",")
+            for j in range(0, len(z[i]), 1):
+                z[i][j] = int(z[i][j])
+
+        target_data = torch.zeros([10, 10])
+        for i in range(0, len(z), 1):
+            target_data[z[i][0]][z[i][1]] = 1
+        target_data[y[1][0]][y[1][1]] = 1
+
+        print(target_data)

basics/load_model.py

+# Remember that you must call model.eval() to set dropout and batch normalization layers to evaluation mode before running inference. Failing to do this will yield inconsistent inference results
+model = TheModelClass(*args, **kwargs)
+model.load_state_dict(torch.load(PATH))
+model.eval()

basics/plotting.py

+import matplotlib.pyplot as plt
+import matplotlib.image as mpimg
+
+
+img = [[1, 0, 0, 0, 0, 0, 0, 0, 0, 0],
+        [0, 1, 0, 0, 0, 0, 0, 0, 0, 0],
+        [0, 0, 1, 0, 0, 0, 0, 0, 0, 0],
+        [0, 0, 0, 1, 0, 0, 0, 0, 0, 0],
+        [0, 0, 0, 0, 1, 0, 0, 0, 0, 0],
+        [0, 0, 0, 0, 0, 1, 0, 0, 0, 0],
+        [0, 0, 0, 0, 0, 0, 1, 0, 0, 0],
+        [0, 0, 0, 0, 0, 0, 0, 1, 0, 0],
+        [0, 0, 0, 0, 0, 0, 0, 0, 1, 0],
+        [0, 0, 0, 0, 0, 0, 0, 0, 0, 1]]
+
+imgplot = plt.matshow(img, cmap=plt.cm.gray)
+plt.show()

basics/test_crop_for_unet.py

+import torch
+from torch import autograd, nn, optim
+import torch.nn.functional as F
+from tqdm import tqdm
+
+
+def crop_tensor(x, cropping_size):
+    if((x.size()[2]/2) >= cropping_size):
+        return x[0:, 0:, cropping_size:-cropping_size, cropping_size:-cropping_size]
+
+if __name__ == '__main__':
+    list = [[[[0, 1, 2, 3, 4, 5, 6, 7, 8, 9],
+    [0, 1, 2, 3, 4, 5, 6, 7, 8, 9],
+    [0, 1, 2, 3, 4, 5, 6, 7, 8, 9],
+    [0, 1, 2, 3, 4, 5, 6, 7, 8, 9],
+    [0, 1, 2, 3, 4, 5, 6, 7, 8, 9],
+    [0, 1, 2, 3, 4, 5, 6, 7, 8, 9],
+    [0, 1, 2, 3, 4, 5, 6, 7, 8, 9],
+    [0, 1, 2, 3, 4, 5, 6, 7, 8, 9],
+    [0, 1, 2, 3, 4, 5, 6, 7, 8, 9],
+    [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]]]]
+
+    tensor = torch.tensor(list)
+
+    tensor = crop_tensor(tensor, 1)
+    print(tensor)

basics/u_net.py

+import torch
+from torch import autograd, nn, optim
+import torch.nn.functional as F
+from tqdm import tqdm
+
+
+class U_Net(nn.Module):
+    def __init__(self):
+        super(U_Net, self).__init__()
+
+        #the way down the u-form:
+
+        #first convolution-block
+        self.down1 = nn.Sequential(
+            nn.Conv2d(2, 64, 3),
+            nn.Conv2d(64, 64, 3),
+            #F.relu()
+            #F.max_pool2d(kernel_size=2)
+        )
+
+        # self.conv1 = nn.Conv2d(2, 64, 3)
+        # self.conv1_1 = nn.Conv2d(64, 64, 3)
+            #max pool2d in forward, also safe the state in the forward function
+
+        #second convolution-block
+        self.down2 = nn.Sequential(
+            nn.Conv2d(64, 128, 3),
+            nn.Conv2d(128, 128, 3),
+            #F.relu()
+            #F.max_pool2d(kernel_size=2)
+        )
+        # self.conv2 = nn.Conv2d(64, 128, 3)
+        # self.conv2_1 = nn.Conv2d(128, 128, 3)
+            #max pool2d in forward, also safe the state in the forward function
+
+        #third convolution-block
+        self.down3 = nn.Sequential(
+            nn.Conv2d(128, 256, 3),
+            nn.Conv2d(256, 256, 3),
+            #F.relu()
+            #F.max_pool2d(kernel_size=2)
+        )
+        # self.conv3 = nn.Conv2d(128, 256, 3)
+        # self.conv3_1 = nn.Conv2d(256, 256, 3)
+            #max pool2d in forward, also safe the state in the forward function
+
+        #forth convolution-block
+        self.down4 = nn.Sequential(
+            nn.Conv2d(256, 512, 3),
+            nn.Conv2d(512, 512, 3),
+            #F.relu()
+            #F.max_pool2d(kernel_size=2)
+        )
+        # self.conv4 = nn.Conv2d(256, 512, 3)
+        # self.conv4_1 = nn.Conv2d(512, 512, 3)
+            #max pool2d in forward, also safe the state in the forward function
+
+        #fifth convolution-block and the bottom of the u-form
+        self.bottom = nn.Sequential(
+            nn.Conv2d(512, 1024, 3),
+            nn.Conv2d(1024, 1024, 3),
+            #F.relu()
+        )
+        # self.conv5 = nn.Conv2d(512, 1024, 3)
+        # self.conv5_1 = nn.Conv2d(1024, 1024, 3)
+
+
+        #the way up the u-form, with the help of upconvolution. Use "ConvTranspose2d" and not(!) "Upsample". Upsample doesn't learn, ConvTranspose2d learns parameters
+
+        self.upconv1 = nn.ConvTranspose2d(1024, 512, 2, stride=2)
+        self.up1 = nn.Sequential(
+            nn.Conv2d(1024, 512, 3),
+            nn.Conv2d(512, 512, 3),
+            #F.relu()
+        )
+
+        self.upconv2 = nn.ConvTranspose2d(512, 256, 2, stride=2)
+        self.up2 = nn.Sequential(
+            # nn.ConvTranspose2d(512, 256, 2, stride=2),
+            nn.Conv2d(512, 256, 3),
+            nn.Conv2d(256, 256, 3),
+            #F.relu()
+        )
+
+        self.upconv3 = nn.ConvTranspose2d(256, 128, 2, stride=2)
+        self.up3 = nn.Sequential(
+            # nn.ConvTranspose2d(256, 128, 2, stride=2),
+            nn.Conv2d(256, 128, 3),
+            nn.Conv2d(128, 128, 3),
+            #F.relu()
+        )
+
+        self.upconv4 = nn.ConvTranspose2d(128, 64, 2, stride=2)
+        self.up4 = nn.Sequential(
+            # nn.ConvTranspose2d(128, 64, 2, stride=2),
+            nn.Conv2d(128, 64, 3),
+            nn.Conv2d(64, 64, 3),
+            #F.relu()
+        )
+
+        self.final_conv = nn.Conv2d(64, 1, kernel_size=1)
+
+    # x represents our data
+    def forward(self, x):
+        #down the u
+        d_1 = self.down1(x)
+        d_1_tmp = d_1
+        d_1 = F.max_pool2d(d_1, kernel_size=2)
+        d_1 = F.relu(d_1)
+        d_2 = self.down2(d_1)
+        d_2_tmp = d_2
+        d_2 = F.max_pool2d(d_2, kernel_size=2)
+        d_2 = F.relu(d_2)
+        d_3 = self.down3(d_2)
+        d_3_tmp = d_3
+        d_3 = F.max_pool2d(d_3, kernel_size=2)
+        d_3 = F.relu(d_3)
+        d_4 = self.down4(d_3)
+        d_4_tmp = d_4
+        d_4 = F.max_pool2d(d_4, kernel_size=2)
+        d_4 = F.relu(d_4)
+
+        #bottom
+        return_value = self.bottom(d_4)
+        return_value = F.relu(return_value)
+
+        #up the u
+        return_value = self.upconv1(return_value)
+        return_value = self.up1(self.concatenate_tensors(return_value, self.crop_tensor(d_4_tmp, 4)))
+        return_value = F.relu(return_value)
+        return_value = self.upconv2(return_value)
+        return_value = self.up2(self.concatenate_tensors(return_value, self.crop_tensor(d_3_tmp, 16)))
+        return_value = F.relu(return_value)
+        return_value = self.upconv3(return_value)
+        return_value = self.up3(self.concatenate_tensors(return_value, self.crop_tensor(d_2_tmp, 40)))
+        return_value = F.relu(return_value)
+        return_value = self.upconv4(return_value)
+        return_value = self.up4(self.concatenate_tensors(return_value, self.crop_tensor(d_1_tmp, 88)))
+
+        return_value = self.final_conv(return_value)
+
+        return return_value
+
+    def num_flat_features(self, x):
+        size = x.size()[1:]  # all dimensions except the batch dimension
+        num_features = 1
+        for s in size:
+            num_features *= s
+        return num_features
+
+    def crop_tensor(self, x, cropping_size):
+        if((x.size()[2]/2) >= cropping_size):
+            return x[0:, 0:, cropping_size:-cropping_size, cropping_size:-cropping_size]
+
+    def concatenate_tensors(self, x, y):
+        return torch.cat((x, y), 1)                 #dimension 1 is the channel-dimension
+
+def init():
+    # torch.cuda.is_available() checks and returns a Boolean True if a GPU is available, else it'll return False
+    is_cuda = torch.cuda.is_available()
+
+    # If we have a GPU available, we'll set our device to GPU. We'll use this device variable later in our code.
+    if is_cuda:
+        device = torch.device("cuda")
+        print("GPU is available")
+    else:
+        device = torch.device("cpu")
+        print("GPU not available, CPU used")
+
+    return device
+
+
+
+if __name__ == '__main__':
+    device = init()
+
+
+    weightmatrix = torch.rand(1,2,572,572)
+
+    target_data = torch.rand(1,1,388,388)
+
+    learning_rate = 0.01
+
+    model_path = 'D:\\Studium\\Bachelorarbeit\\Machine Learning\\ressources\\models'
+
+    nn_pathfinder = U_Net()
+    opt = optim.SGD(params=nn_pathfinder.parameters(), lr=learning_rate)
+    nn_pathfinder.to(device)
+
+    result = nn_pathfinder(weightmatrix)
+    criterion = nn.MSELoss()
+    #print ("Result: ", result)
+
+    print("before training")
+    # print ("Result: ", result.flatten())
+    # print ("Target Data: ",  target_data.flatten())
+    print("loss: ", criterion(result, target_data))
+
+    for epoch in tqdm(range(10)):
+        result = nn_pathfinder(weightmatrix)
+
+        loss = criterion(result, target_data)
+        #if(epoch % 10 == 0):
+        #    print("loss" , loss)
+
+        nn_pathfinder.zero_grad()
+        loss.backward()
+        opt.step()
+
+
+    print("after")
+    # print ("Result: ", result.flatten())
+    # #result = result.type(torch.int8)
+    # result = torch.round(result)
+    # result = result.type(torch.int8)
+    # target_data = target_data.type(torch.int8)
+    # print ("Result: ", result.view(10,10))
+    # print ("Target Data: ",  target_data.view(10,10))
+    print("loss: ", loss)
+
+    # save the model
+    torch.save(nn_pathfinder.state_dict(),
+                model_path + "\\128x128_u-net_nr_" + "1" + ".pth")

basics/use_gpu.py

+# torch.cuda.is_available() checks and returns a Boolean True if a GPU is available, else it'll return False
+is_cuda = torch.cuda.is_available()
+
+# If we have a GPU available, we'll set our device to GPU. We'll use this device variable later in our code.
+if is_cuda:
+    device = torch.device("cuda")
+    print("GPU is available")
+else:
+    device = torch.device("cpu")
+    print("GPU not available, CPU used")
+
+#model is your own, defined NN
+model.to(device)

loss_function/Analyse_stats.py

+import torch
+from torch import autograd, nn, optim
+import torch.nn.functional as F
+from tqdm import tqdm
+import cv2
+import numpy as np
+import matplotlib.pyplot as plt
+
+
+if __name__ == "__main__":
+    f = open("D:\\Studium\\Bachelorarbeit\\Machine Learning\\loss_function\\complete_statistics_changing_clusterweight_gapweight5000.txt", "r")
+    # Using readlines()
+    Lines = f.readlines()
+    in_result = 0
+    tmp_array2 = ''
+    # Strips the newline character
+    for line in Lines:
+        if(in_result == 1):
+            if "]" in line:
+                tmp_array = line.split("]")
+                tmp_array2 += tmp_array[0]
+                in_result = 0
+                array = tmp_array2.split(".,")
+                result = []
+                for value in array:
+                    if "\n" in value:
+                        result.append(int(value.split("\n        ")[1]))
+                    elif " " in value:
+                        if "." in value:
+                            result.append(int(value.split(" ")[1].split(".")[0]))
+                        else:
+                            result.append(int(value.split(" ")[1]))
+                    else:
+                        # if "." in value:
+                        #     result.append(int(value.split(".")[0]))
+                        # else:
+                        result.append(int(value))
+                #print(torch.tensor(result).view(10,10))
+                tmp_array2 = ''
+
+
+                #find the start_cluster_size
+                result = torch.tensor(result, dtype=torch.float32).view(10,10)
+                img = result.round()
+                img = img.detach().numpy()
+                img = np.array(img, dtype=np.uint8)
+                num_labels, labels_im = cv2.connectedComponents(img)
+
+                start_value = labels_im[4][6]
+                end_value = labels_im[0][2]
+
+                #remodel the matrix, so that distance_transform can be used
+                start_cluster = []
+
+                if(start_value != 0 and end_value != 0):
+                    column_nr = len(labels_im)
+                    row_nr = len(labels_im[0])
+                    for i in range(0, column_nr):
+                        for j in range(0, row_nr):
+                            if (labels_im[i][j] == end_value):
+                                labels_im[i][j] = 0
+                            elif (labels_im[i][j] == 0):
+                                labels_im[i][j] = end_value
+                            elif (labels_im[i][j] == start_value):
+                                start_cluster.append([i,j])
+
+                            if(end_value == start_value):
+                                if(labels_im[i][j] == 0):
+                                    start_cluster.append([i,j])
+                print(len(start_cluster))
+
+            else:
+                tmp_array2 += line
+        # if "min_distance" in line:
+        #     #tmp_array = line.split("number_of_pixels: tensor(")
+        #     tmp_array = line.split("min_distance: ")
+        #     tmp_array2 = tmp_array[1].split(",")
+        #     tmp_array2 = tmp_array2[0].split(".")
+        #     print(tmp_array2[0])
+        if "result_rounded:" in line:
+            #tmp_array = line.split("number_of_pixels: tensor(")
+            tmp_array = line.split("result_rounded: tensor([")
+            tmp_array2 = tmp_array[1]
+            in_result = 1

loss_function/own_loss_function_cv2.py

+import torch
+from torch import autograd, nn, optim
+import torch.nn.functional as F
+from tqdm import tqdm
+import cv2
+import numpy as np
+import matplotlib.pyplot as plt
+
+class Net(nn.Module):
+    def __init__(self):
+        super(Net, self).__init__()
+
+        self.fc1 = nn.Linear(100, 250)
+        self.fc2 = nn.Linear(250, 512)
+        self.fc3 = nn.Linear(512, 250)
+        self.fc4 = nn.Linear(250, 100)
+
+    # x represents our data
+    def forward(self, x):
+        # Pass data through fc1
+        x = self.fc1(x)
+        x = F.relu(x)
+
+        x = self.fc2(x)
+        x = F.relu(x)
+
+        x = self.fc3(x)
+        x = F.relu(x)
+
+        x = self.fc4(x)
+
+        x = torch.sigmoid(x)
+        return x
+
+    def num_flat_features(self, x):
+        size = x.size()[1:]  # all dimensions except the batch dimension
+        num_features = 1
+        for s in size:
+            num_features *= s
+        return num_features
+
+class CustomLoss(nn.Module):
+    def init(self):
+        super(CustomLoss,self).__init__()
+
+    def forward(self, result_given, points_given):
+        #variables for easier variation of the loss function
+        weight = 100
+        gap_weight = 10
+        cluster_size_weight = 90
+        result_size = result_given.size()
+        loss = torch.tensor([0], dtype=torch.float32)
+
+        for i in range(0, result_size[0]):
+            result = result_given[i, 0, 0:, 0:]
+            points = points_given[i]
+
+            loss = torch.tensor([0], dtype=torch.float32)
+
+
+            manhattan_distance_start_end = self.estimate_manhattan_distance(points[0][0], points[0][1], points[1][0], points[1][1])
+
+            soa_cells = torch.tensor([0], dtype=torch.float32)                      #sum of all cells
+            soa_cells_inv = torch.tensor([0], dtype=torch.float32)                  #sum of all cells inverted
+            for row in result:
+                for column in row:
+                    soa_cells += column
+                    soa_cells_inv += (1-column)
+
+
+            #set the start and endpoint to 1
+            loss_start = (2-(result[points[0][0]][points[0][1]] + result[points[1][0]][points[1][1]])) * weight
+            loss_start = loss_start.view(1)
+
+            #first compute the clusters
+            img = result.round()
+            img = img.detach().numpy()
+            img = np.array(img, dtype=np.uint8)
+            num_labels, labels_im = cv2.connectedComponents(img)
+
+            start_value = labels_im[points[0][0]][points[0][1]]
+            end_value = labels_im[points[1][0]][points[1][1]]
+
+            #remodel the matrix, so that distance_transform can be used
+            start_cluster = []
+            gap_loss = torch.tensor([0], dtype=torch.float32)
+
+            if(start_value != 0 and end_value != 0):
+                column_nr = len(labels_im)
+                row_nr = len(labels_im[0])
+                for i in range(0, column_nr):
+                    for j in range(0, row_nr):
+                        if (labels_im[i][j] == end_value):
+                            labels_im[i][j] = 0
+                        elif (labels_im[i][j] == 0):
+                            labels_im[i][j] = end_value
+                        elif (labels_im[i][j] == start_value):
+                            start_cluster.append([i,j])
+
+                        if(end_value == start_value):
+                            if(labels_im[i][j] == 0):
+                                start_cluster.append([i,j])
+
+                # perform the distance transform
+                labels_im = np.array(labels_im, dtype=np.uint8)
+                dist_img = cv2.distanceTransform(labels_im, distanceType=cv2.DIST_L1, maskSize=3).astype(np.float32)
+
+                #get the min distance to the end_cluster from the start_cluster
+                min_distance = dist_img[points[0][0]][points[0][1]]
+                for cell in start_cluster:
+                    if dist_img[cell[0]][cell[1]] < min_distance:
+                        min_distance = dist_img[cell[0]][cell[1]]
+
+                gap_loss = min_distance * soa_cells_inv * gap_weight * soa_cells_inv
+
+            else:
+                gap_loss = (2-(result[points[0][0]][points[0][1]] + result[points[1][0]][points[1][1]])) * weight
+                gap_loss = gap_loss.view(1)
+
+            loss = torch.cat((loss, loss_start), 0)
+
+            cluster_cells = torch.tensor([0], dtype=torch.float32)                      #sum of all cluster_cells
+            for cell in start_cluster:
+                cluster_cells += result[cell[0]][cell[1]]
+
+
+            if(len(start_cluster) != 0):
+                cluster_size_penalty = cluster_cells * cluster_size_weight
+            else:
+                cluster_size_penalty = (2-(result[points[0][0]][points[0][1]] + result[points[1][0]][points[1][1]])) * weight
+                cluster_size_penalty = cluster_size_penalty.view(1)
+
+        return sum(loss), gap_loss, cluster_size_penalty
+
+
+
+    def estimate_manhattan_distance(self, start_x, start_y, end_x, end_y):
+        return abs(end_x - start_x) + abs(end_y - start_y)
+
+
+
+
+def get_data():
+        f = open("D:\\Studium\\Bachelorarbeit\\Unity Projekte\\A-Stern Test\\A-Stern Test\\Assets\\Resources\\single_example_data.txt", "r")
+        # Using readlines()
+        Lines = f.readlines()
+
+        weightmatrix = []
+        points = []
+        target_data_return = []
+
+        count = 0
+        for line in Lines:
+            elements = line.split(";")
+
+            #get the weightmatrix
+            x = elements[0].split(", ")
+            x.remove("")
+            for i in range(0, len(x), 1):
+                x[i] = int(x[i])
+
+            weightmatrix.append(x)
+
+            #get the start and endpoints
+            y = elements[1].split(" ")
+            for i in range(0, len(y), 1):
+                y[i] = y[i].strip('[]')
+                y[i] = y[i].split(",")
+                for j in range(0, len(y[i]), 1):
+                    y[i][j] = int(y[i][j])
+
+            points.append(torch.tensor(y))
+
+            #get the path and transform it into a matrix
+            z = elements[2].split(", ")
+            z.remove("")
+
+            for i in range(0, len(z), 1):
+                z[i] = z[i].strip('()')
+                z[i] = z[i].split(",")
+                for j in range(0, len(z[i]), 1):
+                    z[i][j] = int(z[i][j])
+
+            target_data = [[0, 0, 0, 0, 0, 0, 0, 0, 0, 0],
+                        [0, 0, 0, 0, 0, 0, 0, 0, 0, 0],
+                        [0, 0, 0, 0, 0, 0, 0, 0, 0, 0],
+                        [0, 0, 0, 0, 0, 0, 0, 0, 0, 0],
+                        [0, 0, 0, 0, 0, 0, 0, 0, 0, 0],
+                        [0, 0, 0, 0, 0, 0, 0, 0, 0, 0],
+                        [0, 0, 0, 0, 0, 0, 0, 0, 0, 0],
+                        [0, 0, 0, 0, 0, 0, 0, 0, 0, 0],
+                        [0, 0, 0, 0, 0, 0, 0, 0, 0, 0],
+                        [0, 0, 0, 0, 0, 0, 0, 0, 0, 0]]              #torch.zeros([10, 10])
+            for i in range(0, len(z), 1):
+                target_data[z[i][0]][z[i][1]] = 1
+            target_data[y[1][0]][y[1][1]] = 1
+            target_data_return.append(target_data)
+
+            count += 1
+        f.close()
+
+        return weightmatrix, points, target_data_return, count
+
+
+
+weightmatrix, points, target_data, batch_size = get_data()
+weightmatrix = torch.tensor(weightmatrix, dtype=torch.float32)
+weightmatrix = weightmatrix.view(batch_size, 1,10,10)
+
+#mark start and end in the weightmatrix
+for i in range(0, batch_size):
+    weightmatrix[i, 0, points[i][0][0], points[i][0][1]] += 1000
+    weightmatrix[i, 0, points[i][1][0], points[i][1][1]] += 1000
+
+points = [[[3,0],
+        [7,4]]]
+
+
+random_data = autograd.Variable(torch.rand(1, 1, 10, 10))
+model_path = 'D:\\Studium\\Bachelorarbeit\\Machine Learning\\ressources\\models'
+
+criterion = CustomLoss()
+
+my_nn = Net()
+
+opt = optim.SGD(params=my_nn.parameters(), lr=0.1)
+
+result = my_nn(random_data.view(batch_size,1, 1,100))
+
+print(result.view(10,10))
+
+for epoch in tqdm(range(75)):
+    result = my_nn(random_data.view(batch_size,1,1, 100))
+
+    start_loss, gap_loss, cluster_size_loss = criterion(result_given=result.view(batch_size,1,10,10), points_given=points)
+
+
+    my_nn.zero_grad()
+    start_loss.backward(retain_graph=True)
+    gap_loss.backward(retain_graph=True)
+    cluster_size_loss.backward()
+    opt.step()
+    # torch.save(my_nn.state_dict(),
+    #             model_path + "\\own_loss_model_nr_" + "2" + ".pth")
+
+
+print("loss" , start_loss + gap_loss + cluster_size_loss)
+print(result.view(10,10))
+#print(torch.round(result).view(10,10))
+img = result.round().view(10,10)
+imgplot = plt.matshow(img.detach().numpy())
+plt.show()

loss_function/own_loss_function_cv2_2.py

+import torch
+from torch import autograd, nn, optim
+import torch.nn.functional as F
+from tqdm import tqdm
+import cv2
+import numpy as np
+import matplotlib.pyplot as plt
+
+class Net(nn.Module):
+    def __init__(self):
+        super(Net, self).__init__()
+
+        self.fc1 = nn.Linear(100, 250)
+        self.fc2 = nn.Linear(250, 512)
+        self.fc3 = nn.Linear(512, 250)
+        self.fc4 = nn.Linear(250, 100)
+
+    # x represents our data
+    def forward(self, x):
+        # Pass data through fc1
+        x = self.fc1(x)
+        x = F.relu(x)
+
+        x = self.fc2(x)
+        x = F.relu(x)
+
+        x = self.fc3(x)
+        x = F.relu(x)
+
+        x = self.fc4(x)
+
+        x = torch.sigmoid(x)
+        return x
+
+    def num_flat_features(self, x):
+        size = x.size()[1:]  # all dimensions except the batch dimension
+        num_features = 1
+        for s in size:
+            num_features *= s
+        return num_features
+
+class CustomLoss(nn.Module):
+    def init(self):
+        super(CustomLoss,self).__init__()
+
+    def forward(self, result_given, points_given):
+        #variables for easier variation of the loss function
+        weight = 10000
+        gap_weight = 40 #35
+        cluster_size_weight = 3
+        result_size = result_given.size()
+
+        for i in range(0, result_size[0]):
+            result = result_given[i, 0, 0:, 0:]
+            points = points_given[i]
+
+            manhattan_distance_start_end = self.estimate_manhattan_distance(points[0][0], points[0][1], points[1][0], points[1][1])
+
+            soa_cells = sum(sum(result))
+            soa_cells_inv = 100 - sum(sum(result))
+
+            #set the start and endpoint to 1
+            loss_start = torch.tensor([0], dtype=torch.float32, requires_grad = True)
+            if(result[points[0][0]][points[0][1]].round() == 0 or result[points[1][0]][points[1][1]] == 0):
+                loss_start = (2-(result[points[0][0]][points[0][1]] + result[points[1][0]][points[1][1]])) * weight
+                loss_start = loss_start.view(1)
+
+            #first compute the clusters
+            img = result.round()
+            img = img.detach().numpy()
+            img = np.array(img, dtype=np.uint8)
+            num_labels, labels_im = cv2.connectedComponents(img)
+
+            start_value = labels_im[points[0][0]][points[0][1]]
+            end_value = labels_im[points[1][0]][points[1][1]]
+
+            #remodel the matrix, so that distance_transform can be used
+            start_cluster = []
+            gap_loss = torch.tensor([0], dtype=torch.float32)
+
+            if(start_value != 0 and end_value != 0):
+                column_nr = len(labels_im)
+                row_nr = len(labels_im[0])
+                for i in range(0, column_nr):
+                    for j in range(0, row_nr):
+                        if (labels_im[i][j] == end_value):
+                            labels_im[i][j] = 0
+                        elif (labels_im[i][j] == 0):
+                            labels_im[i][j] = end_value
+                        elif (labels_im[i][j] == start_value):
+                            start_cluster.append([i,j])
+
+                        if(end_value == start_value):
+                            if(labels_im[i][j] == 0):
+                                start_cluster.append([i,j])
+
+                # perform the distance transform
+                labels_im = np.array(labels_im, dtype=np.uint8)
+                dist_img = cv2.distanceTransform(labels_im, distanceType=cv2.DIST_L1, maskSize=3).astype(np.float32)
+
+                #get the min distance to the end_cluster from the start_cluster
+                min_distance = dist_img[points[0][0]][points[0][1]]
+                for cell in start_cluster:
+                    if dist_img[cell[0]][cell[1]] < min_distance:
+                        min_distance = dist_img[cell[0]][cell[1]]
+
+                gap_loss = min_distance * soa_cells_inv * gap_weight
+            else:
+                gap_loss = (2-(result[points[0][0]][points[0][1]] + result[points[1][0]][points[1][1]])) * weight
+                gap_loss = gap_loss.view(1)
+
+            cluster_size_penalty = torch.tensor([0], dtype=torch.float32, requires_grad = True)
+            cluster_size_penalty = soa_cells * cluster_size_weight * abs(manhattan_distance_start_end - len(start_cluster))
+
+        return loss_start + cluster_size_penalty + gap_loss
+
+
+    def estimate_manhattan_distance(self, start_x, start_y, end_x, end_y):
+        return abs(end_x - start_x) + abs(end_y - start_y)
+
+
+
+
+def get_data():
+        f = open("D:\\Studium\\Bachelorarbeit\\Unity Projekte\\A-Stern Test\\A-Stern Test\\Assets\\Resources\\single_example_data.txt", "r")
+        # Using readlines()
+        Lines = f.readlines()
+
+        weightmatrix = []
+        points = []
+        target_data_return = []
+
+        count = 0
+        # Strips the newline character
+        for line in Lines:
+            elements = line.split(";")
+
+            #get the weightmatrix
+            x = elements[0].split(", ")
+            x.remove("")
+            for i in range(0, len(x), 1):
+                x[i] = int(x[i])
+
+            weightmatrix.append(x)
+
+            #get the start and endpoints
+            y = elements[1].split(" ")
+            for i in range(0, len(y), 1):
+                y[i] = y[i].strip('[]')
+                y[i] = y[i].split(",")
+                for j in range(0, len(y[i]), 1):
+                    y[i][j] = int(y[i][j])
+
+            points.append(torch.tensor(y))
+
+            #get the path and transform it into a matrix
+            z = elements[2].split(", ")
+            z.remove("")
+
+            for i in range(0, len(z), 1):
+                z[i] = z[i].strip('()')
+                z[i] = z[i].split(",")
+                for j in range(0, len(z[i]), 1):
+                    z[i][j] = int(z[i][j])
+
+            target_data = [[0, 0, 0, 0, 0, 0, 0, 0, 0, 0],
+                        [0, 0, 0, 0, 0, 0, 0, 0, 0, 0],
+                        [0, 0, 0, 0, 0, 0, 0, 0, 0, 0],
+                        [0, 0, 0, 0, 0, 0, 0, 0, 0, 0],
+                        [0, 0, 0, 0, 0, 0, 0, 0, 0, 0],
+                        [0, 0, 0, 0, 0, 0, 0, 0, 0, 0],
+                        [0, 0, 0, 0, 0, 0, 0, 0, 0, 0],
+                        [0, 0, 0, 0, 0, 0, 0, 0, 0, 0],
+                        [0, 0, 0, 0, 0, 0, 0, 0, 0, 0],
+                        [0, 0, 0, 0, 0, 0, 0, 0, 0, 0]]              #torch.zeros([10, 10])
+            for i in range(0, len(z), 1):
+                target_data[z[i][0]][z[i][1]] = 1
+            target_data[y[1][0]][y[1][1]] = 1
+            target_data_return.append(target_data)
+
+            count += 1
+        f.close()
+
+        return weightmatrix, points, target_data_return, count
+
+
+
+weightmatrix, points, target_data, batch_size = get_data()
+weightmatrix = torch.tensor(weightmatrix, dtype=torch.float32)
+weightmatrix = weightmatrix.view(batch_size, 1,10,10)
+
+# points = torch.tensor(points, dtype=torch.float32)
+# points = points.view(batch_size, 1,10,10)
+
+
+
+#mark start and end in the weightmatrix
+for i in range(0, batch_size):
+    weightmatrix[i, 0, points[i][0][0], points[i][0][1]] += 1000
+    weightmatrix[i, 0, points[i][1][0], points[i][1][1]] += 1000
+
+points = [[[3,0],
+        [7,4]]]
+
+
+
+random_data = autograd.Variable(torch.rand(1, 1, 10, 10),requires_grad = True)
+model_path = 'D:\\Studium\\Bachelorarbeit\\Machine Learning\\ressources\\models'
+
+criterion = CustomLoss()
+
+my_nn = Net()
+# comment this lines in to evaluate the model
+#my_nn.load_state_dict(torch.load("D:\\Studium\\Bachelorarbeit\\Machine Learning\\ressources\\models\\own_loss_model_nr_2.pth"))
+#my_nn.eval()
+
+#an optimizer can be chosen
+# opt = optim.SGD(params=my_nn.parameters(), lr=0.01)
+opt = optim.Adam(params=my_nn.parameters(), lr=0.01)
+# opt = optim.AdamW(params=my_nn.parameters(), lr=0.01)
+# opt = optim.ASGD(params=my_nn.parameters(), lr=0.01)
+
+result = my_nn(random_data.view(batch_size,1, 1,100))
+
+print(result.view(10,10))
+
+for epoch in tqdm(range(15000)):
+    result = my_nn(random_data.view(batch_size,1,1, 100))
+
+    loss = criterion(result_given=result.view(batch_size,1,10,10), points_given=points)
+
+    my_nn.zero_grad()
+
+    loss.backward()
+    #Comment the following lines in to see the gradients of the neural net. The gradients of the last layer(4) can be used as an approximation for the change of the result of the NN.
+    # if(sum(my_nn.fc1.bias.grad) != 0):
+    #     print(my_nn.fc1.bias.grad)
+    # if(sum(my_nn.fc2.bias.grad) != 0):
+    #     print(my_nn.fc2.bias.grad)
+    # if(sum(my_nn.fc3.bias.grad) != 0):
+    #     print(my_nn.fc3.bias.grad)
+
+    # if(sum(my_nn.fc4.bias.grad) != 0):
+    #     pass
+        #print(my_nn.fc4.weight.grad)
+        #weights = my_nn.fc4.bias.grad *1000
+    #else:
+        # print(len(weights))
+        # weights_img = weights.view(250,100)
+        # weights_img = weights.view(10,10)
+        # weights_imgplot = plt.matshow(weights_img.detach().numpy())
+        # plt.show()
+    opt.step()
+
+    #Comment this lines in to save the model after every trainingsiteration
+    # torch.save(my_nn.state_dict(),
+    #             model_path + "\\own_loss_model_nr_" + "2" + ".pth")
+
+print("loss" , loss)
+print(result.view(10,10))
+
+#Plot the result of the NN
+img = result.round().view(10,10)
+imgplot = plt.matshow(img.detach().numpy())
+plt.show()

loss_function/own_loss_function_cv2_3.py

+import torch
+from torch import autograd, nn, optim
+import torch.nn.functional as F
+from tqdm import tqdm
+import cv2
+import numpy as np
+import matplotlib.pyplot as plt
+
+class Net(nn.Module):
+    def __init__(self):
+        super(Net, self).__init__()
+
+        self.fc1 = nn.Linear(100, 250)
+        self.fc2 = nn.Linear(250, 512)
+        self.fc3 = nn.Linear(512, 250)
+        self.fc4 = nn.Linear(250, 100)
+
+    # x represents our data
+    def forward(self, x):
+        # Pass data through fc1
+        x = self.fc1(x)
+        x = F.relu(x)
+
+        x = self.fc2(x)
+        x = F.relu(x)
+
+        x = self.fc3(x)
+        x = F.relu(x)
+
+        x = self.fc4(x)
+
+        x = torch.sigmoid(x)
+        return x
+
+    def num_flat_features(self, x):
+        size = x.size()[1:]  # all dimensions except the batch dimension
+        num_features = 1
+        for s in size:
+            num_features *= s
+        return num_features
+
+class CustomLoss(nn.Module):
+    def init(self):
+        super(CustomLoss,self).__init__()
+
+    def forward(self, result_given, points_given, weightmatrix_given):
+        #variables for easier variation of the loss function
+        weight = 20000
+        gap_weight = 3000 #1200
+        cluster_size_weight = 3
+        weight_weight = 1.1  #0.5
+        result_size = result_given.size()
+        loss = torch.tensor([0], dtype=torch.float32, requires_grad = True)
+
+        for i in range(0, result_size[0]):
+            result = result_given[i, 0, 0:, 0:]
+            weightmatrix = weightmatrix_given[i, 0, 0:, 0:]
+
+            points = points_given[i]
+
+            manhattan_distance_start_end = self.estimate_manhattan_distance(points[0][0], points[0][1], points[1][0], points[1][1])
+
+            soa_cells = sum(sum(result))
+            soa_cells_inv = 100 - sum(sum(result))
+
+            #set the start and endpoint to 1
+            loss_start = torch.tensor([0], dtype=torch.float32, requires_grad = True)
+            if(result[points[0][0]][points[0][1]].round() == 0 or result[points[1][0]][points[1][1]] == 0):
+                loss_start = (2-(result[points[0][0]][points[0][1]] + result[points[1][0]][points[1][1]])) * weight
+                loss_start = loss_start.view(1)
+
+            #first compute the clusters
+            img = result.round()
+            img = img.detach().numpy()
+            img = np.array(img, dtype=np.uint8)
+            num_labels, labels_im = cv2.connectedComponents(img)
+
+            start_value = labels_im[points[0][0]][points[0][1]]
+            end_value = labels_im[points[1][0]][points[1][1]]
+
+            #remodel the matrix, so that distance_transform can be used
+            start_cluster = []
+            gap_loss = torch.tensor([0], dtype=torch.float32)
+
+            if(start_value != 0 and end_value != 0):
+                column_nr = len(labels_im)
+                row_nr = len(labels_im[0])
+                for i in range(0, column_nr):
+                    for j in range(0, row_nr):
+                        if (labels_im[i][j] == end_value):
+                            labels_im[i][j] = 0
+                        elif (labels_im[i][j] == 0):
+                            labels_im[i][j] = end_value
+                        elif (labels_im[i][j] == start_value):
+                            start_cluster.append([i,j])
+
+                        if(end_value == start_value):
+                            if(labels_im[i][j] == 0):
+                                start_cluster.append([i,j])
+
+                # perform the distance transform
+                labels_im = np.array(labels_im, dtype=np.uint8)
+                dist_img = cv2.distanceTransform(labels_im, distanceType=cv2.DIST_L1, maskSize=3).astype(np.float32)
+                #print(dist_img[points[0][0]][points[0][1]])
+
+                #get the min distance to the end_cluster from the start_cluster
+                min_distance = dist_img[points[0][0]][points[0][1]]
+                for cell in start_cluster:
+                    if dist_img[cell[0]][cell[1]] < min_distance:
+                        min_distance = dist_img[cell[0]][cell[1]]
+
+                gap_loss = min_distance * soa_cells_inv * gap_weight
+            else:
+                gap_loss = (2-(result[points[0][0]][points[0][1]] + result[points[1][0]][points[1][1]])) * weight
+                gap_loss = gap_loss.view(1)
+
+            cluster_size_penalty = torch.tensor([0], dtype=torch.float32, requires_grad = True)
+            cluster_size_penalty = sum((result*weightmatrix).view(100)) * weight_weight * abs(manhattan_distance_start_end - len(start_cluster))
+
+            #Concatenate the loss with the other losses from the batch
+            loss = torch.cat((loss, loss_start + gap_loss + cluster_size_penalty), 0)
+
+        #return the mean of all losses over the batch
+        return sum(loss)/result_size[0]
+
+
+    def estimate_manhattan_distance(self, start_x, start_y, end_x, end_y):
+        return abs(end_x - start_x) + abs(end_y - start_y)
+
+
+
+
+def get_data():
+        f = open("D:\\Studium\\Bachelorarbeit\\Unity Projekte\\A-Stern Test\\A-Stern Test\\Assets\\Resources\\200_training_data.txt", "r")
+        # Using readlines()
+        Lines = f.readlines()
+
+        weightmatrix = []
+        points = []
+        target_data_return = []
+
+        count = 0
+        # Strips the newline character
+        for line in Lines:
+            elements = line.split(";")
+
+            #get the weightmatrix
+            x = elements[0].split(", ")
+            x.remove("")
+            for i in range(0, len(x), 1):
+                x[i] = int(x[i])
+
+            weightmatrix.append(x)
+
+            #get the start and endpoints
+            y = elements[1].split(" ")
+            for i in range(0, len(y), 1):
+                y[i] = y[i].strip('[]')
+                y[i] = y[i].split(",")
+                for j in range(0, len(y[i]), 1):
+                    y[i][j] = int(y[i][j])
+
+            points.append(torch.tensor(y))
+
+            #get the path and transform it into a matrix
+            z = elements[2].split(", ")
+            z.remove("")
+
+            for i in range(0, len(z), 1):
+                z[i] = z[i].strip('()')
+                z[i] = z[i].split(",")
+                for j in range(0, len(z[i]), 1):
+                    z[i][j] = int(z[i][j])
+
+            target_data = [[0, 0, 0, 0, 0, 0, 0, 0, 0, 0],
+                        [0, 0, 0, 0, 0, 0, 0, 0, 0, 0],
+                        [0, 0, 0, 0, 0, 0, 0, 0, 0, 0],
+                        [0, 0, 0, 0, 0, 0, 0, 0, 0, 0],
+                        [0, 0, 0, 0, 0, 0, 0, 0, 0, 0],
+                        [0, 0, 0, 0, 0, 0, 0, 0, 0, 0],
+                        [0, 0, 0, 0, 0, 0, 0, 0, 0, 0],
+                        [0, 0, 0, 0, 0, 0, 0, 0, 0, 0],
+                        [0, 0, 0, 0, 0, 0, 0, 0, 0, 0],
+                        [0, 0, 0, 0, 0, 0, 0, 0, 0, 0]]              #torch.zeros([10, 10])
+            for i in range(0, len(z), 1):
+                target_data[z[i][0]][z[i][1]] = 1
+            target_data[y[1][0]][y[1][1]] = 1
+            target_data_return.append(target_data)
+
+            count += 1
+        f.close()
+
+        return weightmatrix, points, target_data_return, count
+
+
+
+weightmatrix, points, target_data, batch_size = get_data()
+weightmatrix = torch.tensor(weightmatrix, dtype=torch.float32)
+weightmatrix = weightmatrix.view(batch_size, 1,10,10)
+
+
+
+random_data = autograd.Variable(torch.rand(1, 1, 10, 10),requires_grad = True)
+model_path = 'D:\\Studium\\Bachelorarbeit\\Machine Learning\\ressources\\models'
+
+criterion = CustomLoss()
+
+my_nn = Net()
+# comment this lines in to evaluate the model
+#my_nn.load_state_dict(torch.load("D:\\Studium\\Bachelorarbeit\\Machine Learning\\ressources\\models\\own_loss_model_nr_2.pth"))
+#my_nn.eval()
+
+
+#an optimizer can be chosen
+# opt = optim.SGD(params=my_nn.parameters(), lr=0.01)
+opt = optim.Adam(params=my_nn.parameters(), lr=0.01)
+# opt = optim.AdamW(params=my_nn.parameters(), lr=0.01)
+# opt = optim.ASGD(params=my_nn.parameters(), lr=0.01)
+
+result = my_nn(weightmatrix.view(batch_size,1, 1,100))
+
+for epoch in tqdm(range(15000)):
+    result = my_nn(weightmatrix.view(batch_size,1,1, 100))
+    loss = criterion(result_given=result.view(batch_size,1,10,10), points_given=points, weightmatrix_given=weightmatrix)
+
+
+    my_nn.zero_grad()
+
+    loss.backward()
+    #Comment the following lines in to see the gradients of the neural net. The gradients of the last layer(4) can be used as an approximation for the change of the result of the NN.
+    # if(sum(my_nn.fc1.bias.grad) != 0):
+    #     print(my_nn.fc1.bias.grad)
+    # if(sum(my_nn.fc2.bias.grad) != 0):
+    #     print(my_nn.fc2.bias.grad)
+    # if(sum(my_nn.fc3.bias.grad) != 0):
+    #     print(my_nn.fc3.bias.grad)
+    # if(sum(my_nn.fc4.bias.grad) != 0):
+    #     #pass
+    #     print(my_nn.fc4.weight.grad)
+    #     weights = my_nn.fc4.bias.grad *1000
+        # #else:
+        #     #print(len(weights))
+        #     #weights_img = weights.view(250,100)
+        #     weights_img = weights.view(10,10)
+        #     weights_imgplot = plt.matshow(weights_img.detach().numpy())
+        #     plt.show()
+
+    #save the model
+    opt.step()
+    torch.save(my_nn.state_dict(),
+                model_path + "\\own_loss_model_nr_" + "4" + ".pth")
+
+# print("loss" , loss)
+# print(result.view(10,10))
+# #print(torch.round(result).view(10,10))
+# img = result.round().view(10,10)
+# imgplot = plt.matshow(img.detach().numpy())
+# plt.show()

loss_function/own_loss_function_stats.py

+#This script tracks the stats of the final neural net with an adopted lossfunction
+
+import torch
+from torch import autograd, nn, optim
+import torch.nn.functional as F
+from tqdm import tqdm
+import cv2
+import numpy as np
+import matplotlib.pyplot as plt
+
+class Net(nn.Module):
+    def __init__(self):
+        super(Net, self).__init__()
+
+        self.fc1 = nn.Linear(100, 250)
+        self.fc2 = nn.Linear(250, 512)
+        self.fc3 = nn.Linear(512, 250)
+        self.fc4 = nn.Linear(250, 100)
+
+    # x represents our data
+    def forward(self, x):
+        # Pass data through fc1
+        x = self.fc1(x)
+        x = F.relu(x)
+
+        x = self.fc2(x)
+        x = F.relu(x)
+
+        x = self.fc3(x)
+        x = F.relu(x)
+
+        x = self.fc4(x)
+
+        x = torch.sigmoid(x)
+        return x
+
+    def num_flat_features(self, x):
+        size = x.size()[1:]  # all dimensions except the batch dimension
+        num_features = 1
+        for s in size:
+            num_features *= s
+        return num_features
+
+class CustomLoss(nn.Module):
+    def init(self):
+        super(CustomLoss,self).__init__()
+
+    def forward(self, result_given, points_given, weightmatrix, weight_weight):
+        #variables for easier variation of the loss function
+        weight = 20000
+        gap_weight = 5000 #3000 #1200
+        cluster_size_weight = 3
+        result_size = result_given.size()
+
+        for i in range(0, result_size[0]):
+            result = result_given[i, 0, 0:, 0:]
+            points = points_given[i]
+
+            manhattan_distance_start_end = self.estimate_manhattan_distance(points[0][0], points[0][1], points[1][0], points[1][1])
+            min_distance = manhattan_distance_start_end
+
+            soa_cells = sum(sum(result))
+            soa_cells_inv = 100 - sum(sum(result))
+
+            #set the start and endpoint to 1
+            loss_start = torch.tensor([0], dtype=torch.float32, requires_grad = True)
+            if(result[points[0][0]][points[0][1]].round() == 0 or result[points[1][0]][points[1][1]] == 0):
+                loss_start = (2-(result[points[0][0]][points[0][1]] + result[points[1][0]][points[1][1]])) * weight
+                loss_start = loss_start.view(1)
+
+            #first compute the clusters
+            img = result.round()
+            img = img.detach().numpy()
+            img = np.array(img, dtype=np.uint8)
+            num_labels, labels_im = cv2.connectedComponents(img)
+
+            start_value = labels_im[points[0][0]][points[0][1]]
+            end_value = labels_im[points[1][0]][points[1][1]]
+
+            #remodel the matrix, so that distance_transform can be used
+            start_cluster = []
+            gap_loss = torch.tensor([0], dtype=torch.float32)
+
+            if(start_value != 0 and end_value != 0):
+                column_nr = len(labels_im)
+                row_nr = len(labels_im[0])
+                for i in range(0, column_nr):
+                    for j in range(0, row_nr):
+                        if (labels_im[i][j] == end_value):
+                            labels_im[i][j] = 0
+                        elif (labels_im[i][j] == 0):
+                            labels_im[i][j] = end_value
+                        elif (labels_im[i][j] == start_value):
+                            start_cluster.append([i,j])
+
+                        if(end_value == start_value):
+                            if(labels_im[i][j] == 0):
+                                start_cluster.append([i,j])
+
+                # perform the distance transform
+                labels_im = np.array(labels_im, dtype=np.uint8)
+                dist_img = cv2.distanceTransform(labels_im, distanceType=cv2.DIST_L1, maskSize=3).astype(np.float32)
+
+
+                #get the min distance to the end_cluster from the start_cluster
+                min_distance = dist_img[points[0][0]][points[0][1]]
+                for cell in start_cluster:
+                    if dist_img[cell[0]][cell[1]] < min_distance:
+                        min_distance = dist_img[cell[0]][cell[1]]
+                #print(min_distance)
+
+                gap_loss = min_distance * soa_cells_inv * gap_weight
+            else:
+                gap_loss = (2-(result[points[0][0]][points[0][1]] + result[points[1][0]][points[1][1]])) * weight
+                gap_loss = gap_loss.view(1)
+
+
+            cluster_size_penalty = torch.tensor([0], dtype=torch.float32, requires_grad = True)
+            cluster_size_penalty = sum((result*weightmatrix).view(100)) * weight_weight * abs(manhattan_distance_start_end - len(start_cluster))
+
+        return loss_start + cluster_size_penalty + gap_loss , min_distance
+
+
+    def estimate_manhattan_distance(self, start_x, start_y, end_x, end_y):
+        return abs(end_x - start_x) + abs(end_y - start_y)
+
+
+
+
+def get_data():
+        f = open("D:\\Studium\\Bachelorarbeit\\Unity Projekte\\A-Stern Test\\A-Stern Test\\Assets\\Resources\\single_example_data2.txt", "r")
+        # Using readlines()
+        Lines = f.readlines()
+
+        weightmatrix = []
+        points = []
+        target_data_return = []
+
+        count = 0
+        # Strips the newline character
+        for line in Lines:
+            elements = line.split(";")
+
+            #get the weightmatrix
+            x = elements[0].split(", ")
+            x.remove("")
+            for i in range(0, len(x), 1):
+                x[i] = int(x[i])
+
+            weightmatrix.append(x)
+
+            #get the start and endpoints
+            y = elements[1].split(" ")
+            for i in range(0, len(y), 1):
+                y[i] = y[i].strip('[]')
+                y[i] = y[i].split(",")
+                for j in range(0, len(y[i]), 1):
+                    y[i][j] = int(y[i][j])
+
+            points.append(torch.tensor(y))
+
+            #get the path and transform it into a matrix
+            z = elements[2].split(", ")
+            z.remove("")
+
+            for i in range(0, len(z), 1):
+                z[i] = z[i].strip('()')
+                z[i] = z[i].split(",")
+                for j in range(0, len(z[i]), 1):
+                    z[i][j] = int(z[i][j])
+
+            target_data = [[0, 0, 0, 0, 0, 0, 0, 0, 0, 0],
+                        [0, 0, 0, 0, 0, 0, 0, 0, 0, 0],
+                        [0, 0, 0, 0, 0, 0, 0, 0, 0, 0],
+                        [0, 0, 0, 0, 0, 0, 0, 0, 0, 0],
+                        [0, 0, 0, 0, 0, 0, 0, 0, 0, 0],
+                        [0, 0, 0, 0, 0, 0, 0, 0, 0, 0],
+                        [0, 0, 0, 0, 0, 0, 0, 0, 0, 0],
+                        [0, 0, 0, 0, 0, 0, 0, 0, 0, 0],
+                        [0, 0, 0, 0, 0, 0, 0, 0, 0, 0],
+                        [0, 0, 0, 0, 0, 0, 0, 0, 0, 0]]              #torch.zeros([10, 10])
+            for i in range(0, len(z), 1):
+                target_data[z[i][0]][z[i][1]] = 1
+            target_data[y[1][0]][y[1][1]] = 1
+            target_data_return.append(target_data)
+
+            count += 1
+        f.close()
+
+        return weightmatrix, points, target_data_return, count
+
+
+
+weightmatrix, points, target_data, batch_size = get_data()
+weightmatrix = torch.tensor(weightmatrix, dtype=torch.float32)
+weightmatrix = weightmatrix.view(batch_size, 1,10,10)
+
+
+
+random_data = autograd.Variable(torch.rand(1, 1, 10, 10),requires_grad = True)
+model_path = 'D:\\Studium\\Bachelorarbeit\\Machine Learning\\ressources\\models'
+
+criterion = CustomLoss()
+
+for i in range(0, 100, 1):
+    my_nn = Net()
+
+
+    #an optimizer can be chosen
+    # opt = optim.SGD(params=my_nn.parameters(), lr=0.01)
+    opt = optim.Adam(params=my_nn.parameters(), lr=0.01)
+    # opt = optim.AdamW(params=my_nn.parameters(), lr=0.01)
+    # opt = optim.ASGD(params=my_nn.parameters(), lr=0.01)
+
+    result = my_nn(random_data.view(batch_size,1, 1,100))
+
+    for epoch in tqdm(range(15000)):
+        result = my_nn(random_data.view(batch_size,1,1, 100))
+
+        loss, min_distance = criterion(result_given=result.view(batch_size,1,10,10), points_given=points, weightmatrix=weightmatrix, weight_weight=(i/10))
+
+        my_nn.zero_grad()
+
+        loss.backward()
+
+        opt.step()
+
+
+    print("loss" , loss)
+    print(result.view(10,10))
+    f = open("complete_statistics_changing_clusterweight_gapweight5000.txt", "a")
+    to_write = "clusterweight: " + str(i/10)  + ", min_distance: " + str(min_distance) + ", result: " + str(result.view(100)) + ", result_rounded: " + str(result.round().view(100)) + ", number_of_pixels: " + str(sum(result.round().view(100))) + "\n"
+    f.write(to_write)
+    f.close()

loss_function/statistics_changing_clusterweight.txt

+clusterweight 0: 0.0
+clusterweight 1: 0.0
+clusterweight 2: 0.0
+clusterweight 3: 0.0
+clusterweight 4: 0.0
+clusterweight 5: 0.0
+clusterweight 6: 0.0
+clusterweight 7: 0.0
+clusterweight 8: 0.0
+clusterweight 9: 0.0
+clusterweight 10: 2.0
+clusterweight 11: 0.0
+clusterweight 12: 0.0
+clusterweight 13: 2.0
+clusterweight 14: tensor(8)
+clusterweight 15: 0.0
+clusterweight 16: 0.0
+clusterweight 17: 3.0
+clusterweight 18: 0.0
+clusterweight 19: 0.0
+clusterweight 20: 0.0
+clusterweight 21: 2.0
+clusterweight 22: 0.0
+clusterweight 23: 0.0
+clusterweight 24: 0.0
+clusterweight 25: 0.0
+clusterweight 26: 3.0
+clusterweight 27: 2.0
+clusterweight 28: 0.0
+clusterweight 29: 3.0
+clusterweight 30: 2.0
+clusterweight 31: 0.0
+clusterweight 32: 0.0
+clusterweight 33: 3.0
+clusterweight 34: 0.0
+clusterweight 35: 2.0
+clusterweight 36: 0.0
+clusterweight 37: 0.0
+clusterweight 38: 2.0
+clusterweight 39: 0.0
+clusterweight 40: 8.0
+clusterweight 41: 3.0
+clusterweight 42: 0.0
+clusterweight 43: 8.0
+clusterweight 44: 0.0
+clusterweight 45: 3.0
+clusterweight 46: 8.0
+clusterweight 47: 0.0
+clusterweight 48: 0.0
+clusterweight 49: 0.0
+clusterweight 50: 8.0
+clusterweight 51: 2.0
+clusterweight 52: 2.0
+clusterweight 53: 8.0
+clusterweight 54: 2.0
+clusterweight 55: 2.0
+clusterweight 56: 3.0
+clusterweight 57: 2.0
+clusterweight 58: tensor(8)
+clusterweight 59: 0.0
+clusterweight 60: 2.0
+clusterweight 61: 2.0
+clusterweight 62: 2.0
+clusterweight 63: 2.0
+clusterweight 64: 2.0
+clusterweight 65: 8.0
+clusterweight 66: 2.0
+clusterweight 67: 8.0
+clusterweight 68: 2.0
+clusterweight 69: 2.0
+clusterweight 70: 0.0
+clusterweight 71: 0.0
+clusterweight 72: 2.0
+clusterweight 73: 3.0
+clusterweight 74: 2.0
+clusterweight 75: 0.0
+clusterweight 76: 0.0
+clusterweight 77: 2.0
+clusterweight 78: 0.0
+clusterweight 79: 2.0
+clusterweight 80: 2.0
+clusterweight 81: 0.0
+clusterweight 82: 0.0
+clusterweight 83: 2.0
+clusterweight 84: 3.0
+clusterweight 85: 2.0
+clusterweight 86: 2.0
+clusterweight 87: 2.0
+clusterweight 88: 2.0
+clusterweight 89: 2.0
+clusterweight 90: 8.0
+clusterweight 91: 2.0
+clusterweight 92: 8.0
+clusterweight 93: 2.0
+clusterweight 94: 8.0
+clusterweight 95: 2.0
+clusterweight 96: 2.0
+clusterweight 97: 2.0
+clusterweight 98: 2.0
+clusterweight 99: 8.0

loss_function/statistics_changing_gapweight.txt

+gapweight0: 8.0gapweight100: 4.0gapweight200: 3.0gapweight300: 2.0gapweight400: 2.0gapweight500: tensor(8)gapweight600: 0.0gapweight700: 2.0gapweight800: 2.0gapweight900: 2.0gapweight1000: 0.0gapweight1100: 0.0gapweight1200: 0.0gapweight1300: 2.0gapweight1400: 2.0gapweight1500: 0.0gapweight1600: 0.0gapweight1700: 0.0gapweight1800: 0.0gapweight1900: 0.0gapweight2000: 2.0gapweight2100: 0.0gapweight2200: 0.0gapweight2300: 0.0gapweight2400: 0.0gapweight2500: 0.0gapweight2600: 0.0gapweight2700: 3.0gapweight2800: tensor(8)gapweight2900: 0.0gapweight3000: 0.0gapweight3100: 0.0gapweight3200: 3.0gapweight3300: 0.0gapweight3400: 0.0gapweight3500: 0.0gapweight3600: 0.0gapweight3700: 0.0gapweight3800: 3.0gapweight3900: 0.0gapweight4000: 0.0gapweight4100: 0.0gapweight4200: 0.0gapweight4300: 0.0gapweight4400: 0.0gapweight4500: 0.0gapweight4600: 0.0gapweight4700: 0.0gapweight4800: 0.0gapweight4900: 2.0
\ No newline at end of file