Why are the visual representations of the neural network model being concatenated at the bottom of the output

Question

I am creating a graph neural network model for predicting the flow of water on a terrain. To train the model, I am passing a Digital Elevation Model file as input and extracting a rectangle out of the entire file for training it. Based on the area extracted, I am generating momentum and velocity of water in X and Y directions depending on the physical features. After training, while checking the visualization, I am observing that the arrows representing the flows are being concatenated at the bottom of the output.

# Load the DEM mesh file
mesh_file_path = '/content/drive/MyDrive/meshOp.obj'
mesh = trimesh.load(mesh_file_path)

# Define the bounding box
def calculate_bounding_box(center, size):

    half_size = size / 2
    x_min = center[0] - half_size
    x_max = center[0] + half_size
    y_min = center[1] - half_size
    y_max = center[1] + half_size
    
    return [x_min, y_min, x_max, y_max]

# Example: Set center and size
center = (50, 50)  
size = 200  # Size of the area in square units

# Calculate bounding box based on the given center and size
bounding_box = calculate_bounding_box(center, size)

# Process bounding box function with momentum calculation
def process_bounding_box(mesh, bounding_box):
    bbox_mask = (
        (mesh.vertices[:, 0] >= bounding_box[0]) &
        (mesh.vertices[:, 0] <= bounding_box[2]) &
        (mesh.vertices[:, 1] >= bounding_box[1]) &
        (mesh.vertices[:, 1] <= bounding_box[3])
    )

    bbox_vertices = mesh.vertices[bbox_mask]
    bbox_faces_remapped = []
    vertex_mapping = {old_index: new_index for new_index, old_index in enumerate(np.where(bbox_mask)[0])}
    for face in mesh.faces:
        if all(bbox_mask[face]):
            bbox_faces_remapped.append([vertex_mapping[vertex_index] for vertex_index in face])

    bbox_faces = bbox_faces_remapped
    elevation = bbox_vertices[:, 2]
    slopes = np.zeros(len(bbox_vertices))
    momentum = np.zeros((len(bbox_vertices), 2))  # Momentum in x and y directions

    # Constants for water momentum calculation
    water_density = 1000  # kg/m^3, typical density of water
    gravity = 9.81  # m/s^2, gravitational acceleration
    water_volume_per_vertex = 1.0  # m^3, assumed volume of water at each vertex

    for i, vertex in enumerate(bbox_vertices):
        adjacent_faces = [face for face in bbox_faces if i in face]
        face_slopes = []
        velocity_x, velocity_y = 0, 0
        for face in adjacent_faces:
            v0, v1, v2 = bbox_vertices[face]
            dist1 = distance.euclidean(v0[:2], v1[:2])
            dist2 = distance.euclidean(v1[:2], v2[:2])
            dist3 = distance.euclidean(v2[:2], v0[:2])
            height_diff1 = abs(v0[2] - v1[2])
            height_diff2 = abs(v1[2] - v2[2])
            height_diff3 = abs(v2[2] - v0[2])

            # Calculate slopes
            slope1 = height_diff1 / dist1 if dist1 != 0 else 0
            slope2 = height_diff2 / dist2 if dist2 != 0 else 0
            slope3 = height_diff3 / dist3 if dist3 != 0 else 0
            face_slopes.append(np.mean([slope1, slope2, slope3]))

            # Calculate potential energy converted to kinetic energy to estimate velocity
            velocity_x += np.sign(v1[0] - v0[0]) * gravity * slope1
            velocity_y += np.sign(v1[1] - v0[1]) * gravity * slope2
            velocity_x += np.sign(v2[0] - v1[0]) * gravity * slope2
            velocity_y += np.sign(v2[1] - v1[1]) * gravity * slope3
            velocity_x += np.sign(v0[0] - v2[0]) * gravity * slope3
            velocity_y += np.sign(v0[1] - v2[1]) * gravity * slope1

        if face_slopes:
            slopes[i] = np.mean(face_slopes)

        # Calculate velocity magnitude
        velocity_magnitude = np.sqrt(velocity_x**2 + velocity_y**2)

        # Calculate momentum
        mass = water_density * water_volume_per_vertex
        momentum[i] = mass * np.array([velocity_x, velocity_y]) / max(len(adjacent_faces), 1)

    print(f"Number of vertices in bounding box: {len(bbox_vertices)}")
    tri = Delaunay(bbox_vertices[:, :2])
    edges = []
    for simplex in tri.simplices:
        edges.extend([(simplex[i], simplex[j]) for i in range(len(simplex)) for j in range(i + 1, len(simplex))])

    # Normalize features
    node_features = np.column_stack((elevation, slopes, np.linalg.norm(momentum, axis=1)))

    # Scaling: Standardization (zero mean, unit variance)
    node_features_mean = node_features.mean(axis=0)
    node_features_std = node_features.std(axis=0)
    node_features = (node_features - node_features_mean) / node_features_std

    # Normalize target values
    momentum_mean = momentum.mean(axis=0)
    momentum_std = momentum.std(axis=0)
    momentum = (momentum - momentum_mean) / momentum_std

    node_features = torch.tensor(node_features, dtype=torch.float)
    edge_indices = torch.tensor(edges, dtype=torch.long).t().contiguous()
    momentum_targets = torch.tensor(momentum, dtype=torch.float)

    # Create a PyTorch Geometric data object
    return Data(x=node_features, edge_index=edge_indices, y=momentum_targets)

# Process the bounding box to get training data
data = process_bounding_box(mesh, bounding_box)

# Create a DataLoader
data_loader = DataLoader([data], batch_size=1, shuffle=True)

# Define the GCN model for flow prediction
class FlowGCN(torch.nn.Module):
    def __init__(self):
        super(FlowGCN, self).__init__()
        self.conv1 = GCNConv(3, 64)
        self.conv2 = GCNConv(64, 64)
        self.conv3 = GCNConv(64, 2)  # Output is a vector (momentum in x and y)

    def forward(self, x, edge_index):
        x = F.relu(self.conv1(x, edge_index))
        x = F.relu(self.conv2(x, edge_index))
        x = self.conv3(x, edge_index)
        return x

# Initialize the model, optimizer, and loss function
model = FlowGCN()
optimizer = optim.Adam(model.parameters(), lr=0.01)
loss_fn = nn.MSELoss()

# Training loop
def train(model, data_loader, optimizer, loss_fn, epochs=200):
    model.train()
    for epoch in range(epochs):
        for batch in data_loader:
            optimizer.zero_grad()
            out = model(batch.x, batch.edge_index)
            loss = loss_fn(out, batch.y)
            loss.backward()
            optimizer.step()
        if epoch % 10 == 0:
            print(f"Epoch {epoch}, Loss: {loss.item():.4f}")

# Train the model
train(model, data_loader, optimizer, loss_fn)

def aggregate_predictions_to_faces(vertices, faces, predictions):
    valid_faces = []
    face_predictions = []

    for i, face in enumerate(faces):
        if all(idx < len(predictions) for idx in face):
            # Aggregate vertex predictions for each face by averaging
            face_prediction = np.mean(predictions[face], axis=0)
            face_predictions.append(face_prediction)
            valid_faces.append(i)
    

    print(f"Total faces: {len(faces)}")
    print(f"Valid faces: {len(valid_faces)}")
    print(f"Predictions shape: {predictions.shape}")

    return np.array(face_predictions), valid_faces


# Visualization function updated
def visualize_flow(mesh, faces, face_predictions, valid_faces, title='Predicted Water Flow'):
    plt.figure(figsize=(10, 10))

    # Plot the mesh
    plt.triplot(mesh.vertices[:, 0], mesh.vertices[:, 1], faces, color='gray', alpha=0.5)

    # Scale factor for arrows to make them more visible
    scale_factor = 150.0

    # Define a subset of faces for visualization
    subset_indices = range(0, len(valid_faces), 10)  # Adjust step size as needed

    for i in subset_indices:
        face_index = valid_faces[i]
        face = faces[face_index]
        centroid = mesh.vertices[face].mean(axis=0)
        flow_vector = face_predictions[i]

        # Normalize the flow vector
        magnitude = np.linalg.norm(flow_vector)
        if magnitude > 0:
            flow_vector /= magnitude  # Normalize
            flow_vector *= scale_factor  # Scale

        # Quiver plot for better visualization
        plt.quiver(
            centroid[0], centroid[1],
            flow_vector[0], flow_vector[1],
            angles='xy', scale_units='xy', scale=1, color='blue'
        )

    plt.title(title)
    plt.xlabel('X')
    plt.ylabel('Y')
    plt.grid(True)
    plt.show()


# Test the model and visualize
model.eval()
with torch.no_grad():
    for batch in data_loader:
        predictions = model(batch.x, batch.edge_index).numpy()

        # Aggregate predictions to faces
        face_predictions, valid_faces = aggregate_predictions_to_faces(
            mesh.vertices, mesh.faces, predictions
        )

        visualize_flow(mesh, mesh.faces, face_predictions, valid_faces)

[this is the current output of the visualization where the arrows are concatenated at the bottom](https://i.sstatic.net/pBRI0vdf.png)

Please do let me know if I am taking the input wrong along with any other mistakes I might have made. Thanks for any useful insights in advance. I assumed that the flows will be localized around a set of co-ordinates. Could someone please check through the code and tell me what am I doing wrong as I am fairly new to these concepts.