Changed around line 1
+ # kan-lab1.scroll.pub
+ Website generated by Claude from prompt: Including the code in your prompt is a good idea if you want the developers or team working on the project to implement these specific examples in the backend. By providing the exact implementation, you ensure alignment with the article’s experiments and reduce ambiguity. Here’s how you can integrate it effectively: Improved Prompt with Code Integration: Objective: Develop a website that allows users to configure and experiment with parameters of the Attention-KAN approach detailed in this article: Kolmogorov-Arnold Networks: Exploring Dynamic Weights and Attention Mechanisms. Frontend Requirements: 1. Create an intuitive user interface for configuring the parameters of Kolmogorov-Arnold Networks (KAN) and its dynamic weight integration. • Include all input parameters discussed in the article, such as the number of inner/outer units, grid points, spline degrees, learning rate, and other relevant hyperparameters. • Allow users to adjust these parameters via sliders, dropdowns, or input fields. 2. Provide real-time visualizations: • Display metrics such as AUC, KS Statistic, and Log Loss after the experiments. • Generate a chart of the loss function over training epochs (X-axis = epochs, Y-axis = loss). • Optionally, include visualizations of network behavior or intermediate results if described in the article. Backend Requirements: 1. Implement the following example experiments provided in the article: • Example 1: Training and evaluation of a Dynamic Weight KAN. • Example 2: Training and evaluation of an Improved Standard KAN using spline-based utilities. 2. The backend should: • Use the provided PyTorch code to compute metrics (AUC, KS Statistic, Log Loss) based on user-configured parameters. • Log and output the training loss at each epoch for visualization in the frontend. 3. Ensure the backend is capable of processing user-configured input parameters and reflecting them in the experiment results. (Code provided below should be integrated into the backend implementation.) Below are the code snippet for Backend Implementation of 2 examples ### Code for Example 1: of Dynamic KAN Experiment ==================================================== import torch import torch.nn as nn import torch.optim as optim from sklearn.metrics import roc_auc_score, log_loss, roc_curve from sklearn.model_selection import train_test_split # Dynamic Weight Generator class DynamicWeightGenerator(nn.Module): def __init__(self, input_dim, output_dim): super(DynamicWeightGenerator, self).__init__() self.fc1 = nn.Linear(input_dim, 128) self.fc2 = nn.Linear(128, 64) self.fc_weight = nn.Linear(64, output_dim) self.fc_bias = nn.Linear(64, output_dim) self.softplus = nn.Softplus() def forward(self, x): h = torch.relu(self.fc1(x)) h = torch.relu(self.fc2(h)) weight = self.softplus(self.fc_weight(h)) bias = self.fc_bias(h) return weight, bias # Kolmogorov-Arnold Network (KAN) class KAN(nn.Module): def __init__(self, input_dim, num_inner_units, num_outer_units, dynamic=True): super(KAN, self).__init__() self.dynamic = dynamic self.num_inner_units = num_inner_units self.num_outer_units = num_outer_units if dynamic: self.inner_weight_generators = nn.ModuleList([ DynamicWeightGenerator(input_dim, input_dim) for _ in range(num_inner_units) ]) self.outer_weight_generators = nn.ModuleList([ DynamicWeightGenerator(input_dim, 1) for _ in range(num_outer_units) ]) else: self.inner_weights = nn.ParameterList([ nn.Parameter(torch.randn(input_dim, input_dim)) for _ in range(num_inner_units) ]) self.inner_biases = nn.ParameterList([ nn.Parameter(torch.randn(input_dim)) for _ in range(num_inner_units) ]) self.outer_weights = nn.ParameterList([ nn.Parameter(torch.randn(num_inner_units)) for _ in range(num_outer_units) ]) self.outer_biases = nn.ParameterList([ nn.Parameter(torch.randn(1)) for _ in range(num_outer_units) ]) def forward(self, x): inner_outputs = [] if self.dynamic: for generator in self.inner_weight_generators: weight, bias = generator(x) inner_output = torch.sum(weight * x + bias, dim=-1, keepdim=True) inner_outputs.append(inner_output) else: for weight, bias in zip(self.inner_weights, self.inner_biases): inner_output = torch.sum(torch.matmul(x, weight) + bias, dim=-1, keepdim=True) inner_outputs.append(inner_output) aggregated_inner_output = torch.cat(inner_outputs, dim=-1) outer_outputs = [] if self.dynamic: for generator in self.outer_weight_generators: weight, bias = generator(x) outer_output = torch.sum(weight * aggregated_inner_output + bias, dim=-1, keepdim=True) outer_outputs.append(outer_output) else: for weight, bias in zip(self.outer_weights, self.outer_biases): outer_output = torch.sum(aggregated_inner_output * weight + bias, dim=-1, keepdim=True) outer_outputs.append(outer_output) final_output = torch.sum(torch.cat(outer_outputs, dim=-1), dim=-1, keepdim=True) return final_output # Data generation torch.manual_seed(42) num_samples = 1200 input_dim = 5 # Generate complex non-linear data X = torch.rand((num_samples, input_dim)) * 10 - 5 noise = torch.randn(num_samples) * 0.5 Y = ((torch.sin(X[:, 0]) + torch.cos(X[:, 1]) + 0.3 * X[:, 2] - 0.2 * X[:, 3] ** 2 + noise) > 0).float().unsqueeze(1) # Split data X_train, X_val, Y_train, Y_val = train_test_split(X.numpy(), Y.numpy(), test_size=0.2, random_state=42) # Train Dynamic Weight KAN device = 'cuda' if torch.cuda.is_available() else 'cpu' dynamic_model = KAN(input_dim, num_inner_units=6, num_outer_units=4, dynamic=True).to(device) criterion = nn.BCELoss() optimizer = optim.Adam(dynamic_model.parameters(), lr=0.001) print("\nTraining Dynamic KAN") epochs = 60 for epoch in range(epochs): dynamic_model.train() optimizer.zero_grad() outputs = torch.sigmoid(dynamic_model(torch.tensor(X_train, dtype=torch.float32).to(device))) loss = criterion(outputs, torch.tensor(Y_train, dtype=torch.float32).to(device)) loss.backward() optimizer.step() if (epoch + 1) % 10 == 0: print(f"Dynamic KAN - Epoch [{epoch+1}/{epochs}], Loss: {loss.item():.4f}") # Evaluate Dynamic KAN dynamic_model.eval() with torch.no_grad(): dynamic_predictions = torch.sigmoid(dynamic_model(torch.tensor(X_val, dtype=torch.float32).to(device))).cpu().numpy() dynamic_auc = roc_auc_score(Y_val, dynamic_predictions) dynamic_fpr, dynamic_tpr, _ = roc_curve(Y_val, dynamic_predictions) dynamic_ks = max(dynamic_tpr - dynamic_fpr) dynamic_logloss = log_loss(Y_val, dynamic_predictions) print("\nDynamic KAN Metrics:") print(f"AUC: {dynamic_auc:.4f}") print(f"KS Statistic: {dynamic_ks:.4f}") print(f"Log Loss: {dynamic_logloss:.4f}") # Train Improved Standard KAN # Spline Utilities def B_batch(x, grid, k=3): x = x.unsqueeze(2) grid = grid.unsqueeze(0) if k == 0: value = (x >= grid[:, :, :-1]) * (x < grid[:, :, 1:]) else: B_km1 = B_batch(x[:, :, 0], grid[0], k=k - 1) value = ( (x - grid[:, :, :-(k + 1)]) / (grid[:, :, k:-1] - grid[:, :, :-(k + 1)]) * B_km1[:, :, :-1] + (grid[:, :, k + 1:] - x) / (grid[:, :, k + 1:] - grid[:, :, 1:-k]) * B_km1[:, :, 1:] ) return torch.nan_to_num(value) def coef2curve(x_eval, grid, coef, k): b_splines = B_batch(x_eval, grid, k) return torch.einsum("ijk,jlk->ijl", b_splines, coef.to(b_splines.device)) # Improved Standard KAN class StandardKAN(nn.Module): def __init__(self, input_dim, num_inner_units, num_outer_units, grid_points=20, spline_degree=3): super(StandardKAN, self).__init__() self.input_dim = input_dim self.num_inner_units = num_inner_units self.num_outer_units = num_outer_units self.spline_degree = spline_degree # Adaptive grids self.inner_grid = nn.Parameter(torch.linspace(-10, 10, steps=grid_points).repeat(input_dim, 1), requires_grad=False) self.outer_grid = nn.Parameter(torch.linspace(-10, 10, steps=grid_points).repeat(num_inner_units, 1), requires_grad=False) # Initialize spline coefficients self.inner_coef = nn.Parameter(torch.randn(input_dim, num_inner_units, grid_points - spline_degree - 1) * 0.01) self.outer_coef = nn.Parameter(torch.randn(num_inner_units, num_outer_units, grid_points - spline_degree - 1) * 0.01) def forward(self, x): # Inner layer: Map inputs to hidden features inner_outputs = coef2curve(x, self.inner_grid, self.inner_coef, self.spline_degree) inner_outputs = inner_outputs.sum(dim=1) # Outer layer: Map hidden features to outputs outer_outputs = coef2curve(inner_outputs, self.outer_grid, self.outer_coef, self.spline_degree) final_output = outer_outputs.sum(dim=1).mean(dim=1, keepdim=True) return final_output ###################Standard KAN######################## # Data generation torch.manual_seed(42) num_samples = 1200 input_dim = 5 # Generate complex non-linear data X = torch.rand((num_samples, input_dim)) * 10 - 5 noise = torch.randn(num_samples) * 0.5 Y = ((torch.sin(X[:, 0]) + torch.cos(X[:, 1]) + 0.3 * X[:, 2] - 0.2 * X[:, 3] ** 2 + noise) > 0).float().unsqueeze(1) # Split data X_train, X_val, Y_train, Y_val = train_test_split(X.numpy(), Y.numpy(), test_size=0.2, random_state=42) # Train Improved Standard KAN device = 'cuda' if torch.cuda.is_available() else 'cpu' std_model = StandardKAN(input_dim, num_inner_units=8, num_outer_units=4, grid_points=25, spline_degree=3).to(device) criterion = nn.BCELoss() optimizer = optim.Adam(std_model.parameters(), lr=0.001) scheduler = optim.lr_scheduler.StepLR(optimizer, step_size=20, gamma=0.5) epochs = 60 for epoch in range(epochs): std_model.train() optimizer.zero_grad() outputs = torch.sigmoid(std_model(torch.tensor(X_train, dtype=torch.float32).to(device))) loss = criterion(outputs, torch.tensor(Y_train, dtype=torch.float32).to(device)) loss.backward() optimizer.step() scheduler.step() if (epoch + 1) % 10 == 0: print(f"Improved Standard KAN - Epoch [{epoch+1}/{epochs}], Loss: {loss.item():.4f}") # Evaluate Improved Standard KAN std_model.eval() with torch.no_grad(): std_predictions = torch.sigmoid(std_model(torch.tensor(X_val, dtype=torch.float32).to(device))).cpu().numpy() std_auc = roc_auc_score(Y_val, std_predictions) std_fpr, std_tpr, _ = roc_curve(Y_val, std_predictions) std_ks = max(std_tpr - std_fpr) std_logloss = log_loss(Y_val, std_predictions) print("\nImproved Standard KAN Metrics:") print(f"AUC: {std_auc:.4f}") print(f"KS Statistic: {std_ks:.4f}") print(f"Log Loss: {std_logloss:.4f}") ==================================================== ### Code for Example 2: Attention-Based KAN with three normalization functions: Softmax, Softplus, and Rectify ==================================================== import torch import torch.nn as nn import torch.optim as optim from sklearn.metrics import roc_auc_score, log_loss, roc_curve from sklearn.model_selection import train_test_split def custom_softplus(scores): return torch.log(1 + torch.exp(scores)) / torch.sum(torch.log(1 + torch.exp(scores)), dim=-1, keepdim=True) def custom_softmax(scores): exps = torch.exp(scores - torch.max(scores, dim=-1, keepdim=True)[0]) return exps / exps.sum(dim=-1, keepdim=True) def custom_rectify(scores): relu_scores = torch.relu(scores) return relu_scores / torch.sum(relu_scores, dim=-1, keepdim=True) # Attention Mechanism for KAN with multiple functions class AttentionWeightGenerator(nn.Module): def __init__(self, input_dim, norm_function): super(AttentionWeightGenerator, self).__init__() self.query = nn.Linear(input_dim, input_dim, bias=False) self.key = nn.Linear(input_dim, input_dim, bias=False) self.value = nn.Linear(input_dim, input_dim, bias=False) self.norm_function = norm_function def forward(self, x): Q = self.query(x) # Query K = self.key(x) # Key V = self.value(x) # Value scores = torch.matmul(Q, K.transpose(-1, -2)) / (x.size(-1) ** 0.5) # Scaled dot-product attention_weights = self.norm_function(scores) # Custom normalization output = torch.matmul(attention_weights, V) # Weighted sum of values return output # Kolmogorov-Arnold Network (KAN) with Attention Mechanism class AttentionKAN(nn.Module): def __init__(self, input_dim, num_inner_units, num_outer_units, norm_function): super(AttentionKAN, self).__init__() self.num_inner_units = num_inner_units self.num_outer_units = num_outer_units self.inner_attention_layers = nn.ModuleList([ AttentionWeightGenerator(input_dim, norm_function) for _ in range(num_inner_units) ]) self.outer_attention_layers = nn.ModuleList([ AttentionWeightGenerator(num_inner_units, norm_function) for _ in range(num_outer_units) ]) self.output_layer = nn.Linear(num_outer_units, 1) # Final output layer def forward(self, x): inner_outputs = [] for attention_layer in self.inner_attention_layers: inner_output = attention_layer(x).mean(dim=-1, keepdim=True) # Aggregate inner output inner_outputs.append(inner_output) aggregated_inner_output = torch.cat(inner_outputs, dim=-1) # Combine inner outputs outer_outputs = [] for attention_layer in self.outer_attention_layers: outer_output = attention_layer(aggregated_inner_output).mean(dim=-1, keepdim=True) outer_outputs.append(outer_output) aggregated_outer_output = torch.cat(outer_outputs, dim=-1) final_output = self.output_layer(aggregated_outer_output) # Final prediction return final_output # Data generation torch.manual_seed(42) num_samples = 1200 input_dim = 5 # Generate complex non-linear data X = torch.rand((num_samples, input_dim)) * 10 - 5 noise = torch.randn(num_samples) * 0.5 Y = ((torch.sin(X[:, 0]) + torch.cos(X[:, 1]) + 0.3 * X[:, 2] - 0.2 * X[:, 3] ** 2 + noise) > 0).float().unsqueeze(1) # Split data X_train, X_val, Y_train, Y_val = train_test_split(X.numpy(), Y.numpy(), test_size=0.2, random_state=42) # Train Attention KAN with different functions norm_functions = { "Softmax": custom_softmax, "Softplus": custom_softplus, "Rectify": custom_rectify, } results = {} for name, norm_function in norm_functions.items(): print(f"\nTraining Attention KAN with {name}") device = 'cuda' if torch.cuda.is_available() else 'cpu' attention_model = AttentionKAN(input_dim, num_inner_units=6, num_outer_units=4, norm_function=norm_function).to(device) criterion = nn.BCELoss() optimizer = optim.Adam(attention_model.parameters(), lr=0.001) epochs = 100 for epoch in range(epochs): attention_model.train() optimizer.zero_grad() outputs = torch.sigmoid(attention_model(torch.tensor(X_train, dtype=torch.float32).to(device))) loss = criterion(outputs, torch.tensor(Y_train, dtype=torch.float32).to(device)) loss.backward() optimizer.step() if (epoch + 1) % 10 == 0: print(f"{name} - Epoch [{epoch+1}/{epochs}], Loss: {loss.item():.4f}") # Evaluate Attention KAN attention_model.eval() with torch.no_grad(): attention_predictions = torch.sigmoid(attention_model(torch.tensor(X_val, dtype=torch.float32).to(device))).cpu().numpy() attention_auc = roc_auc_score(Y_val, attention_predictions) attention_fpr, attention_tpr, _ = roc_curve(Y_val, attention_predictions) attention_ks = max(attention_tpr - attention_fpr) attention_logloss = log_loss(Y_val, attention_predictions) results[name] = { "AUC": attention_auc, "KS Statistic": attention_ks, "Log Loss": attention_logloss } print(f"\n{name} Metrics:") print(f"AUC: {attention_auc:.4f}") print(f"KS Statistic: {attention_ks:.4f}") print(f"Log Loss: {attention_logloss:.4f}") # Print final results print("\nComparison of Normalization Functions:") for name, metrics in results.items(): print(f"{name}: AUC={metrics['AUC']:.4f}, KS={metrics['KS Statistic']:.4f}, Log Loss={metrics['Log Loss']:.4f}") ====================================================
