this

main.py

@@ -1,4 +1,5 @@
 # import pandas as pd
+import os
 from matplotlib import pyplot as plt
 import numpy as np
 from sklearn.model_selection import train_test_split  # type: ignore[import]
@@ -35,8 +36,11 @@ from visualize import (
 )
 
 plt.ioff()
+data_file = "hanoi_weather_data.csv"
+model_file = "theta.npy"
+
 # Step 1: Load data
-data = load_weather_data("hanoi_weather_data.csv")
+data = load_weather_data(data_file)
 
 plot_value_distribution(data)
 plot_time_series(data, feature_names=["QV2M"], title="QV2M Time Series")
@@ -69,35 +73,48 @@ date_train, date_test = train_test_split(date_index, test_size=test_size, shuffl
 # linear_lr = lambda epoch: linear_decay_schedule(
 #     epoch, initial_lr=0.01, min_lr=0.001, total_epochs=1000
 # )
-# cycle_lr = lambda epoch: cyclical_lr_schedule(
+cycle_lr = lambda epoch: cyclical_lr_schedule(
-#     epoch, base_lr=0.001, max_lr=0.01, step_size=200
+    epoch, base_lr=0.001, max_lr=0.01, step_size=200
-# )
+)
 # poly_lr = lambda epoch: polynomial_decay_schedule(
 #     epoch, initial_lr=0.01, min_lr=0.001, total_epochs=1000, power=2
 # )
 # warmup_lr = lambda epoch: warmup_schedule(
 #     epoch, warmup_epochs=10, initial_lr=0.001, max_lr=0.01
 # )
+trained = None
+if os.path.isfile(model_file):
+    if input("found model, train anyway? ").strip().lower() in "y":
+        print("found model but training")
+        trained = False
+    else:
+        print("found model")
+        trained = True
+else:
+    print("model not found, training")
 
 # Step 4: Train model with gradient descent
+if not trained:
     theta, losses = train_linear_regression_gd(
         X_train,
         y_train,
-    lr_schedule=lambda epoch: cyclical_lr_schedule(
+        lr_schedule=cycle_lr,
-    epoch, base_lr=0.0005, max_lr=0.01, step_size=150
-),
         n_epochs=1000,
         batch_size=32,
-    patience=80,
+        patience=50,
-    l2_reg=0.001,  # Enable L2 regularization
-    l1_reg=0.0,  # Disable L1 regularization
     )
 
+    # Save model
+    np.save(model_file, theta)
+else:
+    # Load model
+    theta = np.load(model_file)
+
 predictions = predict_linear_regression(X_test, theta)
 
 # Step 5: Evaluate
-mse, r2 = evaluate_model(y_test, predictions)
+mse, r2, mae = evaluate_model(y_test, predictions)
-print(f"MSE: {mse:.4f}, R2: {r2:.4f}")
+print(f"MSE: {mse:.4f}, R2: {r2:.4f}, MAE: {mae:.4f}")
 
 # Step 6: Rescale predictions
 y_test_rescaled = inverse_scale(y_test, scaler, data.shape[1])
@@ -115,6 +132,7 @@ plot_error_hist(errors)
 plot_scatter(y_test.reshape(-1, 1), predictions.reshape(-1, 1))
 plot_residual_scatter(y_test, predictions)
 # == inspect_residuals(predictions, residuals)
+if losses:
     plot_loss_curve_with_annotations(losses)
 plot_feature_importance(theta, feature_names=[f"QV2M_t-{i}" for i in range(seq_length)])
 

model_train.py

@@ -3,7 +3,7 @@ import time
 
 import numpy as np
 from matplotlib import pyplot as plt
-from sklearn.metrics import (mean_squared_error,  # type: ignore[import]
+from sklearn.metrics import (mean_absolute_error, mean_squared_error,  # type: ignore[import]
                              r2_score)
 
 from preprocessing import inverse_scale
@@ -83,7 +83,7 @@ def train_linear_regression_gd(
 
         # Log statistics
         epoch_time = time.time() - start_time
-        if epoch % 100 == 0 or epoch == n_epochs - 1:
+        if epoch % 50 == 0 or epoch == n_epochs - 1:
             print(
                 f"Epoch {epoch}: Loss = {epoch_loss:.6f}, LR: {lr:.6f}, "
                 f"Gradient Norm = {gradient_norm:.4f}, Weight Norm = {weight_norm:.4f}, "
@@ -104,7 +104,8 @@ def evaluate_model(y_true, y_pred):
     """Evaluate using sklearn's built-in MSE and R² functions."""
     mse = mean_squared_error(y_true, y_pred)
     r2 = r2_score(y_true, y_pred)
-    return mse, r2
+    mae = mean_absolute_error(y_true, y_pred)
+    return iter((mse, r2, mae))
 
 
 import math