Fatigue_Life/surrogateAI/utilities/rollout_vis.py

import pickle
import numpy as np
import pandas as pd
import os
import re
from sklearn.metrics import mean_squared_error, mean_absolute_error, r2_score

# --- 1. Define the base directory for all evaluation results ---
base_results_dir = "Evaluation_results"

# --- 2. Manually set the k-value and epoch number for this run ---
k_value = 60
epoch_num = 500
# --- You can change the two values above for each run ---


# Load the trajectory operations dictionary from the .pkl file
pkl_path = '/home/gd_user1/AnK/project_PINN/Project_Fatigue/Fatigue_Life/output_shaft_extra/regDGCNN_seg/shaft_low_extra2_/EXPERIMENT_multi_feat_exponent_k60_500_shaft_low_extra2/Thu-Oct--9-10-04-58-2025/rollout/rollout_epoch_500.pkl'

# Create a unique, descriptive folder name based on the manual inputs
run_name = f"k{k_value}_epoch{epoch_num}_exponent_multi_metrics_2"
run_dir = os.path.join(base_results_dir, run_name)
os.makedirs(run_dir, exist_ok=True)

print(f"Results for this run will be saved in: {run_dir}")


try:
    with open(pkl_path, 'rb') as f:
        traj_ops = pickle.load(f)
except FileNotFoundError:
    print(f"Error: The file was not found at {pkl_path}")
    exit()  # Exit the script if the file doesn't exist

total_trajectories = len(traj_ops)
print(f"Loaded {total_trajectories} trajectory operations from the file.")
all_metrics=[]
for i in range(total_trajectories):
    # Extract tensors (log10 values)
    gt_life_log = traj_ops[i]['gt_fatigue_life'].cpu().numpy().flatten()
    pred_life_log = traj_ops[i]['pred_fatigue_life'].cpu().numpy().flatten()
    
    cells= traj_ops[i]['cells'].cpu().numpy()
    mesh_pos= traj_ops[i]['mesh_pos'].cpu().numpy()

    # Save the geometry to a separate, efficient .npz file
    geometry_path = os.path.join(run_dir, f"geometry_sample_{i+1}.npz")
    np.savez_compressed(geometry_path, mesh_pos=mesh_pos, cells=cells)
    print(f"  -> Saved geometry to {geometry_path}")

    # Convert back to non-log values
    gt_life_real = np.power(10.0, gt_life_log)
    pred_life_real = np.power(10.0, pred_life_log)

    # --- CALCULATE ERRORS ---
    abs_error = np.abs(gt_life_real - pred_life_real)
    percentage_error = (abs_error / (gt_life_real + 1e-9)) * 100

    # Create a DataFrame for easy analysis and saving
    df = pd.DataFrame({
        'node': np.arange(len(gt_life_log)),

        # Log values
        'gt_fatigue_life_log10': gt_life_log,
        'pred_fatigue_life_log10': pred_life_log,

        # Real (non-log) values
        'gt_fatigue_life_real': gt_life_real,
        'pred_fatigue_life_real': pred_life_real,

        # Errors in real values
        'absolute_error': abs_error,
        'percentage_error': percentage_error
    })
    
    # Metrics on real (non-log) fatigue life
    mse = mean_squared_error(gt_life_real, pred_life_real)
    rmse = np.sqrt(mse)
    mae = mean_absolute_error(gt_life_real, pred_life_real)
    mape = np.mean(percentage_error)  # Mean of percentage error
    r2 = r2_score(gt_life_real, pred_life_real)

    metrics_dict = {
        "test_sample": i+1,
        "MSE": mse,
        "RMSE": rmse,
        "MAE": mae,
        "MAPE": mape,
        "R2": r2
    }
    all_metrics.append(metrics_dict)

    # Save each trajectory to a CSV
    csv_filename = f"prediction_sample{i+1}.csv"
    csv_path = os.path.join(run_dir, csv_filename)

    df.to_csv(csv_path, index=False)

    print(f"Saved fatigue life evaluation to {csv_path}")
    
metrics_df = pd.DataFrame(all_metrics)
metrics_csv_path = os.path.join(run_dir, "evaluation_summary.csv")
metrics_df.to_csv(metrics_csv_path, index=False)
print(f"Saved summary metrics to {metrics_csv_path}")