model_analysis.py

import matplotlib.pyplot as plt 
import numpy as np 
from sklearn.metrics import auc


def loss(model):
    plt.plot(model.train_loss, label='Training Loss')
    plt.plot(model.val_loss, label='Validation Loss')
    plt.legend()
    plt.xlabel('Epoch')
    plt.ylabel('Loss')
    plt.title('Training and Validation Losses')
    plt.savefig('Training and Validation Losses')
    plt.clf()

    
def plot_anomaly_score(model):
    plt.figure()
    bins = 50  
    range_ = (min(np.min(model.anomaly_scores), np.min(model.test_scores)),
            max(np.max(model.anomaly_scores), np.max(model.test_scores)))

    weights_anomaly = np.ones_like(model.anomaly_scores.flatten()) / len(model.anomaly_scores.flatten())
    weights_test = np.ones_like(model.test_scores.flatten()) / len(model.test_scores.flatten())

    plt.hist(model.anomaly_scores.flatten(), bins=bins, range=range_, weights=weights_anomaly,
            color='red', alpha=0.5, label='anomalous', density=False)
    plt.hist(model.test_scores.flatten(), bins=bins, range=range_, weights=weights_test,
            color='blue', alpha=0.5, label='not anomalous', density=False)

    plt.xlabel('Anomaly Score')
    plt.title('Anomaly Score Distribution')
    plt.legend()
    plt.savefig('Anomaly Score Distribution')
    plt.clf()


def roc(model, test_data, anomaly_data):
        all_data = np.concatenate((test_data, anomaly_data), axis=0)
        data_pred = model.predict(all_data)
        data_loss = np.mean(np.square(all_data - data_pred), axis=(1,2,3))
        thresholds = np.linspace(np.min(data_loss[len(test_data):]), np.max(data_loss[:len(test_data)]), num=500)
        tprs = []
        fprs = []
        for threshold in thresholds:
            pred_signal = (data_loss > threshold)
            true_signal = np.ones_like(pred_signal)
            true_signal[:len(test_data)] = 0  # The first len(light_data) events are background, the rest are signal
            tp = np.sum(np.logical_and(pred_signal, true_signal))
            fp = np.sum(np.logical_and(pred_signal, 1 - true_signal))
            tn = np.sum(np.logical_and(1 - pred_signal, 1 - true_signal))
            fn = np.sum(np.logical_and(1 - pred_signal, true_signal))
            tpr = tp / (tp + fn)
            fpr = fp / (fp + tn)
            tprs.append(tpr)
            fprs.append(fpr)

        plt.figure()
        plt.plot(fprs, tprs, label='Autoencoder')
        plt.plot([0, 1], [0, 1], 'k--', label='Random guess')
        plt.xlabel('False positive rate')
        plt.ylabel('True positive rate')
        plt.title('Receiver operating characteristic (ROC) curve')
        plt.legend()
        plt.savefig("ROC")
        plt.clf()

        auc_score = auc(fprs, tprs)
        print('AUC: {:.3f}'.format(auc_score))