Paul wang

dnn/dnn3.py

@@ -65,4 +65,3 @@
 
 
 
-

dnn/mft.py

@@ -0,0 +1,115 @@
+import os
+import pandas as pd
+import numpy as np
+import tensorflow as tf
+from tensorflow.keras import Input, Model
+from tensorflow.keras.layers import Dense, Dropout, BatchNormalization, Activation, LayerNormalization, MultiHeadAttention, GlobalAveragePooling1D, Reshape
+from tensorflow.keras.callbacks import ModelCheckpoint, CSVLogger, EarlyStopping, Callback
+from sklearn.preprocessing import StandardScaler
+from sklearn.metrics import precision_score, recall_score, f1_score, accuracy_score, classification_report, confusion_matrix
+
+
+np.random.seed(1337)
+tf.random.set_seed(1337)
+
+
+train_df = pd.read_csv('kdd/binary/Training.csv', header=None)
+test_df = pd.read_csv('kdd/binary/Testing.csv', header=None)
+
+X_train_raw = train_df.iloc[:, 1:42].values
+y_train = train_df.iloc[:, 0].values
+X_test_raw = test_df.iloc[:, 1:42].values
+y_test = test_df.iloc[:, 0].values
+
+scaler = StandardScaler()
+X_train = scaler.fit_transform(X_train_raw)
+X_test = scaler.transform(X_test_raw)
+
+
+os.makedirs("kddresults/transformer_mft", exist_ok=True)
+os.makedirs("dnn/chung_results_transformer", exist_ok=True)
+os.makedirs("dnnres", exist_ok=True)
+
+
+def build_mft_model(input_dim):
+    inputs = Input(shape=(input_dim,))
+    x = Dense(256)(inputs)
+    x = BatchNormalization()(x)
+    x = Activation('relu')(x)
+    x = Dropout(0.2)(x)
+    x = Reshape((1, 256))(x)
+
+    attn = MultiHeadAttention(num_heads=8, key_dim=256)(x, x)
+    x = LayerNormalization()(x + attn)
+
+    ff = Dense(512, activation='relu')(x)
+    mem = Dense(256, activation='tanh')(ff)
+    x = LayerNormalization()(x + mem)
+
+    ff2 = Dense(256, activation='relu')(x)
+    x = LayerNormalization()(x + ff2)
+
+    x = GlobalAveragePooling1D()(x)
+    x = Dropout(0.4)(x)
+    outputs = Dense(1, activation='sigmoid')(x)
+    return Model(inputs, outputs)
+
+
+def focal_loss_fixed(y_true, y_pred):
+    y_pred = tf.keras.backend.clip(y_pred, 1e-7, 1. - 1e-7)
+    pos = tf.cast(tf.equal(y_true, 1), tf.float32)
+    neg = 1.0 - pos
+    loss = - (0.3 * pos * tf.pow(1 - y_pred, 2) * tf.math.log(y_pred) +
+              0.7 * neg * tf.pow(y_pred, 2) * tf.math.log(1 - y_pred))
+    return tf.reduce_mean(loss)
+
+
+class MetricsLogger(Callback):
+    def on_epoch_end(self, epoch, logs=None):
+        if (epoch + 1) % 10 == 0:
+            y_pred = (self.model.predict(X_test) > 0.2).astype(int)
+            acc = accuracy_score(y_test, y_pred)
+            pre = precision_score(y_test, y_pred)
+            rec = recall_score(y_test, y_pred)
+            f1 = f1_score(y_test, y_pred)
+            print(f"\n[Epoch {epoch+1}] Accuracy: {acc:.4f}, Precision: {pre:.4f}, Recall: {rec:.4f}, F1: {f1:.4f}")
+
+
+model = build_mft_model(41)
+model.compile(loss=focal_loss_fixed, optimizer='adam', metrics=['accuracy'])
+
+checkpointer = ModelCheckpoint("kddresults/transformer_mft/checkpoint-{epoch:02d}.keras", save_best_only=True)
+csv_logger = CSVLogger("kddresults/transformer_mft/training_log.csv", separator=',', append=False)
+earlystop = EarlyStopping(monitor='val_loss', patience=10, restore_best_weights=True)
+
+model.fit(X_train, y_train,
+          batch_size=64,
+          epochs=100,
+          validation_split=0.3,
+          callbacks=[checkpointer, csv_logger, MetricsLogger(), earlystop],
+          verbose=1)
+
+
+model.save("dnn/chung_results_transformer/final_model.hdf5")
+
+scores, names = [], []
+for file in os.listdir("kddresults/transformer_mft"):
+    if file.endswith(".keras"):
+        model.load_weights(os.path.join("kddresults/transformer_mft", file))
+        y_pred = (model.predict(X_test) > 0.2).astype(int)
+        scores.append(f1_score(y_test, y_pred))
+        names.append(file)
+
+best_model = names[np.argmax(scores)]
+model.load_weights(os.path.join("kddresults/transformer_mft", best_model))
+y_pred = (model.predict(X_test) > 0.2).astype(int)
+y_proba = model.predict(X_test)
+
+np.savetxt("dnnres/transformer_predicted.txt", y_pred, fmt='%d')
+np.savetxt("dnnres/transformer_probability.txt", y_proba, fmt='%.6f')
+
+print("\n===== Final Evaluation on Best Transformer Checkpoint (Threshold = 0.32) =====")
+print("Confusion Matrix:")
+print(confusion_matrix(y_test, y_pred))
+print("\nClassification Report:")
+print(classification_report(y_test, y_pred, digits=4))