我正在使用 CNN 和 MobileNet 模型構建一個模型，根據影像資料集將手語分類為字母。因此，它是一個多類分類模型。但是，在編譯和擬合模型之后。我的準確率很高（98%）。但是當我想可視化混淆矩陣時，我真的錯過了矩陣。這是否意味著模型過度擬合？以及如何修復它以獲得更好的矩陣？
train_path = 'train'
test_path = 'test'
train_batches = ImageDataGenerator(preprocessing_function=tf.keras.applications.mobilenet.preprocess_input).flow_from_directory(
directory=train_path, target_size=(64,64), batch_size=10)
test_batches = ImageDataGenerator(preprocessing_function=tf.keras.applications.mobilenet.preprocess_input).flow_from_directory(
directory=test_path, target_size=(64,64), batch_size=10)
mobile = tf.keras.applications.mobilenet.MobileNet()
x = mobile.layers[-6].output
output = Dense(units=32, activation='softmax')(x)
model = Model(inputs=mobile.input, outputs=output)
for layer in model.layers[:-23]:
layer.trainable = False
model.compile(optimizer=Adam(learning_rate=0.0001), loss='categorical_crossentropy', metrics=['accuracy'])
class myCallback(tf.keras.callbacks.Callback):
def on_epoch_end(self,epoch,logs={}):
if(logs.get('val_accuracy')>=0.98):
print('\n Reached to good accuracy')
self.model.stop_training=True
callbacks=myCallback()
model.fit(train_batches,
steps_per_epoch=len(train_batches),
validation_data=test_batches,
validation_steps=len(test_batches),
epochs=10,callbacks=[callbacks])
Epoch 1/10
4498/4498 [==============================] - 979s 217ms/step - loss: 1.3062 - accuracy: 0.6530 - val_loss: 0.1528 - val_accuracy: 0.9594
Epoch 2/10
4498/4498 [==============================] - 992s 221ms/step - loss: 0.1777 - accuracy: 0.9491 - val_loss: 0.1164 - val_accuracy: 0.9691
Epoch 3/10
4498/4498 [==============================] - 998s 222ms/step - loss: 0.1117 - accuracy: 0.9654 - val_loss: 0.0925 - val_accuracy: 0.9734
Epoch 4/10
4498/4498 [==============================] - 1000s 222ms/step - loss: 0.0789 - accuracy: 0.9758 - val_loss: 0.0992 - val_accuracy: 0.9750
Epoch 5/10
4498/4498 [==============================] - 1001s 223ms/step - loss: 0.0626 - accuracy: 0.9805 - val_loss: 0.0818 - val_accuracy: 0.9783
Epoch 6/10
4498/4498 [==============================] - 1007s 224ms/step - loss: 0.0521 - accuracy: 0.9834 - val_loss: 0.0944 - val_accuracy: 0.9789
Epoch 7/10
4498/4498 [==============================] - 1004s 223ms/step - loss: 0.0475 - accuracy: 0.9863 - val_loss: 0.0935 - val_accuracy: 0.9795
Epoch 8/10
4498/4498 [==============================] - 1013s 225ms/step - loss: 0.0371 - accuracy: 0.9880 - val_loss: 0.0854 - val_accuracy: 0.9781
Epoch 9/10
4498/4498 [==============================] - 896s 199ms/step - loss: 0.0365 - accuracy: 0.9879 - val_loss: 0.0766 - val_accuracy: 0.9806
Reached to good accuracy
test_labels = test_batches.classes
predictions = model.predict(x=test_batches, steps=len(test_batches),verbose=0)
cm = confusion_matrix(y_true=test_labels, y_pred=predictions.argmax(axis=1))
cm_plot_labels = ['0','1','2','3','4','5','6','7','8','9','10','11','12','13','14','15','16',
'17','18','19','20','21','22','23','24','25','26','27','28','29','30','31'
]
plot_confusion_matrix(cm=cm, classes=cm_plot_labels, title='Confusion Matrix')
結果混淆矩陣
uj5u.com熱心網友回復：
有一些技巧可以幫助解決 orver 擬合問題：
- 添加資料增強，這種方法會在每次輸入時進行旋轉、隨機裁剪等輕微變換，模型將看到更多相同影像的示例，這將有助于模型更好地泛化。
- 添加dropout 層，該層會在訓練程序中將輸入單元隨機設定為 0，這樣模型在過擬合之前會產生更多的 epoch。
- L1 和 L2正則化，此方法將通過將權重的絕對值加到總損失中來懲罰權重的絕對值。（在此處輸入鏈接描述
- 最好使用 更改您的回呼
callback = tf.keras.callbacks.EarlyStopping(monitor='val_accuracy', patience=3)，我認為您的模型在仍有改進空間時就停止了。
轉載請註明出處，本文鏈接：https://www.uj5u.com/ruanti/363852.html
上一篇：使用HuggingfaceTransformers的聊天機器人
下一篇：如何使用提供的需要tf.Tensor的preprocess_input函式預處理tf.data.Dataset？