1

我已经训练了一种自定义 CNN 模型，如下所示：

class MyModel(tf.keras.Model):
    def __init__(self):
        self.layers = [Conv2D(filters=4,kernel_size=(12,12),input_shape=28,28,1),padding='same',activation='relu'),\
            BatchNormalization(axis=-1 if tf.keras.backend.image_data_format() == 'channels_last' else 1),\
            Conv2D(filters=16,kernel_size=(2,2),padding='same',activation='sigmoid'),\
            BatchNormalization(axis=-1 if tf.keras.backend.image_data_format() == 'channels_last' else 1),\
            Conv2D(filters=1,kernel_size=(3,3),padding='same',activation='relu'),\
            Conv2D(28, (9, 9), input_shape=(28,28,2), activation='sigmoid'), \
            BatchNormalization(axis=-1 if tf.keras.backend.image_data_format() == 'channels_last' else 1), Conv2D(28, (3, 3), activation='sigmoid'), \
            Conv2D(14, (3, 3), activation='relu'), BatchNormalization(axis=-1 if tf.keras.backend.image_data_format() == 'channels_last' else 1), \
            MaxPooling2D((2, 2)), Conv2D(56, (3, 3), activation='relu'), BatchNormalization(axis=-1 if tf.keras.backend.image_data_format() == 'channels_last' else 1), \
            Conv2D(56, (3, 3), activation='relu'), MaxPooling2D((2, 2)), Flatten(),\
            Dense(200, activation="relu"), Dense(100, activation="relu"), Dense(10, activation='softmax', name='class_output')]

    def call(self, inputs, training=False):
        x = self.layers[0](inputs, training=training)
        for lyr in range(1,len(self.layers)):
            x = lyr(x, training=training)

有时（并非总是如此！）将training 变量设置为False.

如果我用类似的东西测试训练有素的模型：

def test_model(model, test_dataset, training=False):
    accuracy = tf.keras.metrics.Accuracy(name='Test Accuracy')
    for (x, y) in test_dataset:
        test_preds = model(x, training=training) 
        accuracy.update_state(y, tf.argmax(test_preds, 1))
    print("{}: {}\n".format(accuracy.name, accuracy.result()*100))

例如，训练 = 假时我得到 0.018%，但训练 =真时得到 98.9%。知道什么可能导致这种行为吗？它实际上似乎并不固定于批量规范化本身，因为我第一次经历了使用 Dropout 层而不是批量规范化层的类似行为。

4

0 回答 0