人工智能深度学习:TensorFlow2.0如何保持和读取模型?
Posted peijz
tags:
篇首语:本文由小常识网(cha138.com)小编为大家整理,主要介绍了人工智能深度学习:TensorFlow2.0如何保持和读取模型?相关的知识,希望对你有一定的参考价值。
导入数据
(train_images, train_labels), (test_images, test_labels) = tf.keras.datasets.mnist.load_data()
train_labels = train_labels[:1000]
test_labels = test_labels[:1000]
train_images = train_images[:1000].reshape(-1, 28 * 28) / 255.0
test_images = test_images[:1000].reshape(-1, 28 * 28) / 255.0
1.定义一个模型
def create_model():
model = keras.Sequential([
keras.layers.Dense(128, activation=‘relu‘, input_shape=(784,)),
keras.layers.Dropout(0.5),
keras.layers.Dense(10, activation=‘softmax‘)
])
model.compile(optimizer=‘adam‘,
loss=keras.losses.sparse_categorical_crossentropy,
metrics=[‘accuracy‘])
return model
model = create_model()
model.summary()
Model: "sequential_2"
_________________________________________________________________
Layer (type) Output Shape Param #
=================================================================
dense_4 (Dense) (None, 128) 100480
_________________________________________________________________
dropout_2 (Dropout) (None, 128) 0
_________________________________________________________________
dense_5 (Dense) (None, 10) 1290
=================================================================
Total params: 101,770
Trainable params: 101,770
Non-trainable params: 0
_________________________________________________________________
2.checkpoint回调
check_path = ‘106save/model.ckpt‘
check_dir = os.path.dirname(check_path)
cp_callback = tf.keras.callbacks.ModelCheckpoint(check_path,
save_weights_only=True, verbose=1)
model = create_model()
model.fit(train_images, train_labels, epochs=10,
validation_data=(test_images, test_labels),
callbacks=[cp_callback])
Train on 1000 samples, validate on 1000 samples
Epoch 1/10
544/1000 [===============>..............] - ETA: 0s - loss: 2.0658 - accuracy: 0.2831
...
Epoch 00010: saving model to 106save/model.ckpt
1000/1000 [==============================] - 0s 128us/sample - loss: 0.2701 - accuracy: 0.9170 - val_loss: 0.4465 - val_accuracy: 0.8620
<tensorflow.python.keras.callbacks.History at 0x7fbcd872fbe0>
!ls {check_dir}
checkpoint model.ckpt.data-00000-of-00001 model.ckpt.index
model = create_model()
loss, acc = model.evaluate(test_images, test_labels)
print("Untrained model, accuracy: {:5.2f}%".format(100*acc))
1000/1000 [==============================] - 0s 69us/sample - loss: 2.4036 - accuracy: 0.0890
Untrained model, accuracy: 8.90%
model.load_weights(check_path)
loss, acc = model.evaluate(test_images, test_labels)
print("Untrained model, accuracy: {:5.2f}%".format(100*acc))
1000/1000 [==============================] - 0s 47us/sample - loss: 0.4465 - accuracy: 0.8620
Untrained model, accuracy: 86.20%
3.设置checkpoint回调
check_path = ‘106save02/cp-{epoch:04d}.ckpt‘
check_dir = os.path.dirname(check_path)
cp_callback = tf.keras.callbacks.ModelCheckpoint(check_path,save_weights_only=True,
verbose=1, period=5) # 每5
model = create_model()
model.fit(train_images, train_labels, epochs=10,
validation_data=(test_images, test_labels),
callbacks=[cp_callback])
Train on 1000 samples, validate on 1000 samples
Epoch 1/10
1000/1000 [==============================] - 1s 1ms/sample - loss: 1.7242 - accuracy: 0.4490 - val_loss: 1.2205 - val_accuracy: 0.6890
....
Epoch 00010: saving model to 106save02/cp-0010.ckpt
1000/1000 [==============================] - 0s 120us/sample - loss: 0.2845 - accuracy: 0.9220 - val_loss: 0.4402 - val_accuracy: 0.8580
<tensorflow.python.keras.callbacks.History at 0x7fbc5c911b38>
!ls {check_dir}
checkpoint cp-0010.ckpt.data-00000-of-00001
cp-0005.ckpt.data-00000-of-00001 cp-0010.ckpt.index
cp-0005.ckpt.index
载入最新版模型
latest = tf.train.latest_checkpoint(check_dir)
print(latest)
106save02/cp-0010.ckpt
model = create_model()
model.load_weights(latest)
loss, acc = model.evaluate(test_images, test_labels)
print(‘restored model accuracy: {:5.2f}%‘.format(acc*100))
1000/1000 [==============================] - 0s 78us/sample - loss: 0.4402 - accuracy: 0.8580
restored model accuracy: 85.80%
5.手动保持权重
model.save_weights(‘106save03/manually_model.ckpt‘)
model = create_model()
model.load_weights(‘106save03/manually_model.ckpt‘)
loss, acc = model.evaluate(test_images, test_labels)
print(‘restored model accuracy: {:5.2f}%‘.format(acc*100))
1000/1000 [==============================] - 0s 69us/sample - loss: 0.4402 - accuracy: 0.8580
restored model accuracy: 85.80%
6.保持整个模型
model = create_model()
model.fit(train_images, train_labels, epochs=10,
validation_data=(test_images, test_labels),
)
model.save(‘106save03.h5‘)
Train on 1000 samples, validate on 1000 samples
Epoch 1/10
1000/1000 [==============================] - 0s 240us/sample - loss: 1.7636 - accuracy: 0.4460 - val_loss: 1.2041 - val_accuracy: 0.7230
...
Epoch 10/10
1000/1000 [==============================] - 0s 90us/sample - loss: 0.2574 - accuracy: 0.9290 - val_loss: 0.4674 - val_accuracy: 0.8540
new_model = keras.models.load_model(‘106save03.h5‘)
new_model.summary()
Model: "sequential_11"
_________________________________________________________________
Layer (type) Output Shape Param #
=================================================================
dense_22 (Dense) (None, 128) 100480
_________________________________________________________________
dropout_11 (Dropout) (None, 128) 0
_________________________________________________________________
dense_23 (Dense) (None, 10) 1290
=================================================================
Total params: 101,770
Trainable params: 101,770
Non-trainable params: 0
_________________________________________________________________
loss, acc = model.evaluate(test_images, test_labels)
print(‘restored model accuracy: {:5.2f}%‘.format(acc*100))
1000/1000 [==============================] - 1s 810us/sample - loss: 0.4674 - accuracy: 0.8540
restored model accuracy: 85.40%
7.其他导出模型的方法
import time
saved_model_path = "./saved_models/{}".format(int(time.time()))
tf.keras.experimental.export_saved_model(model, saved_model_path)
saved_model_path
‘./saved_models/1553601639‘
new_model = tf.keras.experimental.load_from_saved_model(saved_model_path)
new_model.summary()
Model: "sequential_11"
_________________________________________________________________
Layer (type) Output Shape Param #
=================================================================
dense_22 (Dense) (None, 128) 100480
_________________________________________________________________
dropout_11 (Dropout) (None, 128) 0
_________________________________________________________________
dense_23 (Dense) (None, 10) 1290
=================================================================
Total params: 101,770
Trainable params: 101,770
Non-trainable params: 0
_________________________________________________________________
# 该方法必须先运行compile函数
new_model.compile(optimizer=model.optimizer, # keep the optimizer that was loaded
loss=‘sparse_categorical_crossentropy‘,
metrics=[‘accuracy‘])
# Evaluate the restored model.
loss, acc = new_model.evaluate(test_images, test_labels)
print("Restored model, accuracy: {:5.2f}%".format(100*acc))
1000/1000 [==============================] - 0s 131us/sample - loss: 0.4674 - accuracy: 0.8540
Restored model, accuracy: 85.40%
以上是关于人工智能深度学习:TensorFlow2.0如何保持和读取模型?的主要内容,如果未能解决你的问题,请参考以下文章
人工智能深度学习:TensorFlow2.0如何保持和读取模型?
神经网络与深度学习-基于Tensorflow2.0的手势识别(数据集和训练集)
深度学习TensorFlow2.0:如何用keras构建自己的网络层?
TensorFlow2.0教程-Keras 快速入门:用于构建和训练深度学习模型的高阶 API