1 保存序列模型或函數式模型
In [1]:
# 構建一個簡單的模型並訓練
from __future__ import absolute_import, division, print_function import tensorflow as tf tf.keras.backend.clear_session() from tensorflow import keras from tensorflow.keras import layers inputs = keras.Input(shape=(784,), name='digits') x = layers.Dense(64, activation='relu', name='dense_1')(inputs) x = layers.Dense(64, activation='relu', name='dense_2')(x) outputs = layers.Dense(10, activation='softmax', name='predictions')(x) model = keras.Model(inputs=inputs, outputs=outputs, name='3_layer_mlp') model.summary() (x_train, y_train), (x_test, y_test) = keras.datasets.mnist.load_data() x_train = x_train.reshape(60000, 784).astype('float32') / 255 x_test = x_test.reshape(10000, 784).astype('float32') / 255 model.compile(loss='sparse_categorical_crossentropy', optimizer=keras.optimizers.RMSprop()) history = model.fit(x_train, y_train, batch_size=64, epochs=1) predictions = model.predict(x_test)
1.1 保存整個模型
可以對整個模型進行保存,其保持的內容包括:
- 該模型的架構
- 模型的權重(在訓練期間學到的)
- 模型的訓練配置(傳遞給編譯的)
- 優化器及其狀態(這使您可以從中斷的地方重新啟動訓練)
In [2]:
import numpy as np # 模型保存 model.save('the_save_model.h5') # 導入模型 new_model = keras.models.load_model('the_save_model.h5') new_prediction = new_model.predict(x_test) np.testing.assert_allclose(predictions, new_prediction, atol=1e-6) # 預測結果一樣
1.2 導出為SavedModel文件
SavedModel是Tensorflow對象的獨立序列化格式,支持使用Tensorflow Serving server來部署模型,支持其他語言讀取。
In [4]:
# 導出為tf的SavedModel文件
model.save('save_model', save_format='tf') # 從SavedModel文件中導入模型 new_model = keras.models.load_model('save_model') new_prediction = new_model.predict(x_test) np.testing.assert_allclose(predictions, new_prediction, atol=1e-6) # 預測結果一樣
SaveModel創建的文件包含:
- 權重
- 網絡圖
1.3 僅保存網絡結構
僅保持網絡結構,這樣導出的模型並未包含訓練好的參數
In [5]:
# 獲取網絡結構配置
config = model.get_config() reinitialized_model = keras.Model.from_config(config) new_prediction = reinitialized_model.predict(x_test) assert abs(np.sum(predictions-new_prediction)) >0
也可以使用json保存網絡結構
In [6]:
# 將網絡結構導出為json格式
json_config = model.to_json() reinitialized_model = keras.models.model_from_json(json_config) new_prediction = reinitialized_model.predict(x_test) assert abs(np.sum(predictions-new_prediction)) >0
1.4 僅保存網絡權重參數
In [7]:
# 獲取網絡權重
weights = model.get_weights() # 對網絡權重進行賦值 model.set_weights(weights)
可以把結構和參數保存結合起來
In [8]:
config = model.get_config() weights = model.get_weights() new_model = keras.Model.from_config(config) # config只能用keras.Model的這個api new_model.set_weights(weights) new_predictions = new_model.predict(x_test) np.testing.assert_allclose(predictions, new_predictions, atol=1e-6)
1.5 完整的模型保存方法
In [10]:
# 導出網絡結構和權重
json_config = model.to_json() with open('model_config.json', 'w') as json_file: json_file.write(json_config) model.save_weights('path_to_my_weights.h5') # 載入網絡結構和權重 with open('model_config.json') as json_file: json_config = json_file.read() new_model = keras.models.model_from_json(json_config) new_model.load_weights('path_to_my_weights.h5') new_predictions = new_model.predict(x_test) np.testing.assert_allclose(predictions, new_predictions, atol=1e-6)
In [11]:
# 當然也可以一步到位
model.save('path_to_my_model.h5') del model model = keras.models.load_model('path_to_my_model.h5')
1.6 權重保存格式
有.h5或.keras后綴時保存為keras HDF5格式文件,否則默認為TensorFlow Checkpoint格式文件。可以使用save_format顯式確定。
In [12]:
model.save_weights('weight_tf_savedmodel') model.save_weights('weight_tf_savedmodel.h5')
In [13]:
model.save_weights('weight_tf_savedmodel_tf', save_format='tf') model.save_weights('weight_tf_savedmodel_h5', save_format='h5')
1.7 子類模型權重保存
子類模型的結構無法保存和序列化,只能保持參數
In [16]:
# 構建模型
class ThreeLayerMLP(keras.Model): def __init__(self, name=None): super(ThreeLayerMLP, self).__init__(name=name) self.dense_1 = layers.Dense(64, activation='relu', name='dense_1') self.dense_2 = layers.Dense(64, activation='relu', name='dense_2') self.pred_layer = layers.Dense(10, activation='softmax', name='predictions') def call(self, inputs): x = self.dense_1(inputs) x = self.dense_2(x) return self.pred_layer(x) def get_model(): return ThreeLayerMLP(name='3_layer_mlp') model = get_model()
首先,無法保存從未使用過的子類模型。
這是因為需要在某些數據上調用子類模型才能創建其權重。
In [17]:
# 訓練模型
(x_train, y_train), (x_test, y_test) = keras.datasets.mnist.load_data() x_train = x_train.reshape(60000, 784).astype('float32') / 255 x_test = x_test.reshape(10000, 784).astype('float32') / 255 model.compile(loss='sparse_categorical_crossentropy', optimizer=keras.optimizers.RMSprop()) history = model.fit(x_train, y_train, batch_size=64, epochs=1)
推薦的保存子類模型的方法是使用save_weights創建TensorFlow SavedModel檢查點。
該檢查點將包含與模型關聯的所有變量的值:
- 圖層的權重
- 優化器的狀態
- 與有狀態模型指標關聯的任何變量
In [18]:
# 保持權重參數
model.save_weights('my_model_weights', save_format='tf') # 輸出結果,供后面對比 predictions = model.predict(x_test) first_batch_loss = model.train_on_batch(x_train[:64], y_train[:64])
要還原模型,將需要訪問創建模型對象的代碼。
請注意,為了恢復優化器狀態和任何有狀態度量的狀態,應該先編譯模型(使用與以前完全相同的參數)。
In [19]:
# 讀取保存的模型參數
new_model = get_model() new_model.compile(loss='sparse_categorical_crossentropy', optimizer=keras.optimizers.RMSprop()) #new_model.train_on_batch(x_train[:1], y_train[:1]) new_model.load_weights('my_model_weights') new_predictions = new_model.predict(x_test) np.testing.assert_allclose(predictions, new_predictions, atol=1e-6) new_first_batch_loss = new_model.train_on_batch(x_train[:64], y_train[:64]) assert first_batch_loss == new_first_batch_loss
In [ ]: