关于keras模型分步训练的一些探索

 将模型分为了两部分来训练，观察相互之间的影响，结果表明，模型1，2在分开训练的情况下，组合后依然有效，最终计算结果相同。

from keras import optimizers
from keras.datasets import mnist
from keras.layers import Dense, Input
from keras.models import Model
from keras.utils import np_utils

(X_train, y_train1), (X_test, y_test1) = mnist.load_data()

X_train = X_train.reshape(X_train.shape[0], -1) / 255.   # normalize
X_test = X_test.reshape(X_test.shape[0], -1) / 255.      # normalize
y_train = np_utils.to_categorical(y_train1, num_classes=10)
y_test = np_utils.to_categorical(y_test1, num_classes=10)

input = Input([784])

dnn1 = Dense(64,activation='relu')(input)
dnn11 = Dense(10,activation='softmax')(dnn1)

in2 = Input([64])
dnn2 = Dense(10,activation='softmax')(in2)

m1 = Model(inputs = input,outputs = dnn1)
m11 = Model(inputs = input,outputs = dnn11)
m2 = Model(inputs = in2,outputs = dnn2)


m2.compile(optimizer=optimizers.RMSprop(lr=0.001, rho=0.9, epsilon=1e-08, decay=0.0),
              loss='categorical_crossentropy',
              metrics=['accuracy'])
b = m1.predict(X_test)
a = m2.evaluate(b,y_test)
print(a)


rmsprop = optimizers.RMSprop(lr=0.001, rho=0.9, epsilon=1e-08, decay=0.0)
m11.compile(optimizer=rmsprop,
              loss='categorical_crossentropy',
              metrics=['accuracy'])
m11.fit(X_train, y_train, epochs=2, batch_size=32,validation_data=(X_test,y_test))

a = m2.evaluate(m1.predict(X_test),y_test)
print(a)

m2.fit(m1.predict(X_train),y_train, epochs=2, batch_size=32,validation_data=(m1.predict(X_test),y_test))

a = m2.evaluate(m1.predict(X_test),y_test)
print(a)

输出：

10000/10000 [==============================] - 0s 20us/step
[2.374014566040039, 0.0989]
Train on 60000 samples, validate on 10000 samples
Epoch 1/2
60000/60000 [==============================] - 4s 68us/step - loss: 0.3074 - acc: 0.9135 - val_loss: 0.1863 - val_acc: 0.9443
Epoch 2/2
60000/60000 [==============================] - 4s 67us/step - loss: 0.1626 - acc: 0.9529 - val_loss: 0.1344 - val_acc: 0.9614
10000/10000 [==============================] - 0s 19us/step
[3.976517904663086, 0.1419]
Train on 60000 samples, validate on 10000 samples
Epoch 1/2
60000/60000 [==============================] - 3s 54us/step - loss: 0.3351 - acc: 0.9049 - val_loss: 0.1436 - val_acc: 0.9580
Epoch 2/2
60000/60000 [==============================] - 3s 51us/step - loss: 0.1275 - acc: 0.9634 - val_loss: 0.1278 - val_acc: 0.9624
10000/10000 [==============================] - 0s 18us/step
[0.12776271086111665, 0.9624]