python - tf.session.run 的网络输出与使用 keras.Model.predict 获得的网络输出有很大不同

Question

我正在尝试通过 Tensorflow session 使用 Keras 模型。但是结果形式 model.predict 和 sess.run 不同。有什么方法可以通过 Tensorflow session 使用 Kers 模型吗？

Tensorflow version: 1.4.0
Keras version: 2.1.1

from sklearn.datasets.samples_generator import make_circles
from keras.models import Sequential
from keras.layers import Dense
from keras.optimizers import SGD
import numpy as np
import tensorflow as tf
from keras import backend as K

sess = tf.Session()
K.tensorflow_backend.set_session(sess)

X, y = make_circles(n_samples=1000,
                    noise=0.1,
                    factor=0.2,
                    random_state=0)

model = Sequential()
model.add(Dense(4, input_shape=(2,), activation='tanh'))
model.add(Dense(1, activation='sigmoid'))
model.compile(SGD(lr=0.5), 'binary_crossentropy', metrics=['accuracy'])
model.fit(X, y, epochs=20)

print("Keras model. First prediction: " + str(model.predict(np.c_[0, 0])))
print("Keras model. Second prediction: " + str(model.predict(np.c_[1.5, 1.5])))

with sess.as_default():

    y_tensor = model.outputs[0]
    x_tensor = model.inputs[0]
    sess.run(tf.global_variables_initializer())

    print("TF model. First prediction: " + str(sess.run(y_tensor, feed_dict={x_tensor: np.c_[0, 0]} )))
    print("TF model. Second prediction: " + str(sess.run(y_tensor, feed_dict={x_tensor: np.c_[1.5, 1.5]} )))

Answer 1

好的，它是 K.set_session(s) 而不是 K.tensorflow_backend.set_session(s)。

第二:sess.run(tf.global_variables_initializer()) 使用各自的初始化器重置所有变量，包括网络权重(它们默认使用 xavier 初始化器)。

所以你是:

1. Training the keras model
2. Printing the prediction for the keras model
3. Reseting to random weights
4. Printing the predictions for the same model

注释 sess.run(tf.global_variables_initializer()) 修复了问题:

Keras model. First prediction: [[ 0.99195099]]
Keras model. Second prediction: [[ 0.03110269]]
TF model. First prediction: [[ 0.99195099]]
TF model. Second prediction: [[ 0.03110269]]