Keras 是一个用 Python 编写的高级神经网络 API,它能够以 TensorFlow, CNTK, 或者 Theano 作为后端运行。Keras 的开发重点是支持快速的实验。有时候我们在使用keras设计好模型后,需要在其他平台进行运行,这时候我们就需要将keras h5 model转换为TensorFlow pb model,因为keras只是一个Python的高级库,而TensorFlow能够支持多平台的运行。
环境
Python 3.6
Keras 2.2.2
Tensorflow-gpu 1.8.0
Keras to Tensorflow
测试数据:
from keras.datasets import imdb
def get_data():
max_features = 20000
# cut texts after this number of words
# (among top max_features most common words)
maxlen = 100
print('Loading data...')
(x_train, y_train), (x_test, y_test) = imdb.load_data(num_words=max_features)
print(x_train.shape, 'train sequences')
print(x_test.shape, 'test sequences')
print('Pad sequences (samples x time)')
x_train = sequence.pad_sequences(x_train, maxlen=maxlen)
x_test = sequence.pad_sequences(x_test, maxlen=maxlen)
print('x_train shape:', x_train.shape)
print('x_test shape:', x_test.shape)
y_train = np.array(y_train)
y_test = np.array(y_test)
return x_train, x_test, y_train, y_test
生成一个keras模型进行训练,获得模型和对应的权重文件:
from keras.layers import Conv1D, GlobalMaxPooling1D, Embedding, Dense, Dropout
from keras.datasets import imdb
from keras.preprocessing import sequence
from keras.models import Sequential
def gen_keras_model(x_train, x_test, y_train, y_test, train=False):
inp = Input(shape=(100,))
x = Embedding(20000, 50)(inp)
x = Dropout(0.2)(x)
x = Conv1D(250, 3, padding='valid', activation='relu', strides=1)(x)
x = GlobalMaxPooling1D()(x)
x = Dense(250, activation='relu')(x)
x = Dropout(0.2)(x)
x = Dense(1, activation='sigmoid')(x)
model = Model(inputs=inp, outputs=x)
if train:
model.compile(loss='binary_crossentropy',
optimizer='adam',
metrics=['accuracy'])
model.fit(x_train, y_train,
batch_size=32,
epochs=2,
validation_data=(x_test, y_test))
model.save_weights('model.h5')
return model
if __name__ == '__main__':
x_train, x_test, y_train, y_test = get_data()
model = gen_keras_model(x_train, x_test, y_train, y_test, True)
下面的函数将keras model转换为Tensorflow pb文件:
首先构建一个Session与空的计算图,将这个计算图设置为默认的计算图。
获取keras model的输出节点,将这个输出节点与节点名在这个计算图中进行绑定。
使用convert_variables_to_constants函数保存数输出节点,函数会自动推导计算图并将计算图中的变量取值以常量的形式保存。在保存模型文件的时候,我们只是导出了GraphDef部分,GraphDef保存了从输入层到输出层的计算过程。
最后向指定目录写入pb文件。
如果你的graph使用了Keras的learning phase(在训练和测试中行为不同),你首先要做的事就是在graph中硬编码你的工作模式(设为0,即测试模式),该工作通过:1)使用Keras的后端注册一个learning phase常量,2)重新构建模型,来完成。
import tensorflow as tf
from keras import backend as K
from tensorflow.python.framework import graph_util, graph_io
def export_graph(model, export_path):
input_names = model.input_names
if not tf.gfile.Exists(export_path):
tf.gfile.MakeDirs(export_path)
with K.get_session() as sess:
init_graph = sess.graph
with init_graph.as_default():
out_nodes = []
for i in range(len(model.outputs)):
out_nodes.append("output_" + str(i + 1))
tf.identity(model.output[i], "output_" + str(i + 1))
init_graph = sess.graph.as_graph_def()
main_graph = graph_util.convert_variables_to_constants(sess, init_graph, out_nodes)
graph_io.write_graph(main_graph, export_path, name='model.pb', as_text=False)
return input_names, out_nodes
if __name__ == '__main__':
x_train, x_test, y_train, y_test = get_data()
learning_phase = 0
K.set_learning_phase(learning_phase)
model = gen_keras_model(x_train, x_test, y_train, y_test, learning_phase)
model.load_weights('model.h5')
input_names, output_names = export_graph(model, 'model')
在Python Tensorflow环境下进行测试
首先在Session与Graph中读入pb文件,构建计算图。
然后根据输入张量与输出张量的张量名来获取到对应的张量,这里一定要加上:0。比如input_1:0是张量的名称而input_1表示的是节点的名称。
最后使用常规的Tensorflow操作来运行模型。
import numpy as np
import tensorflow as tf
from sklearn.metrics import accuracy_score
def run_graph(pb_file_path, input_name, output_name, x_test, y_test):
tf.reset_default_graph()
sess = tf.Session()
with tf.gfile.FastGFile(pb_file_path, 'rb') as f:
graph_def = tf.GraphDef()
graph_def.ParseFromString(f.read())
sess.graph.as_default()
tf.import_graph_def(graph_def, name='')
#输入
input_x = sess.graph.get_tensor_by_name('{}:0'.format(input_name))
#输出
op = sess.graph.get_tensor_by_name('{}:0'.format(output_name))
#预测结果
pred = []
for x in x_test:
res = sess.run(op, {input_x: x.reshape(1, -1)})
pred.append(res[0])
pred = np.array([1 if p > 0.5 else 0 for p in pred])
acc = accuracy_score(y_test, pred)
print('Accuracy:{}'.format(acc))
if __name__ == '__main__':
x_train, x_test, y_train, y_test = get_data()
learning_phase = 0
K.set_learning_phase(learning_phase)
model = gen_keras_model(x_train, x_test, y_train, y_test, learning_phase)
model.load_weights('model.h5')
input_names, output_names = export_graph(model, 'model')
pred = run_graph('model\model.pb', input_names[0], output_names[0], x_test, y_test)
输出如下:
Using TensorFlow backend.
Loading data...
(25000,) train sequences
(25000,) test sequences
Pad sequences (samples x time)
x_train shape: (25000, 100)
x_test shape: (25000, 100)
INFO:tensorflow:Froze 7 variables.
Converted 7 variables to const ops.
Accuracy:0.84388
在JavaTensorflow环境下进行测试
在 Windows 上安装按照以下步骤在 Windows 上安装适用于 Java 的 TensorFlow:
下载 libtensorflow.jar,这是 TensorFlow Java 归档 (JAR)。
解压缩该 .zip 文件。
配置到IDEA的External Libraries中。
setting
在Java中使用PB文件的代码如下,我们随机生成一个数组作为输入的张量进行测试。整个流程与Python下类似,需要注意的是生成输入张量时数组类型需要定义为float类型,不然会出现以下错误:
Exception in thread "main" java.lang.IllegalArgumentException: Expects arg[0] to be float but double is provided
Java下的测试代码:
import org.tensorflow.Graph;
import org.tensorflow.Session;
import org.tensorflow.Tensor;
import java.io.IOException;
import java.nio.file.Files;
import java.nio.file.Paths;
import java.util.Arrays;
public class TFTest {
public static void main(String[] args) throws IOException {
String path = "E:\\Documents\\Desktop\\code\\glu\\model\\model.pb";
float[][] input = new float[1][100];
for (int i=0; i < 100; i++){
input[0][i] = (float) (Math.random() * 100);
}
try (Graph graph = new Graph()){
graph.importGraphDef(Files.readAllBytes(Paths.get(path)));
try (Session sess = new Session(graph)){
try (Tensor x = Tensor.create(input);
Tensor y = sess.runner().feed("input_1", x).fetch("output_1").run().get(0)){
float[] res = (float[]) y.copyTo(new float[1]);
System.out.println(Arrays.toString(y.shape()));
System.out.println(Arrays.toString(res));
}
}
}
}
}
输出结果如下:
[1]
[0.088513985]
Process finished with exit code 0