Java调用Keras、Tensorflow模型

https://blog.csdn.net/rabbit_judy/article/details/80054085

https://blog.csdn.net/Butertfly/article/details/80952987

https://www.jianshu.com/p/0016a34c82c8

 

1、Keras的h5模型转为pb模型

在Keras中使用model.save(model.h5)保存当前模型为HDF5格式的文件中。
Keras的后端框架使用的是tensorflow，所以先把模型导出为pb模型。在Java中只需要调用模型进行预测，所以将当前的graph中的Variable全部变成Constant，并且使用训练后的weight。以下是freeze graph的代码：

    def freeze_session(session, keep_var_names=None, output_names=None, clear_devices=True):
        """
        :param session: 需要转换的tensorflow的session
        :param keep_var_names:需要保留的variable，默认全部转换constant
        :param output_names:output的名字
        :param clear_devices:是否移除设备指令以获得更好的可移植性
        :return:
        """
        from tensorflow.python.framework.graph_util import convert_variables_to_constants
        graph = session.graph
        with graph.as_default():
            freeze_var_names = list(set(v.op.name for v in tf.global_variables()).difference(keep_var_names or []))
            output_names = output_names or []
            # 如果指定了output名字，则复制一个新的Tensor，并且以指定的名字命名
            if len(output_names) > 0:
                for i in range(output_names):
                    # 当前graph中复制一个新的Tensor，指定名字
                    tf.identity(model.model.outputs[i], name=output_names[i])
            output_names += [v.op.name for v in tf.global_variables()]
            input_graph_def = graph.as_graph_def()
            if clear_devices:
                for node in input_graph_def.node:
                    node.device = ""
            frozen_graph = convert_variables_to_constants(session, input_graph_def,
                                                          output_names, freeze_var_names)
            return frozen_graph

该方法可以将tensor为Variable的graph全部转为constant并且使用训练后的weight。注意output_name比较重要，后面Java调用模型的时候会用到。

在Keras中，模型是这么定义的:

    def create_model(self):
        input_tensor = Input(shape=(self.maxlen,), name="input")
        x = Embedding(len(self.text2id) + 1, 200)(input_tensor)
        x = Bidirectional(LSTM(128))(x)
        x = Dense(256, activation="relu")(x)
        x = Dropout(self.dropout)(x)
        x = Dense(len(self.id2class), activation='softmax', name="output_softmax")(x)
        model = Model(inputs=input_tensor, outputs=x)
        model.compile(loss='categorical_crossentropy',
                      optimizer='adam',
                      metrics=['accuracy'])

下面的代码可以查看定义好的Keras模型的输入、输出的name，这对之后Java调用有帮助。

print(model.input.op.name)
print(model.output.op.name)

训练好Keras模型后，转换为pb模型：

from keras import backend as K
import tensorflow as tf

model.load_model("model.h5")
print(model.input.op.name)
print(model.output.op.name)
# 自定义output_names
frozen_graph = freeze_session(K.get_session(), output_names=["output"])
tf.train.write_graph(frozen_graph, "./", "model.pb", as_text=False)

### 输出：
# input
# output_softmax/Softmax
# 如果不自定义output_name，则生成的pb模型的output_name为output_softmax/Softmax，如果自定义则以自定义名为output_name

运行之后会生成model.pb的模型，这将是之后调用的模型。

2、Java调用

新建一个maven项目，pom里面导入tensorflow包：

<dependency>
            <groupId>org.tensorflow</groupId>
            <artifactId>tensorflow</artifactId>
            <version>1.6.0</version>
</dependency>

核心代码:

public void predict() throws Exception {
        try (Graph graph = new Graph()) {
            graph.importGraphDef(Files.readAllBytes(Paths.get(
                    "path/to/model.pb"
            )));
            try (Session sess = new Session(graph)) {
                // 自己构造一个输入
                float[][] input = {{56, 632, 675, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0}};
                try (Tensor x = Tensor.create(input);
                    // input是输入的name，output是输出的name
                    Tensor y = sess.runner().feed("input", x).fetch("output").run().get(0)) {
                    float[][] result = new float[1][y.shape[1]];
                    y.copyTo(result);
                    System.out.println(Arrays.toString(y.shape()));
                    System.out.println(Arrays.toString(result[0]));
                }
            }
        }
    }

Graph和Tensor对象都是需要通过close()方法显式地释放占用的资源，代码中使用了try-with-resources的方法实现的。

至此，已经可以实现Keras离线训练，Java在线预测的功能。

 

 

from __future__ import print_function
import keras
from keras.datasets import mnist
from keras.models import Sequential
from keras.layers import Dense, Dropout, Flatten
from keras.layers import Conv2D, MaxPooling2D
from keras import backend as K
import tensorflow as tf
import numpy as np
数据预处理
batch_size = 128
num_classes = 10
epochs = 30
img_rows, img_cols = 28, 28

(x_train, y_train), (x_test, y_test) = mnist.load_data()

x_train = x_train.reshape(x_train.shape[0], img_rows, img_cols, 1)
x_test = x_test.reshape(x_test.shape[0], img_rows, img_cols, 1)
input_shape = (img_rows, img_cols, 1)

x_train = x_train.astype('float32')
x_test = x_test.astype('float32')

x_train /= 255
x_test /= 255
print('x_train shape:', x_train.shape)
print(x_train.shape[0], 'train samples')
print(x_test.shape[0], 'test samples')

y_train = keras.utils.to_categorical(y_train, num_classes)
y_test = keras.utils.to_categorical(y_test, num_classes)

keras模型搭建
model = Sequential()
model.add(Conv2D(32, kernel_size=(3, 3),
activation='relu',
input_shape=input_shape))
model.add(Conv2D(64, (3, 3), activation='relu'))
model.add(MaxPooling2D(pool_size=(2, 2)))
model.add(Dropout(0.25))
model.add(Flatten())
model.add(Dense(128, activation='relu'))
model.add(Dropout(0.5))
model.add(Dense(num_classes, activation='softmax'))

model.summary()

训练模型
model.compile(loss=keras.losses.categorical_crossentropy,
optimizer=keras.optimizers.Adadelta(),
metrics=['accuracy'])

model.fit(x_train, y_train,
batch_size=batch_size,
epochs=epochs,
verbose=1,
validation_data=(x_test, y_test))
kera模型保存为pb文件
sess = K.get_session()

frozen_graph_def = tf.graph_util.convert_variables_to_constants(
sess,
sess.graph_def,
output_node_names=["y/Softmax"])

# 保存图为pb文件
# with open('model.pb', 'wb') as f:
# f.write(frozen_graph_def.SerializeToString())

tf.train.write_graph(frozen_graph_def, 'model', 'test_model.pb', as_text=False)

 

模型预测
a = np.array(
[0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0.2, 0.517647, 0.839216, 0.992157, 0.996078, 0.992157, 0.796079, 0.635294, 0.160784,
0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0.4, 0.556863, 0.796079, 0.796079, 0.992157, 0.988235, 0.992157,
0.988235, 0.592157, 0.27451, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0.996078, 0.992157, 0.956863,
0.796079, 0.556863, 0.4, 0.321569, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0.67451,
0.988235, 0.796079, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0.0823529, 0.87451,
0.917647, 0.117647, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0.478431, 0.992157,
0.196078, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0.482353, 0.996078, 0.356863,
0.2, 0.2, 0.2, 0.0392157, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0.0823529, 0.87451, 0.992157,
0.988235, 0.992157, 0.988235, 0.992157, 0.67451, 0.321569, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
0.0823529, 0.839216, 0.992157, 0.796079, 0.635294, 0.4, 0.4, 0.796079, 0.87451, 0.996078, 0.992157, 0.2, 0.0392157,
0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0.239216, 0.992157, 0.670588, 0, 0, 0, 0, 0, 0.0784314, 0.439216,
0.752941, 0.992157, 0.831373, 0.160784, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
0.4, 0.796079, 0.917647, 0.2, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0.0784314,
0.835294, 0.909804, 0.321569, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0.243137,
0.796079, 0.917647, 0.439216, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0.0784314,
0.835294, 0.988235, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0.6, 0.992157, 0,
0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0.160784, 0.913726, 0.831373, 0, 0, 0, 0,
0, 0, 0, 0, 0, 0, 0, 0, 0.443137, 0.360784, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0.121569, 0.678431, 0.956863, 0.156863,
0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0.321569, 0.992157, 0.592157, 0, 0, 0, 0, 0, 0, 0.0823529, 0.4, 0.4, 0.717647,
0.913726, 0.831373, 0.317647, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0.321569, 1.0, 0.992157, 0.917647, 0.596078, 0.6,
0.756863, 0.678431, 0.992157, 0.996078, 0.992157, 0.996078, 0.835294, 0.556863, 0.0784314, 0, 0, 0, 0, 0, 0, 0, 0,
0, 0, 0, 0, 0, 0, 0.278431, 0.592157, 0.592157, 0.909804, 0.992157, 0.831373, 0.752941, 0.592157, 0.513726,
0.196078, 0.196078, 0.0392157, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0.0])

print(model.predict(a.reshape(1, img_rows, img_cols, 1)))
# print(model.input.name)
# print(model.output.name)
python调用pb文件
with tf.Graph().as_default():
output_graph_def = tf.GraphDef()

with open('model/test_model.pb', "rb") as f:
output_graph_def.ParseFromString(f.read())
_ = tf.import_graph_def(output_graph_def, name="")

with tf.Session() as sess:
init = tf.global_variables_initializer()
sess.run(init)

input_x = sess.graph.get_tensor_by_name("x_input:0")
output = sess.graph.get_tensor_by_name("y/Softmax:0")

print(sess.run(output, feed_dict={input_x: a.reshape(1, img_rows, img_cols, 1)})）
java调用pb文件
maven相关依赖

<dependency>
<groupId>commons-io</groupId>
<artifactId>commons-io</artifactId>
<version>2.6</version>
</dependency>
<dependency>
<groupId>org.tensorflow</groupId>
<artifactId>tensorflow</artifactId>
<version>1.7.0</version>
</dependency>
import org.apache.commons.io.IOUtils;
import org.tensorflow.Graph;
import org.tensorflow.Session;
import org.tensorflow.Tensor;

import java.io.FileInputStream;
import java.io.IOException;
import java.nio.FloatBuffer;
import java.util.List;
public class Test {
public static String PB_FILE_PATH = "test_model.pb";
public static String INPUT_TENSOR_NAME = "x_input";
public static String OUTPUT_TENSOR_NAME = "y/Softmax";

public static void main(String[] args) throws IOException {
try (Graph graph = new Graph()) {
//导入图
byte[] graphBytes = IOUtils.toByteArray(new FileInputStream(PB_FILE_PATH));
graph.importGraphDef(graphBytes);

float[] a = new float[]{
0f, 0f, 0f, 0f, 0f, 0f, 0f, 0f, 0f, 0f, 0f, 0f, 0f, 0f, 0f, 0f, 0f, 0f, 0f, 0f, 0f, 0f, 0f, 0f, 0f, 0f, 0f, 0f, 0f, 0f, 0f, 0f, 0f, 0f, 0f, 0f, 0f, 0f, 0f, 0f, 0f, 0f, 0f, 0f, 0f, 0f, 0f, 0f, 0f, 0f, 0f, 0f, 0f, 0f, 0f, 0f, 0f, 0f, 0f, 0f, 0f, 0f, 0f, 0f, 0f, 0f, 0f, 0f, 0f, 0f, 0f, 0f, 0f, 0f, 0f, 0f, 0f, 0f, 0f, 0f, 0f, 0f, 0f, 0f, 0f, 0f, 0f, 0f, 0f, 0f, 0f, 0f, 0f, 0f, 0f, 0f, 0f, 0f, 0f, 0f, 0f, 0f, 0f, 0f, 0f, 0f, 0f, 0f, 0f, 0f, 0f, 0f, 0f, 0f, 0f, 0f, 0f, 0f, 0f, 0f, 0f, 0f, 0f, 0f, 0.2f, 0.517647f, 0.839216f, 0.992157f, 0.996078f, 0.992157f, 0.796079f, 0.635294f, 0.160784f, 0f, 0f, 0f, 0f, 0f, 0f, 0f, 0f, 0f, 0f, 0f, 0f, 0f, 0f, 0f, 0.4f, 0.556863f, 0.796079f, 0.796079f, 0.992157f, 0.988235f, 0.992157f, 0.988235f, 0.592157f, 0.27451f, 0f, 0f, 0f, 0f, 0f, 0f, 0f, 0f, 0f, 0f, 0f, 0f, 0f, 0f, 0f, 0f, 0f, 0f, 0.996078f, 0.992157f, 0.956863f, 0.796079f, 0.556863f, 0.4f, 0.321569f, 0f, 0f, 0f, 0f, 0f, 0f, 0f, 0f, 0f, 0f, 0f, 0f, 0f, 0f, 0f, 0f, 0f, 0f, 0f, 0f, 0f, 0.67451f, 0.988235f, 0.796079f, 0f, 0f, 0f, 0f, 0f, 0f, 0f, 0f, 0f, 0f, 0f, 0f, 0f, 0f, 0f, 0f, 0f, 0f, 0f, 0f, 0f, 0f, 0f, 0f, 0f, 0.0823529f, 0.87451f, 0.917647f, 0.117647f, 0f, 0f, 0f, 0f, 0f, 0f, 0f, 0f, 0f, 0f, 0f, 0f, 0f, 0f, 0f, 0f, 0f, 0f, 0f, 0f, 0f, 0f, 0f, 0f, 0f, 0.478431f, 0.992157f, 0.196078f, 0f, 0f, 0f, 0f, 0f, 0f, 0f, 0f, 0f, 0f, 0f, 0f, 0f, 0f, 0f, 0f, 0f, 0f, 0f, 0f, 0f, 0f, 0f, 0f, 0f, 0.482353f, 0.996078f, 0.356863f, 0.2f, 0.2f, 0.2f, 0.0392157f, 0f, 0f, 0f, 0f, 0f, 0f, 0f, 0f, 0f, 0f, 0f, 0f, 0f, 0f, 0f, 0f, 0f, 0f, 0f, 0f, 0.0823529f, 0.87451f, 0.992157f, 0.988235f, 0.992157f, 0.988235f, 0.992157f, 0.67451f, 0.321569f, 0f, 0f, 0f, 0f, 0f, 0f, 0f, 0f, 0f, 0f, 0f, 0f, 0f, 0f, 0f, 0f, 0f, 0f, 0.0823529f, 0.839216f, 0.992157f, 0.796079f, 0.635294f, 0.4f, 0.4f, 0.796079f, 0.87451f, 0.996078f, 0.992157f, 0.2f, 0.0392157f, 0f, 0f, 0f, 0f, 0f, 0f, 0f, 0f, 0f, 0f, 0f, 0f, 0f, 0f, 0f, 0.239216f, 0.992157f, 0.670588f, 0f, 0f, 0f, 0f, 0f, 0.0784314f, 0.439216f, 0.752941f, 0.992157f, 0.831373f, 0.160784f, 0f, 0f, 0f, 0f, 0f, 0f, 0f, 0f, 0f, 0f, 0f, 0f, 0f, 0f, 0f, 0f, 0f, 0f, 0f, 0f, 0f, 0f, 0f, 0f, 0f, 0.4f, 0.796079f, 0.917647f, 0.2f, 0f, 0f, 0f, 0f, 0f, 0f, 0f, 0f, 0f, 0f, 0f, 0f, 0f, 0f, 0f, 0f, 0f, 0f, 0f, 0f, 0f, 0f, 0f, 0f, 0f, 0.0784314f, 0.835294f, 0.909804f, 0.321569f, 0f, 0f, 0f, 0f, 0f, 0f, 0f, 0f, 0f, 0f, 0f, 0f, 0f, 0f, 0f, 0f, 0f, 0f, 0f, 0f, 0f, 0f, 0f, 0f, 0f, 0.243137f, 0.796079f, 0.917647f, 0.439216f, 0f, 0f, 0f, 0f, 0f, 0f, 0f, 0f, 0f, 0f, 0f, 0f, 0f, 0f, 0f, 0f, 0f, 0f, 0f, 0f, 0f, 0f, 0f, 0f, 0f, 0.0784314f, 0.835294f, 0.988235f, 0f, 0f, 0f, 0f, 0f, 0f, 0f, 0f, 0f, 0f, 0f, 0f, 0f, 0f, 0f, 0f, 0f, 0f, 0f, 0f, 0f, 0f, 0f, 0f, 0f, 0f, 0.6f, 0.992157f, 0f, 0f, 0f, 0f, 0f, 0f, 0f, 0f, 0f, 0f, 0f, 0f, 0f, 0f, 0f, 0f, 0f, 0f, 0f, 0f, 0f, 0f, 0f, 0f, 0f, 0.160784f, 0.913726f, 0.831373f, 0f, 0f, 0f, 0f, 0f, 0f, 0f, 0f, 0f, 0f, 0f, 0f, 0.443137f, 0.360784f, 0f, 0f, 0f, 0f, 0f, 0f, 0f, 0f, 0f, 0f, 0.121569f, 0.678431f, 0.956863f, 0.156863f, 0f, 0f, 0f, 0f, 0f, 0f, 0f, 0f, 0f, 0f, 0f, 0.321569f, 0.992157f, 0.592157f, 0f, 0f, 0f, 0f, 0f, 0f, 0.0823529f, 0.4f, 0.4f, 0.717647f, 0.913726f, 0.831373f, 0.317647f, 0f, 0f, 0f, 0f, 0f, 0f, 0f, 0f, 0f, 0f, 0f, 0f, 0.321569f, 1.0f, 0.992157f, 0.917647f, 0.596078f, 0.6f, 0.756863f, 0.678431f, 0.992157f, 0.996078f, 0.992157f, 0.996078f, 0.835294f, 0.556863f, 0.0784314f, 0f, 0f, 0f, 0f, 0f, 0f, 0f, 0f, 0f, 0f, 0f, 0f, 0f, 0f, 0.278431f, 0.592157f, 0.592157f, 0.909804f, 0.992157f, 0.831373f, 0.752941f, 0.592157f, 0.513726f, 0.196078f, 0.196078f, 0.0392157f, 0f, 0f, 0f, 0f, 0f, 0f, 0f, 0f, 0f, 0f, 0f, 0f, 0f, 0f, 0f, 0f, 0f, 0f, 0f, 0f, 0f, 0f, 0f, 0f, 0f, 0f, 0f, 0f, 0f, 0f, 0f, 0f, 0f, 0f, 0f, 0f, 0f, 0f, 0f, 0f, 0f, 0f, 0f, 0f, 0f, 0f, 0f, 0f, 0f, 0f, 0f, 0f, 0f, 0f, 0f, 0f, 0f, 0f, 0f, 0f, 0f, 0f, 0f, 0f, 0f, 0f, 0f, 0f, 0f, 0f, 0f, 0f, 0f, 0f, 0f, 0f, 0f, 0f, 0f, 0f, 0f, 0f, 0f, 0f, 0f, 0f, 0f, 0f, 0f, 0f, 0f, 0f, 0f, 0f, 0f, 0f, 0f, 0f, 0f, 0f, 0f, 0f, 0f, 0f, 0f, 0f, 0f, 0f, 0f, 0f, 0f, 0f, 0f, 0f, 0f, 0f, 0f, 0f, 0f, 0.0f};
long[] shape = new long[]{1, 28, 28, 1};

Tensor data = Tensor.create(shape, FloatBuffer.wrap(a));

//根据图建立Session
try (Session session = new Session(graph)) {
//相当于TensorFlow Python中的sess.run(z, feed_dict = {'x': 10.0})
Tensor<?> out = session.runner()
.feed(INPUT_TENSOR_NAME, data)
.fetch(OUTPUT_TENSOR_NAME).run().get(0);

// Tensor结果转换
float[][] t = new float[1][10];
out.copyTo(t);
float max = 0f;
float[] result = t[0];

int label = 0;
for (int i = 0; i < result.length; i++) {
float score = result[i];
System.out.println(score);
if (score > max) {
max = score;
label = i;
}
}
System.out.println(label);
}

}
}
}