1. 安装tensorflow和golang(参考https://tensorflow.google.cn/install/install_go)
2. python训练模型,这里以keras example的imdb_cnn.py为例:
# coding:utf-8
import tensorflow as tf
from keras.models import Sequential
from keras.layer import Embedding, Dropout, Conv1D, Dense, GlobalMaxPooling1D
from keras.preprocessing import sequence
from keras.datasets import imdb
from keras import backend as K
# 代码源于keras example的 imdb_cnn.py
max_features = 5000
maxlen = 20
batch_size = 32
embedding_dims = 50
filters = 250
kernel_size = 3
hidden_dims = 250
epochs = 2
# 读取数据
print('Loading data...')
(x_train, y_train), (x_test, y_test) = imdb.load_data(num_words=max_features)
print(len(x_train), 'train sequences')
print(len(x_test), '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)
# 定义模型
# 这里对每个层都添加一个name
model = Sequential()
model.add(Embedding(max_features + 1, 100, name="input_layer"))
model.add(Dropout(0.2, name="dropout_layer1"))
model.add(Conv1D(filters=256, kernel_size=5, padding="valid", activation="relu", strides=1, name="cnn_layer"))
model.add(Dropout(0.2, name="dropout_layer2"))
model.add(GlobalMaxPooling1D(name="maxpooling_layer"))
model.add(Dense(256, activation='relu', name="dense_layer1"))
model.add(Dropout(0.2, name="dropout_layer3"))
model.add(Dense(2, activation='softmax', name="output_layer"))
# 模型训练
sess = tf.Session()
K.set_session(sess)
# 这步找到input_layer和output_layer的完整路径,在golang中使用时需要用来定义输入输出node
for n in sess.graph.as_graph_def().node:
if 'input_layer' in n.name:
print(n.name)
if 'output_layer' in n.name:
print(n.name)
model.fit(x_train,
y_train,
batch_size=batch_size,
epochs=epochs,
validation_data=(x_test, y_test),
shuffle=1)
# 以下是关键代码
# Use TF to save the graph model instead of Keras save model to load it in Golang
builder = tf.saved_model.builder.SavedModelBuilder("cnnModel")
# Tag the model, required for Go
builder.add_meta_graph_and_variables(sess, ["myTag"])
builder.save()
sess.close()
生成cnnModel文件夹,里面包括了一个.pd文件和variables文件夹。
3. golang中使用训练好的模型
package main
import (
"fmt"
tf "github.com/tensorflow/tensorflow/tensorflow/go"
)
func main() {
// 句子最大长度
const MAXLEN int = 20
// 将文本转换为id序列,为了实验方便直接使用转换好的ID序列即可
input_data := [1][MAXLEN]float32{{0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 208.0, 659.0, 180.0, 408.0, 42.0, 547.0, 829.0, 285.0, 334.0, 42.0, 642.0, 81.0, 800.0}}
tensor, err := tf.NewTensor(input_data)
if err != nil {
fmt.Printf("Error NewTensor: err: %s", err.Error())
return
}
//读取模型
model, err := tf.LoadSavedModel("cnnModel", []string{"myTag"}, nil)
if err != nil {
fmt.Printf("Error loading Saved Model: %s\n", err.Error())
return
}
// 识别
result, err := model.Session.Run(
map[tf.Output]*tf.Tensor{
// python版tensorflow/keras中定义的输入层input_layer
model.Graph.Operation("input_layer").Output(0): tensor,
},
[]tf.Output{
// python版tensorflow/keras中定义的输出层output_layer
model.Graph.Operation("output_layer/Softmax").Output(0),
},
nil,
)
if err != nil {
fmt.Printf("Error running the session with input, err: %s ", err.Error())
return
}
// 输出结果,interface{}格式
fmt.Printf("Result value: %v", result[0].Value())
}
4. 预测性能对比:
任务:文本二分类
测试样本数:6万
padding长度:200
平台:只使用CPU
python用时110秒,golang用时30秒,据说图像识别速度会相差10倍以上。