Tensorflow Python 转 Java（一）

一、背景描述

最近python的tensorflow项目要弄到线上去。网络用的Tensorflow现成的包。数据用kaggle中的数据为例子。

数据地址：

https://www.kaggle.com/johnfarrell/gpu-example-from-prepared-data-try-deepfm

二、Python代码

1、Python Code

 # author: adrian.wu
 from __future__ import absolute_import
 from __future__ import division
 from __future__ import print_function
 
 import tensorflow as tf
 
 tf.logging.set_verbosity(tf.logging.INFO)
 # Set to INFO for tracking training, default is WARN
 
 print("Using TensorFlow version %s" % (tf.__version__))
 
 CATEGORICAL_COLUMNS = ["workclass", "education",
                        "marital.status", "occupation",
                        "relationship", "race",
                        "sex", "native.country"]
 
 # Columns of the input csv file
 COLUMNS = ["age", "workclass", "fnlwgt", "education",
            "education.num", "marital.status",
            "occupation", "relationship", "race",
            "sex", "capital.gain", "capital.loss",
            "hours.per.week", "native.country", "income"]
 
 FEATURE_COLUMNS = ["age", "workclass", "education",
                    "education.num", "marital.status",
                    "occupation", "relationship", "race",
                    "sex", "capital.gain", "capital.loss",
                    "hours.per.week", "native.country"]
 
 import pandas as pd
 
 df = pd.read_csv("/Users/adrian.wu/Desktop/learn/kaggle/adult-census-income/adult.csv")
 
 from sklearn.model_selection import train_test_split
 
 BATCH_SIZE = 40
 
 num_epochs = 1
 shuffle = True
 
 y = df["income"].apply(lambda x: ">50K" in x).astype(int)
 del df["fnlwgt"]  # Unused column
 del df["income"]  # Labels column, already saved to labels variable
 X = df
 
 print(X.describe())
 
 X_train, X_test, y_train, y_test = train_test_split(X, y, test_size=0.20)
 
 train_input_fn = tf.estimator.inputs.pandas_input_fn(
     x=X_train,
     y=y_train,
     batch_size=BATCH_SIZE,
     num_epochs=num_epochs,
     shuffle=shuffle)
 
 eval_input_fn = tf.estimator.inputs.pandas_input_fn(
     x=X_test,
     y=y_test,
     batch_size=BATCH_SIZE,
     num_epochs=num_epochs,
     shuffle=shuffle)
 
 def generate_input_fn(filename, num_epochs=None, shuffle=True, batch_size=BATCH_SIZE):
     df = pd.read_csv(filename)  # , header=None, names=COLUMNS)
     labels = df["income"].apply(lambda x: ">50K" in x).astype(int)
     del df["fnlwgt"]  # Unused column
     del df["income"]  # Labels column, already saved to labels variable
 
     type(df['age'].iloc[3])
 
     return tf.estimator.inputs.pandas_input_fn(
         x=df,
         y=labels,
         batch_size=batch_size,
         num_epochs=num_epochs,
         shuffle=shuffle)
 
 sex = tf.feature_column.categorical_column_with_vocabulary_list(
     key="sex",
     vocabulary_list=["female", "male"])
 race = tf.feature_column.categorical_column_with_vocabulary_list(
     key="race",
     vocabulary_list=["Amer-Indian-Eskimo",
                      "Asian-Pac-Islander",
                      "Black", "Other", "White"])
 
 # 先对categorical的列做hash
 education = tf.feature_column.categorical_column_with_hash_bucket(
     "education", hash_bucket_size=1000)
 marital_status = tf.feature_column.categorical_column_with_hash_bucket(
     "marital.status", hash_bucket_size=100)
 relationship = tf.feature_column.categorical_column_with_hash_bucket(
     "relationship", hash_bucket_size=100)
 workclass = tf.feature_column.categorical_column_with_hash_bucket(
     "workclass", hash_bucket_size=100)
 occupation = tf.feature_column.categorical_column_with_hash_bucket(
     "occupation", hash_bucket_size=1000)
 native_country = tf.feature_column.categorical_column_with_hash_bucket(
     "native.country", hash_bucket_size=1000)
 
 print('Categorical columns configured')
 
 age = tf.feature_column.numeric_column("age")
 deep_columns = [
     # Multi-hot indicator columns for columns with fewer possibilities
     tf.feature_column.indicator_column(workclass),
     tf.feature_column.indicator_column(marital_status),
     tf.feature_column.indicator_column(sex),
     tf.feature_column.indicator_column(relationship),
     tf.feature_column.indicator_column(race),
     # Embeddings for categories with more possibilities. Should have at least (possibilties)**(0.25) dims
     tf.feature_column.embedding_column(education, dimension=8),
     tf.feature_column.embedding_column(native_country, dimension=8),
     tf.feature_column.embedding_column(occupation, dimension=8),
     age
 ]
 
 m2 = tf.estimator.DNNClassifier(
     model_dir="model/dir",
     feature_columns=deep_columns,
     hidden_units=[100, 50])
 
 m2.train(input_fn=train_input_fn)
 
 start, end = 0, 5
 data_predict = df.iloc[start:end]
 predict_labels = y.iloc[start:end]
 print(predict_labels)
 print(data_predict.head(12))  # show this before deleting, so we know what the labels
 predict_input_fn = tf.estimator.inputs.pandas_input_fn(
     x=data_predict,
     batch_size=1,
     num_epochs=1,
     shuffle=False)
 
 predictions = m2.predict(input_fn=predict_input_fn)
 
 for prediction in predictions:
     print("Predictions:    {} with probabilities {}\n".format(prediction["classes"], prediction["probabilities"]))
 
 def column_to_dtype(column):
     if column in CATEGORICAL_COLUMNS:
         return tf.string
     else:
         return tf.float32
 
 # 什么数据要喂给输入
 FEATURE_COLUMNS_FOR_SERVE = ["workclass", "education",
                              "marital.status", "occupation",
                              "relationship", "race",
                              "sex", "native.country", "age"]
 
 serving_features = {column: tf.placeholder(shape=[1], dtype=column_to_dtype(column), name=column) for column in
                     FEATURE_COLUMNS_FOR_SERVE}
 # serving_input_receiver_fn有很多种方式
 export_dir = m2.export_savedmodel(export_dir_base="models/export",
                                   serving_input_receiver_fn=tf.estimator.export.build_raw_serving_input_receiver_fn(
                                       serving_features), as_text=True)
 export_dir = export_dir.decode("utf8")

2、通过 export_savedmodel这个函数生成了variables变量和pbtxt文件。如图所示：

3、先打开saved_model.pbtxt文件浏览一下，会发现这是对tensorflow 的一个个描述。包含了node name， operation name，dtype等信息。在套用java时需要明确node的name。

node {
      name: "dnn/head/predictions/probabilities"
      op: "Softmax"
      input: "dnn/head/predictions/two_class_logits"
      attr {
        key: "T"
        value {
          type: DT_FLOAT
        }
      }
      attr {
        key: "_output_shapes"
        value {
          list {
            shape {
              dim {
                size: -1
              }
              dim {
                size: 2
              }
            }
          }
        }

三、Java代码

1、先将variable和pbtxt文件放到resources下面。

2、Java代码

 import org.tensorflow.SavedModelBundle;
 import org.tensorflow.Session;
 import org.tensorflow.Tensor;
 
 /**
  * Created by adrian.wu on 2019/3/14.
  */
 public class TestAdultIncome {
 
     public static void main(String[] args) throws Exception {
 
         SavedModelBundle model = SavedModelBundle.load("/Users/adrian.wu/Desktop/sc/adrian_test/src/main/resources/adultincomemodel", "serve");
         Session sess = model.session();
 
         String sex = "Female";
         String workclass = "?";
         String education = "HS-grad";
         String ms = "Widowed";
         String occupation = "?";
         String relationship = "Not-in-family";
         String race = "White";
         String nc = "United-States";
 
         //不能将string直接喂给create()接口
         Tensor sexTensor = Tensor.create(new byte[][]{sex.getBytes()});
         Tensor workclassTensor = Tensor.create(new byte[][]{workclass.getBytes()});
         Tensor eduTensor = Tensor.create(new byte[][]{education.getBytes()});
         Tensor msTensor = Tensor.create(new byte[][]{ms.getBytes()});
         Tensor occuTensor = Tensor.create(new byte[][]{occupation.getBytes()});
         Tensor ralaTensor = Tensor.create(new byte[][]{relationship.getBytes()});
         Tensor raceTensor = Tensor.create(new byte[][]{race.getBytes()});
         Tensor ncTesnsor = Tensor.create(new byte[][]{nc.getBytes()});
 
         float[][] age = {{90f}};
 
         Tensor ageTensor = Tensor.create(age);
 
         //根据pbtxt文件，查看operation name。
         Tensor result = sess.runner()
                 .feed("workclass", workclassTensor)
                 .feed("education", eduTensor)
                 .feed("marital.status", msTensor)
                 .feed("relationship", ralaTensor)
                 .feed("race", raceTensor)
                 .feed("sex", sexTensor)
                 .feed("native.country", ncTesnsor)
                 .feed("occupation",occuTensor)
                 .feed("age", ageTensor)
                 .fetch("dnn/head/predictions/probabilities")
                 .run()
                 .get(0);
 
         float[][] buffer = new float[1][2];
         result.copyTo(buffer);
         System.out.println("" + String.valueOf(buffer[0][0]));
     }
 
 }

四、结果对比

python和java结果：

 java: 0.9432887
 python: 0.9432887