Create DNN Using Ready-to-Use TensorFlow Estimator

TensorFlow has some ready-to-use classifiers and regressors: DNNClassifier, DNNRegressor, etc.

Libraries used in this example:

• tensorflow: The machine learning library
• numpy: Library to create inputs for TensorFlow
• matplotlib.pyplot: Utility library to draw, eg. loss curve, accuracy curve, etc.

Utility functions:

• log: Print a log line using tf.logging.info
• log_sep: Print separator in log

TensorFlow functions:

• tf.feature_column.numeric_column: Create a feature of numeric input values
• tf.estimator.inputs.numpy_input_fn: Create TensorFlow input function from numpy array

TensorFlow classes:

• tf.estimator.DNNClassifier: The ready-to-use DNN classifier class

#core
import sys,json;

#libs
import tensorflow        as tf;
import numpy             as np;
import matplotlib.pyplot as pyplot;

'''
\brief Log
'''
def log(*Args):
  Str = "";
  for I in range(len(Args)):
    Str += str(Args[I]);
    if (I<len(Args)-1): Str += "\x20";
      
  tf.logging.info(Str);
#end def

'''
\brief Log a separator
'''
def log_sep():
  tf.logging.info("\n"+"="*80);
#end def

'''
\brief Train/eval using ready-to-use estimator DNNClassifier
'''
def train_eval_estimator(X,Y):
  Feature    = tf.feature_column.numeric_column(key="X",shape=[2]);
  Classifier = tf.estimator.DNNClassifier(
    feature_columns = [Feature],
    hidden_units    = [8,4],
    n_classes       = 2
  );
  
  Train_Input_Fn = tf.estimator.inputs.numpy_input_fn(
    x          = {"X":X},
    y          = Y,
    batch_size = 4,
    num_epochs = None,
    shuffle    = True
  );
  
  Eval_Input_Fn = tf.estimator.inputs.numpy_input_fn(
    x          = {"X":X},
    y          = Y,
    batch_size = 4,
    num_epochs = 1,
    shuffle    = False
  );
  
  STEPS  = 1000;
  Losses = [];
  
  for I in range(10):
    log_sep();
    log("\nTraining steps from {}...".format(I*STEPS));
    Classifier.train(input_fn=Train_Input_Fn, steps=STEPS);
    
    log_sep();
    log("\nEvaluating after {} steps...".format((I+1)*STEPS));
    Result = Classifier.evaluate(input_fn=Eval_Input_Fn);
    
    Losses += [Result["loss"]];
  #end for
  
  log_sep();
  log("\nLosses:",Losses);
  pyplot.plot(Losses);
#end def

#PROGRAMME ENTRY POINT==========================================================
tf.logging.set_verbosity(tf.logging.INFO);
log("TensorFlow version:",tf.__version__);

#data
X = [[0,0],[0,1],[1,0],[1,1]];
Y = [ 0,    1,    1,    0   ];
X = np.array(X);
Y = np.array(Y);

train_eval_estimator(X,Y);
#eof

Colab link:
https://colab.research.google.com/drive/1CkpVsVbHKxrgLFqoGee-B6iLpyRdiode