tmp.py

predictors = df_train.iloc[:,FEATURE_COLS].values
targets = df_train[TARGET_COL].values
shutil.rmtree('taxi_model', ignore_errors=True) # start fresh each time
modelprefix = 'taxi_model'
with tf.Session() as sess:
  npredictors = len(FEATURE_COLS)
  noutputs = 1
  feature_data = tf.placeholder("float", [None, npredictors])
  target_data = tf.placeholder("float", [None, noutputs])
  hidden = tf.contrib.layers.stack(feature_data, 
                                  tf.contrib.layers.fully_connected, 
                                  [64, 8],
                                  activation_fn=tf.nn.relu,
                                  # biases_initializer=tf.ones,
                                  scope='fc')
  model = tf.contrib.layers.fully_connected(hidden, 1, activation_fn=None)
  cost = tf.nn.l2_loss(model - target_data)
  # tf.scalar_summary("cost", cost)
  tf.summary.scalar("cost", cost)
  training_step = tf.contrib.layers.optimize_loss(cost, 
                                                  tf.contrib.framework.get_global_step(),
                                                  optimizer='Adam', learning_rate=0.1)

#   summary_writer = tf.train.SummaryWriter('taxi_model',graph=sess.graph)
  summary_writer = tf.summary.FileWriter('taxi_model',graph=sess.graph)

  tf.initialize_all_variables().run()
#   merged_summary_op = tf.merge_all_summaries()
  merged_summary_op = tf.summary.merge_all()
  for iter in xrange(0, 1000):
    _, trainerr, summary = sess.run([training_step, cost, merged_summary_op], feed_dict = {
        feature_data : predictors,
        target_data : targets.reshape(len(predictors), noutputs)
      })
    summary_writer.add_summary(summary)
    if (iter%100 == 1):
      trmse = np.sqrt(trainerr/len(predictors))     
      print 'iter={0} train_error={1}'.format(iter, trmse)