Python learning.create_train_op函数代码示例

  def testTrainingSubsetsOfVariablesOnlyUpdatesThoseVariables(self):
    # First, train only the weights of the model.
    with ops.Graph().as_default():
      random_seed.set_random_seed(0)
      total_loss = self.ModelLoss()
      optimizer = gradient_descent.GradientDescentOptimizer(learning_rate=1.0)
      weights, biases = variables_lib2.get_variables()

      train_op = learning.create_train_op(total_loss, optimizer)
      train_weights = learning.create_train_op(
          total_loss, optimizer, variables_to_train=[weights])
      train_biases = learning.create_train_op(
          total_loss, optimizer, variables_to_train=[biases])

      with session.Session() as sess:
        # Initialize the variables.
        sess.run(variables_lib.global_variables_initializer())

        # Get the initial weights and biases values.
        weights_values, biases_values = sess.run([weights, biases])
        self.assertGreater(np.linalg.norm(weights_values), 0)
        self.assertAlmostEqual(np.linalg.norm(biases_values), 0)

        # Update weights and biases.
        loss = sess.run(train_op)
        self.assertGreater(loss, .5)
        new_weights, new_biases = sess.run([weights, biases])

        # Check that the weights and biases have been updated.
        self.assertGreater(np.linalg.norm(weights_values - new_weights), 0)
        self.assertGreater(np.linalg.norm(biases_values - new_biases), 0)

        weights_values, biases_values = new_weights, new_biases

        # Update only weights.
        loss = sess.run(train_weights)
        self.assertGreater(loss, .5)
        new_weights, new_biases = sess.run([weights, biases])

        # Check that the weights have been updated, but biases have not.
        self.assertGreater(np.linalg.norm(weights_values - new_weights), 0)
        self.assertAlmostEqual(np.linalg.norm(biases_values - new_biases), 0)
        weights_values = new_weights

        # Update only biases.
        loss = sess.run(train_biases)
        self.assertGreater(loss, .5)
        new_weights, new_biases = sess.run([weights, biases])

        # Check that the biases have been updated, but weights have not.
        self.assertAlmostEqual(np.linalg.norm(weights_values - new_weights), 0)
        self.assertGreater(np.linalg.norm(biases_values - new_biases), 0)

  def testResumeTrainAchievesRoughlyTheSameLoss(self):
    logdir = os.path.join(
        tempfile.mkdtemp(prefix=self.get_temp_dir()), 'tmp_logs')
    number_of_steps = [300, 301, 305]

    for i in range(len(number_of_steps)):
      with ops.Graph().as_default():
        random_seed.set_random_seed(i)
        tf_inputs = constant_op.constant(self._inputs, dtype=dtypes.float32)
        tf_labels = constant_op.constant(self._labels, dtype=dtypes.float32)

        tf_predictions = LogisticClassifier(tf_inputs)
        loss_ops.log_loss(tf_predictions, tf_labels)
        total_loss = loss_ops.get_total_loss()

        optimizer = gradient_descent.GradientDescentOptimizer(learning_rate=1.0)

        train_op = learning.create_train_op(total_loss, optimizer)

        loss = learning.train(
            train_op,
            logdir,
            number_of_steps=number_of_steps[i],
            log_every_n_steps=10)
        self.assertIsNotNone(loss)
        self.assertLess(loss, .015)

  def testTrainWithTrace(self):
    logdir = os.path.join(
        tempfile.mkdtemp(prefix=self.get_temp_dir()), 'tmp_logs')
    with ops.Graph().as_default():
      random_seed.set_random_seed(0)
      tf_inputs = constant_op.constant(self._inputs, dtype=dtypes.float32)
      tf_labels = constant_op.constant(self._labels, dtype=dtypes.float32)

      tf_predictions = LogisticClassifier(tf_inputs)
      loss_ops.log_loss(tf_predictions, tf_labels)
      total_loss = loss_ops.get_total_loss()
      summary.scalar('total_loss', total_loss)

      optimizer = gradient_descent.GradientDescentOptimizer(learning_rate=1.0)

      train_op = learning.create_train_op(total_loss, optimizer)

      loss = learning.train(
          train_op,
          logdir,
          number_of_steps=300,
          log_every_n_steps=10,
          trace_every_n_steps=100)
    self.assertIsNotNone(loss)
    for trace_step in [1, 101, 201]:
      trace_filename = 'tf_trace-%d.json' % trace_step
      self.assertTrue(os.path.isfile(os.path.join(logdir, trace_filename)))

  def testNoneGlobalStep(self):
    with ops.Graph().as_default():
      random_seed.set_random_seed(0)
      tf_inputs = constant_op.constant(self._inputs, dtype=dtypes.float32)
      tf_labels = constant_op.constant(self._labels, dtype=dtypes.float32)

      tf_predictions = BatchNormClassifier(tf_inputs)
      loss_ops.log_loss(tf_predictions, tf_labels)
      total_loss = loss_ops.get_total_loss()
      optimizer = gradient_descent.GradientDescentOptimizer(learning_rate=1.0)

      train_op = learning.create_train_op(
          total_loss, optimizer, global_step=None)

      global_step = variables_lib2.get_or_create_global_step()

      with session.Session() as sess:
        # Initialize all variables
        sess.run(variables_lib.global_variables_initializer())

        for _ in range(10):
          sess.run([train_op])
        global_step = global_step.eval()
        # Since train_op don't use global_step it shouldn't change.
        self.assertAllClose(global_step, 0)

  def testTrainWithSessionWrapper(self):
    """Test that slim.learning.train can take `session_wrapper` args.

    One of the applications of `session_wrapper` is the wrappers of TensorFlow
    Debugger (tfdbg), which intercept methods calls to `tf.Session` (e.g., run)
    to achieve debugging. `DumpingDebugWrapperSession` is used here for testing
    purpose.
    """
    dump_root = tempfile.mkdtemp()

    def dumping_wrapper(sess):  # pylint: disable=invalid-name
      return dumping_wrapper_lib.DumpingDebugWrapperSession(sess, dump_root)

    with ops.Graph().as_default():
      random_seed.set_random_seed(0)
      tf_inputs = constant_op.constant(self._inputs, dtype=dtypes.float32)
      tf_labels = constant_op.constant(self._labels, dtype=dtypes.float32)

      tf_predictions = LogisticClassifier(tf_inputs)
      loss_ops.log_loss(tf_predictions, tf_labels)
      total_loss = loss_ops.get_total_loss()

      optimizer = gradient_descent.GradientDescentOptimizer(learning_rate=1.0)

      train_op = learning.create_train_op(total_loss, optimizer)

      loss = learning.train(
          train_op, None, number_of_steps=1, session_wrapper=dumping_wrapper)
    self.assertIsNotNone(loss)

    run_root = glob.glob(os.path.join(dump_root, 'run_*'))[-1]
    dump = debug_data.DebugDumpDir(run_root)
    self.assertAllEqual(0,
                        dump.get_tensors('global_step', 0, 'DebugIdentity')[0])

  def testEmptyUpdateOps(self):
    with ops.Graph().as_default():
      random_seed.set_random_seed(0)
      tf_inputs = constant_op.constant(self._inputs, dtype=dtypes.float32)
      tf_labels = constant_op.constant(self._labels, dtype=dtypes.float32)

      tf_predictions = BatchNormClassifier(tf_inputs)
      loss_ops.log_loss(tf_predictions, tf_labels)
      total_loss = loss_ops.get_total_loss()
      optimizer = gradient_descent.GradientDescentOptimizer(learning_rate=1.0)

      train_op = learning.create_train_op(total_loss, optimizer, update_ops=[])

      moving_mean = variables_lib2.get_variables_by_name('moving_mean')[0]
      moving_variance = variables_lib2.get_variables_by_name('moving_variance')[
          0]

      with session.Session() as sess:
        # Initialize all variables
        sess.run(variables_lib.global_variables_initializer())
        mean, variance = sess.run([moving_mean, moving_variance])
        # After initialization moving_mean == 0 and moving_variance == 1.
        self.assertAllClose(mean, [0] * 4)
        self.assertAllClose(variance, [1] * 4)

        for _ in range(10):
          sess.run([train_op])
        mean = moving_mean.eval()
        variance = moving_variance.eval()
        # Since we skip update_ops the moving_vars are not updated.
        self.assertAllClose(mean, [0] * 4)
        self.assertAllClose(variance, [1] * 4)

  def testTrainAllVarsHasLowerLossThanTrainSubsetOfVars(self):
    logdir1 = os.path.join(
        tempfile.mkdtemp(prefix=self.get_temp_dir()), 'tmp_logs1')

    # First, train only the weights of the model.
    with ops.Graph().as_default():
      random_seed.set_random_seed(0)
      total_loss = self.ModelLoss()
      optimizer = gradient_descent.GradientDescentOptimizer(learning_rate=1.0)
      weights = variables_lib2.get_variables_by_name('weights')

      train_op = learning.create_train_op(
          total_loss, optimizer, variables_to_train=weights)

      loss = learning.train(
          train_op, logdir1, number_of_steps=200, log_every_n_steps=10)
      self.assertGreater(loss, .015)
      self.assertLess(loss, .05)

    # Next, train the biases of the model.
    with ops.Graph().as_default():
      random_seed.set_random_seed(1)
      total_loss = self.ModelLoss()
      optimizer = gradient_descent.GradientDescentOptimizer(learning_rate=1.0)
      biases = variables_lib2.get_variables_by_name('biases')

      train_op = learning.create_train_op(
          total_loss, optimizer, variables_to_train=biases)

      loss = learning.train(
          train_op, logdir1, number_of_steps=300, log_every_n_steps=10)
      self.assertGreater(loss, .015)
      self.assertLess(loss, .05)

    # Finally, train both weights and bias to get lower loss.
    with ops.Graph().as_default():
      random_seed.set_random_seed(2)
      total_loss = self.ModelLoss()
      optimizer = gradient_descent.GradientDescentOptimizer(learning_rate=1.0)

      train_op = learning.create_train_op(total_loss, optimizer)
      loss = learning.train(
          train_op, logdir1, number_of_steps=400, log_every_n_steps=10)

      self.assertIsNotNone(loss)
      self.assertLess(loss, .015)

  def testRecordTrainOpInCollection(self):
    with ops.Graph().as_default():
      random_seed.set_random_seed(0)
      tf_inputs = constant_op.constant(self._inputs, dtype=dtypes.float32)
      tf_labels = constant_op.constant(self._labels, dtype=dtypes.float32)

      tf_predictions = LogisticClassifier(tf_inputs)
      loss_ops.log_loss(tf_predictions, tf_labels)
      total_loss = loss_ops.get_total_loss()

      optimizer = gradient_descent.GradientDescentOptimizer(learning_rate=1.0)
      train_op = learning.create_train_op(total_loss, optimizer)

      # Make sure the training op was recorded in the proper collection
      self.assertTrue(train_op in ops.get_collection(ops.GraphKeys.TRAIN_OP))

  def testTrainWithNoneAsLogdirWhenUsingTraceRaisesError(self):
    with ops.Graph().as_default():
      random_seed.set_random_seed(0)
      tf_inputs = constant_op.constant(self._inputs, dtype=dtypes.float32)
      tf_labels = constant_op.constant(self._labels, dtype=dtypes.float32)

      tf_predictions = LogisticClassifier(tf_inputs)
      loss_ops.log_loss(tf_predictions, tf_labels)
      total_loss = loss_ops.get_total_loss()

      optimizer = gradient_descent.GradientDescentOptimizer(learning_rate=1.0)

      train_op = learning.create_train_op(total_loss, optimizer)

      with self.assertRaises(ValueError):
        learning.train(
            train_op, None, number_of_steps=300, trace_every_n_steps=10)

  def testTrainWithNoneAsInitWhenUsingVarsRaisesError(self):
    logdir = os.path.join(
        tempfile.mkdtemp(prefix=self.get_temp_dir()), 'tmp_logs')
    with ops.Graph().as_default():
      random_seed.set_random_seed(0)
      tf_inputs = constant_op.constant(self._inputs, dtype=dtypes.float32)
      tf_labels = constant_op.constant(self._labels, dtype=dtypes.float32)

      tf_predictions = LogisticClassifier(tf_inputs)
      loss_ops.log_loss(tf_predictions, tf_labels)
      total_loss = loss_ops.get_total_loss()

      optimizer = gradient_descent.GradientDescentOptimizer(learning_rate=1.0)

      train_op = learning.create_train_op(total_loss, optimizer)

      with self.assertRaises(RuntimeError):
        learning.train(train_op, logdir, init_op=None, number_of_steps=300)

  def create_train_op(self, learning_rate=1.0, gradient_multiplier=1.0):
    tf_inputs = constant_op.constant(self._inputs, dtype=dtypes.float32)
    tf_labels = constant_op.constant(self._labels, dtype=dtypes.float32)

    tf_predictions = LogisticClassifier(tf_inputs)
    loss_ops.log_loss(tf_predictions, tf_labels)
    total_loss = loss_ops.get_total_loss()

    optimizer = gradient_descent.GradientDescentOptimizer(
        learning_rate=learning_rate)

    if gradient_multiplier != 1.0:
      variables = variables_lib.trainable_variables()
      gradient_multipliers = {var: gradient_multiplier for var in variables}
    else:
      gradient_multipliers = None

    return learning.create_train_op(
        total_loss, optimizer, gradient_multipliers=gradient_multipliers)

  def testTrainWithNoInitAssignCanAchieveZeroLoss(self):
    logdir = os.path.join(
        tempfile.mkdtemp(prefix=self.get_temp_dir()), 'tmp_logs')
    with ops.Graph().as_default():
      random_seed.set_random_seed(0)
      tf_inputs = constant_op.constant(self._inputs, dtype=dtypes.float32)
      tf_labels = constant_op.constant(self._labels, dtype=dtypes.float32)

      tf_predictions = LogisticClassifier(tf_inputs)
      loss_ops.log_loss(tf_predictions, tf_labels)
      total_loss = loss_ops.get_total_loss()

      optimizer = gradient_descent.GradientDescentOptimizer(learning_rate=1.0)

      train_op = learning.create_train_op(total_loss, optimizer)

      loss = learning.train(
          train_op, logdir, number_of_steps=300, log_every_n_steps=10)
      self.assertIsNotNone(loss)
      self.assertLess(loss, .015)

  def testTrainWithSessionConfig(self):
    with ops.Graph().as_default():
      random_seed.set_random_seed(0)
      tf_inputs = constant_op.constant(self._inputs, dtype=dtypes.float32)
      tf_labels = constant_op.constant(self._labels, dtype=dtypes.float32)

      tf_predictions = LogisticClassifier(tf_inputs)
      loss_ops.log_loss(tf_predictions, tf_labels)
      total_loss = loss_ops.get_total_loss()

      optimizer = gradient_descent.GradientDescentOptimizer(learning_rate=1.0)

      train_op = learning.create_train_op(total_loss, optimizer)

      session_config = config_pb2.ConfigProto(allow_soft_placement=True)
      loss = learning.train(
          train_op,
          None,
          number_of_steps=300,
          log_every_n_steps=10,
          session_config=session_config)
    self.assertIsNotNone(loss)
    self.assertLess(loss, .015)

  def testTrainWithEpochLimit(self):
    logdir = os.path.join(tempfile.mkdtemp(prefix=self.get_temp_dir()),
                          'tmp_logs')
    with ops.Graph().as_default():
      random_seed.set_random_seed(0)
      tf_inputs = constant_op.constant(self._inputs, dtype=dtypes.float32)
      tf_labels = constant_op.constant(self._labels, dtype=dtypes.float32)
      tf_inputs_limited = input_lib.limit_epochs(tf_inputs, num_epochs=300)
      tf_labels_limited = input_lib.limit_epochs(tf_labels, num_epochs=300)

      tf_predictions = LogisticClassifier(tf_inputs_limited)
      loss_ops.log_loss(tf_predictions, tf_labels_limited)
      total_loss = loss_ops.get_total_loss()

      optimizer = gradient_descent.GradientDescentOptimizer(learning_rate=1.0)

      train_op = learning.create_train_op(total_loss, optimizer)

      loss = learning.train(train_op, logdir, log_every_n_steps=10)
    self.assertIsNotNone(loss)
    self.assertLess(loss, .015)
    self.assertTrue(os.path.isfile('{}/model.ckpt-300.index'.format(logdir)))
    self.assertTrue(os.path.isfile('{}/model.ckpt-300.data-00000-of-00001'.format(logdir)))

  def testUseUpdateOps(self):
    with ops.Graph().as_default():
      random_seed.set_random_seed(0)
      tf_inputs = constant_op.constant(self._inputs, dtype=dtypes.float32)
      tf_labels = constant_op.constant(self._labels, dtype=dtypes.float32)

      expected_mean = np.mean(self._inputs, axis=(0))
      expected_var = np.var(self._inputs, axis=(0))
      expected_var = self._addBesselsCorrection(16, expected_var)

      tf_predictions = BatchNormClassifier(tf_inputs)
      loss_ops.log_loss(tf_predictions, tf_labels)
      total_loss = loss_ops.get_total_loss()
      optimizer = gradient_descent.GradientDescentOptimizer(learning_rate=1.0)

      train_op = learning.create_train_op(total_loss, optimizer)

      moving_mean = variables_lib2.get_variables_by_name('moving_mean')[0]
      moving_variance = variables_lib2.get_variables_by_name('moving_variance')[
          0]

      with session.Session() as sess:
        # Initialize all variables
        sess.run(variables_lib.global_variables_initializer())
        mean, variance = sess.run([moving_mean, moving_variance])
        # After initialization moving_mean == 0 and moving_variance == 1.
        self.assertAllClose(mean, [0] * 4)
        self.assertAllClose(variance, [1] * 4)

        for _ in range(10):
          sess.run([train_op])
        mean = moving_mean.eval()
        variance = moving_variance.eval()
        # After 10 updates with decay 0.1 moving_mean == expected_mean and
        # moving_variance == expected_var.
        self.assertAllClose(mean, expected_mean)
        self.assertAllClose(variance, expected_var)