path: root/tensorflow/contrib/distribute/python/minimize_loss_test.py

# Copyright 2018 The TensorFlow Authors. All Rights Reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
#     http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
# ==============================================================================
"""Tests for running legacy optimizer code with DistributionStrategy."""

from __future__ import absolute_import
from __future__ import division
from __future__ import print_function

from absl.testing import parameterized
import numpy

from tensorflow.contrib.distribute.python import combinations
from tensorflow.contrib.distribute.python import mirrored_strategy
from tensorflow.contrib.distribute.python.single_loss_example import batchnorm_example
from tensorflow.contrib.distribute.python.single_loss_example import minimize_loss_example
from tensorflow.python.data.ops import dataset_ops
from tensorflow.python.eager import context
from tensorflow.python.eager import test
from tensorflow.python.framework import constant_op
from tensorflow.python.framework import ops
from tensorflow.python.layers import core
from tensorflow.python.ops import array_ops
from tensorflow.python.ops import control_flow_ops
from tensorflow.python.ops import math_ops
from tensorflow.python.ops import variable_scope
from tensorflow.python.ops import variables as variables_lib
from tensorflow.python.ops.losses import losses_impl


class MinimizeLossStepTest(test.TestCase, parameterized.TestCase):

  @combinations.generate(
      combinations.times(
          combinations.distributions_and_v1_optimizers(),
          combinations.combine(mode=["graph"], use_callable_loss=[True, False])
          + combinations.combine(mode=["eager"], use_callable_loss=[True])) +
      combinations.combine(
          distribution=[combinations.tpu_strategy],
          optimizer_fn=combinations.optimizers_v1,
          mode=["graph"],
          use_callable_loss=[True, False]))
  def testTrainNetwork(self, distribution, optimizer_fn, use_callable_loss):
    with distribution.scope():
      model_fn, dataset_fn, layer = minimize_loss_example(
          optimizer_fn, use_bias=True, use_callable_loss=use_callable_loss)

      def step_fn(ctx, *inputs):
        del ctx  # Unused
        return distribution.group(
            distribution.call_for_each_tower(
                model_fn, *inputs, run_concurrently=layer.built))

      iterator = distribution.distribute_dataset(
          dataset_fn).make_one_shot_iterator()

      def run_step():
        return distribution.run_steps_on_dataset(
            step_fn, iterator, iterations=2).run_op

      self.evaluate(distribution.initialize())
      if not context.executing_eagerly():
        with self.cached_session() as sess:
          run_step = sess.make_callable(run_step())
      self.evaluate(variables_lib.global_variables_initializer())

      weights, biases = [], []
      for _ in range(5):
        run_step()

        weights.append(self.evaluate(layer.kernel))
        biases.append(self.evaluate(layer.bias))

      self.evaluate(distribution.finalize())

      error = abs(numpy.add(numpy.squeeze(weights), numpy.squeeze(biases)) - 1)
      is_not_increasing = all(y <= x for x, y in zip(error, error[1:]))
      self.assertTrue(is_not_increasing)

  @combinations.generate(
      combinations.times(
          combinations.distributions_and_v1_optimizers(),
          combinations.combine(mode=["graph"], use_callable_loss=[True, False])
          + combinations.combine(mode=["eager"], use_callable_loss=[True])))
  def testTrainNetworkByCallForEachTower(self, distribution, optimizer_fn,
                                         use_callable_loss):
    with distribution.scope():
      model_fn, dataset_fn, layer = minimize_loss_example(
          optimizer_fn, use_bias=True, use_callable_loss=use_callable_loss)

      iterator = distribution.distribute_dataset(
          dataset_fn).make_one_shot_iterator()

      def run_step():
        return distribution.group(
            distribution.call_for_each_tower(
                model_fn, iterator.get_next(), run_concurrently=layer.built))

      if not context.executing_eagerly():
        with self.cached_session() as sess:
          run_step = sess.make_callable(run_step())
        self.evaluate(variables_lib.global_variables_initializer())

      weights, biases = [], []
      for _ in range(10):
        run_step()

        weights.append(self.evaluate(layer.kernel))
        biases.append(self.evaluate(layer.bias))

      error = abs(numpy.add(numpy.squeeze(weights), numpy.squeeze(biases)) - 1)
      is_not_increasing = all(y <= x for x, y in zip(error, error[1:]))
      self.assertTrue(is_not_increasing)

  @combinations.generate(
      combinations.times(
          combinations.distributions_and_v1_optimizers() +
          combinations.distributions_and_v2_optimizers(),
          combinations.combine(mode=["graph", "eager"])) +
      combinations.combine(
          distribution=[combinations.tpu_strategy],
          optimizer_fn=combinations.optimizers_v1+combinations.optimizers_v2,
          mode=["graph"]))
  def testOptimizerInsideModelFn(self, distribution, optimizer_fn):
    created_variables = []
    trainable_variables = []

    def appending_creator(next_creator, *args, **kwargs):
      v = next_creator(*args, **kwargs)
      created_variables.append(v.name)
      if "trainable" in kwargs and kwargs["trainable"]:
        trainable_variables.append(v.name)
      return v

    # Creator scope needs to be set before it's used inside
    # `distribution.scope`.
    with variable_scope.variable_creator_scope(
        appending_creator), distribution.scope():
      model_fn, dataset_fn, layer = minimize_loss_example(
          optimizer_fn,
          use_bias=True,
          use_callable_loss=True,
          create_optimizer_inside_model_fn=True)

      def step_fn(ctx, *inputs):
        del ctx  # Unused
        return distribution.group(
            distribution.call_for_each_tower(
                model_fn, *inputs, run_concurrently=layer.built))

      iterator = distribution.distribute_dataset(
          dataset_fn).make_one_shot_iterator()

      def run_step():
        return distribution.run_steps_on_dataset(
            step_fn, iterator, iterations=1).run_op

      self.evaluate(distribution.initialize())
      if not context.executing_eagerly():
        with self.cached_session() as sess:
          run_step = sess.make_callable(run_step())
      self.evaluate(variables_lib.global_variables_initializer())

      run_step()

      self.evaluate(distribution.finalize())

      def get_expected_variables(optimizer_fn, num_parameter_devices):
        variables_map = {
            "GradientDescent": ["dense/kernel", "dense/bias"],
            "Adagrad": [
                "dense/kernel/Adagrad", "dense/kernel",
                "dense/bias/Adagrad", "dense/bias"
            ]
        }
        variables = variables_map[optimizer_fn().get_name()]
        variables.extend([
            v + "/replica_{}".format(replica)
            for v in variables
            for replica in range(1, num_parameter_devices)
        ])
        return set([v + ":0" for v in variables])

      self.assertEqual(
          get_expected_variables(optimizer_fn,
                                 len(distribution.parameter_devices)),
          set(created_variables))

  @combinations.generate(
      combinations.times(
          combinations.combine(momentum=[0.8, 0.9, 0.99], renorm=[False, True]),
          combinations.times(
              combinations.distributions_and_v1_optimizers(),
              combinations.combine(
                  mode=["graph", "eager"],
                  # TODO(isaprykin):  Allow False here.  Currently subsequent
                  # towers will re-execute UPDATE_OPS of previous towers.
                  update_ops_in_cross_tower_mode=[True])) +
          combinations.combine(
              distribution=[combinations.tpu_strategy],
              optimizer_fn=combinations.optimizers_v1,
              mode=["graph"],
              update_ops_in_cross_tower_mode=[False])))
  def testTrainNetworkWithBatchNorm(self, distribution, optimizer_fn, momentum,
                                    renorm, update_ops_in_cross_tower_mode):
    """Verifies that moving mean updates are reduced across towers."""
    with distribution.scope():
      num_towers = len(distribution.worker_devices)
      model_fn, dataset_fn, batchnorm = batchnorm_example(
          optimizer_fn,
          batch_per_epoch=num_towers,
          momentum=momentum,
          renorm=renorm,
          update_ops_in_tower_mode=not update_ops_in_cross_tower_mode)

      # Make sure prefetching is disabled since that makes the
      # specific input on each device to be non deterministic, and
      # this test relies on specific input being on each device.
      if isinstance(distribution, mirrored_strategy.MirroredStrategy):
        self.assertFalse(distribution._prefetch_on_device)

      def step_fn(ctx, *inputs):
        del ctx  # Unused
        fetches = distribution.unwrap(
            distribution.call_for_each_tower(
                model_fn, *inputs, run_concurrently=batchnorm.built))
        if update_ops_in_cross_tower_mode:
          fetches += ops.get_collection(ops.GraphKeys.UPDATE_OPS)
        return control_flow_ops.group(fetches)

      iterator = distribution.distribute_dataset(
          dataset_fn).make_one_shot_iterator()

      def run_step():
        return distribution.run_steps_on_dataset(
            step_fn, iterator, iterations=1).run_op

      self.evaluate(distribution.initialize())
      if not context.executing_eagerly():
        with self.cached_session() as sess:
          run_step = sess.make_callable(run_step())
      self.evaluate(variables_lib.global_variables_initializer())

      expected_moving_means = [0.] * 8

      def averaged_batch_mean(i):
        # Each batch has shape [16, 8] where the ith element in jth list is
        # (8 * j + i + tower_id * 100). So the batch mean in each tower is
        # (60 + i + tower_id * 100). So here comes its batch mean over all
        # towers:
        return 60. + i + (num_towers - 1.) / 2. * 100.

      for _ in range(10):
        run_step()
        moving_means = self.evaluate(batchnorm.moving_mean)

        # We make sure that the moving_mean is updated as if the sample mean is
        # calculated over all towers.
        for i, expected_moving_mean in enumerate(expected_moving_means):
          expected_moving_means[i] -= ((
              expected_moving_mean - averaged_batch_mean(i)) * (1.0 - momentum))
          self.assertNear(expected_moving_means[i], moving_means[i], 0.0001)

      self.evaluate(distribution.finalize())

  @combinations.generate(
      combinations.times(
          combinations.combine(
              optimizer_fn=[
                  combinations.gradient_descent_optimizer_v1_fn,
                  combinations.gradient_descent_optimizer_v2_fn
              ],
              loss_reduction=[
                  losses_impl.Reduction.SUM, losses_impl.Reduction.MEAN,
                  losses_impl.Reduction.SUM_OVER_BATCH_SIZE,
                  losses_impl.Reduction.SUM_OVER_NONZERO_WEIGHTS
              ]),
          combinations.times(
              combinations.combine(
                  distribution=[
                      combinations.one_device_strategy,
                      combinations.mirrored_strategy_with_gpu_and_cpu,
                      combinations.mirrored_strategy_with_two_gpus
                  ]),
              combinations.combine(
                  mode=["graph"], use_callable_loss=[True, False]) +
              combinations.combine(mode=["eager"], use_callable_loss=[True])) +
          combinations.combine(
              distribution=[combinations.tpu_strategy],
              mode=["graph"],
              use_callable_loss=[True, False])))
  def testMeanVsSum(self, distribution, optimizer_fn, loss_reduction,
                    use_callable_loss):
    with distribution.scope():
      all_vars = []

      def model_fn(x, y):

        def loss_fn():
          # Use fixed initialization to make the steps deterministic.
          w = variable_scope.get_variable("w", initializer=[[2.]])
          all_vars.append(w)
          predict = math_ops.matmul(x, w)
          return losses_impl.mean_squared_error(
              y, predict, reduction=loss_reduction)

        optimizer = optimizer_fn()  # GradientDescent with 0.2 learning rate

        if use_callable_loss:
          return optimizer.minimize(loss_fn)
        else:
          return optimizer.minimize(loss_fn())

      def dataset_fn():
        features = dataset_ops.Dataset.from_tensors([[2.], [7.]])
        labels = dataset_ops.Dataset.from_tensors([[6.], [21.]])
        return dataset_ops.Dataset.zip((features, labels)).repeat()

      def step_fn(ctx, x, y):
        del ctx  # Unused
        return distribution.group(
            distribution.call_for_each_tower(
                model_fn, x, y, run_concurrently=False))

      iterator = distribution.distribute_dataset(
          dataset_fn).make_one_shot_iterator()

      def run_step():
        return distribution.run_steps_on_dataset(
            step_fn, iterator, iterations=1).run_op

      self.evaluate(distribution.initialize())
      if not context.executing_eagerly():
        with self.cached_session() as sess:
          run_step = sess.make_callable(run_step())
      self.evaluate(variables_lib.global_variables_initializer())

      run_step()

      v = all_vars[0]
      self.assertTrue(all([v is vi for vi in all_vars[1:]]))
      weight = numpy.squeeze(self.evaluate(v))
      # Our model is:
      #   predict = x * w
      #   loss = (predict - y)^2
      #   dloss/dpredict = 2*(predict - y)
      #   dloss/dw = 2 * x^T @ (predict - y)
      # For our batch size of 2, assuming sum loss reduction:
      #   x = [2, 7]
      #   y = [6, 21]
      #   w_initial = 2
      #   predict = [4, 14]
      #   predict - y = [-2, -7]
      #   dloss/dw = 2 <[2, 7], [-2, -7]> = - 2(4 + 49) = -106
      # So unreplicated the update to w with lr=0.2 is -0.2 * -106 = 21.2
      # with sum loss reduction, or 10.6 with mean.
      if loss_reduction == losses_impl.Reduction.SUM:
        # Note that the "distribution.num_towers" factor will go away once
        # we split the input across towers, instead of pulling a complete
        # batch of input per tower.
        self.assertNear(weight, 2 + 21.2 * distribution.num_towers, 0.0001)
      else:
        # One of the mean loss reductions.
        self.assertNear(weight, 2 + 10.6, 0.0001)

      self.evaluate(distribution.finalize())

  @combinations.generate(
      combinations.times(
          combinations.distributions_and_v1_optimizers(),
          combinations.combine(mode=["graph", "eager"]),
          combinations.combine(is_tpu=[False])) +
      combinations.combine(
          distribution=[combinations.tpu_strategy],
          optimizer_fn=combinations.optimizers_v1,
          mode=["graph"],
          is_tpu=[True]))
  def testRunStepsWithOutputContext(self, distribution, optimizer_fn, is_tpu):
    with distribution.scope():
      def dataset_fn():
        dataset = dataset_ops.Dataset.from_tensors([[1.]]).repeat()
        # TODO(priyag): batch with drop_remainder=True causes shapes to be
        # fully defined for TPU. Remove this when XLA supports dynamic shapes.
        return dataset.batch(batch_size=1, drop_remainder=True)

      optimizer = optimizer_fn()
      layer = core.Dense(1, use_bias=True)

      key1 = "foo"
      value1 = "bar"

      def model_fn(output_context, x):
        """A very simple model written by the user."""
        def loss_fn():
          y = array_ops.reshape(layer(x), []) - constant_op.constant(1.)
          return y * y

        train_op = optimizer.minimize(loss_fn)
        loss = loss_fn()
        output_context.set_last_step_output(
            name="tower_loss_agg",
            output=loss,
            aggregation=variables_lib.VariableAggregation.MEAN)
        output_context.set_non_tensor_output(key1, value1)
        return (train_op, loss)

      def step_fn(output_context, *inputs):
        (train_op, loss) = distribution.call_for_each_tower(
            model_fn, output_context, *inputs, run_concurrently=False)
        output_context.set_last_step_output(
            name="cross_tower_loss_agg",
            output=loss,
            aggregation=variables_lib.VariableAggregation.MEAN)
        output_context.set_last_step_output(
            name="cross_tower_loss_noagg",
            output=loss)
        return distribution.group(train_op)

      iterator = distribution.distribute_dataset(
          dataset_fn).make_one_shot_iterator()

      def run_step():
        initial_loss = lambda: constant_op.constant(1e7)
        # Initial values corresponding to aggregated losses are just single
        # tensors. But for non aggregated losses, we need to have initial
        # values that are of the same structure as non reduced losses. In
        # MirroredStrategy, this will be a list of losses, in TPUStrategy
        # it will be single tensor. Using `broadcast` followed by `unwrap`
        # gives us the desired initial value structure.
        initial_loop_values = {
            "tower_loss_agg": initial_loss(),
            "cross_tower_loss_agg": initial_loss(),
            "cross_tower_loss_noagg":
            distribution.unwrap(distribution.broadcast(initial_loss()))
        }
        ctx = distribution.run_steps_on_dataset(
            step_fn, iterator, iterations=2,
            initial_loop_values=initial_loop_values)

        self.assertEqual({key1: [value1]}, ctx.non_tensor_outputs)
        self._verify_loss_output(
            initial_loss(),
            loss_output=ctx.last_step_outputs["tower_loss_agg"],
            aggregated=True, distribution=distribution)
        self._verify_loss_output(
            initial_loss(),
            loss_output=ctx.last_step_outputs["cross_tower_loss_agg"],
            aggregated=True, distribution=distribution)
        self._verify_loss_output(
            initial_loss(),
            loss_output=ctx.last_step_outputs["cross_tower_loss_noagg"],
            aggregated=False, distribution=distribution)
        return (ctx.run_op, ctx.last_step_outputs["tower_loss_agg"])

      self.evaluate(distribution.initialize())
      if not context.executing_eagerly():
        with self.cached_session() as sess:
          run_step = sess.make_callable(run_step())
      self.evaluate(variables_lib.global_variables_initializer())

      weights, biases, losses = [], [], []
      for _ in range(5):
        _, loss = run_step()
        losses.append(loss)
        weights.append(self.evaluate(layer.kernel))
        biases.append(self.evaluate(layer.bias))

      self.evaluate(distribution.finalize())

      loss_is_not_increasing = all(y <= x for x, y in zip(losses, losses[1:]))
      self.assertTrue(loss_is_not_increasing)

      error = abs(
          numpy.add(numpy.squeeze(weights), numpy.squeeze(biases)) - 1)
      error_is_not_increasing = all(y <= x for x, y in zip(error, error[1:]))
      self.assertTrue(error_is_not_increasing)

  def _verify_loss_output(self, initial_loss, loss_output, aggregated,
                          distribution):
    if not aggregated:
      self.assertEqual(distribution.num_towers,
                       len(distribution.unwrap(loss_output)))
      loss_output = distribution.reduce(
          aggregation=variables_lib.VariableAggregation.MEAN,
          value=loss_output, destinations="/device:CPU:0")

    unwrapped_output = distribution.unwrap(loss_output)
    self.assertEqual(1, len(unwrapped_output))
    loss_tensor = unwrapped_output[0]
    self.assertEqual(initial_loss.dtype, loss_tensor.dtype)
    self.assertEqual(initial_loss.shape, loss_tensor.shape)

if __name__ == "__main__":
  test.main()