sdks/python/apache_beam/ml/inference/tensorflow_inference_test.py

#
# Licensed to the Apache Software Foundation (ASF) under one or more
# contributor license agreements.  See the NOTICE file distributed with
# this work for additional information regarding copyright ownership.
# The ASF licenses this file to You 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.
#

# pytype: skip-file

import os
import shutil
import tempfile
import unittest
import uuid
from collections.abc import Iterable
from collections.abc import Sequence
from typing import Any
from typing import Optional
from typing import Union

import numpy
import pytest

import apache_beam as beam
from apache_beam.ml.inference import utils
from apache_beam.ml.inference.base import KeyedModelHandler
from apache_beam.ml.inference.base import PredictionResult
from apache_beam.ml.inference.base import RunInference
from apache_beam.testing.test_pipeline import TestPipeline
from apache_beam.testing.util import assert_that
from apache_beam.testing.util import equal_to

# pylint: disable=ungrouped-imports
try:
  import tensorflow as tf

  from apache_beam.ml.inference import tensorflow_inference
  from apache_beam.ml.inference.sklearn_inference_test import _compare_prediction_result
  from apache_beam.ml.inference.tensorflow_inference import TFModelHandlerNumpy
  from apache_beam.ml.inference.tensorflow_inference import TFModelHandlerTensor
except ImportError:
  raise unittest.SkipTest(
      'Tensorflow dependencies are not installed. ' +
      'Make sure you have both tensorflow and tensorflow_hub installed.')


class FakeTFNumpyModel:
  def predict(self, input: numpy.ndarray):
    return numpy.multiply(input, 10)


class FakeTFTensorModel:
  def predict(self, input: tf.Tensor, add=False):
    if add:
      return tf.math.add(tf.math.multiply(input, 10), 10)
    return tf.math.multiply(input, 10)


def _create_mult2_model():
  inputs = tf.keras.Input(shape=(3, ))
  outputs = tf.keras.layers.Lambda(lambda x: x * 2, dtype='float32')(inputs)
  return tf.keras.Model(inputs=inputs, outputs=outputs)


def _compare_tensor_prediction_result(x, y):
  return tf.reduce_all(tf.math.equal(x.inference, y.inference))


def fake_inference_fn(
    model: tf.Module,
    batch: Union[Sequence[numpy.ndarray], Sequence[tf.Tensor]],
    inference_args: dict[str, Any],
    model_id: Optional[str] = None) -> Iterable[PredictionResult]:
  predictions = model.predict(batch, **inference_args)
  return utils._convert_to_result(batch, predictions, model_id)


@pytest.mark.uses_tf
class TFRunInferenceTest(unittest.TestCase):
  def setUp(self):
    self.tmpdir = tempfile.mkdtemp()

  def tearDown(self):
    shutil.rmtree(self.tmpdir)

  def test_predict_numpy(self):
    fake_model = FakeTFNumpyModel()
    inference_runner = TFModelHandlerNumpy(
        model_uri='unused', inference_fn=fake_inference_fn)
    batched_examples = [numpy.array([1]), numpy.array([10]), numpy.array([100])]
    expected_predictions = [
        PredictionResult(numpy.array([1]), 10),
        PredictionResult(numpy.array([10]), 100),
        PredictionResult(numpy.array([100]), 1000)
    ]
    inferences = inference_runner.run_inference(batched_examples, fake_model)
    for actual, expected in zip(inferences, expected_predictions):
      self.assertTrue(_compare_prediction_result(actual, expected))

  def test_predict_tensor(self):
    fake_model = FakeTFTensorModel()
    inference_runner = TFModelHandlerTensor(
        model_uri='unused', inference_fn=fake_inference_fn)
    batched_examples = [
        tf.convert_to_tensor(numpy.array([1])),
        tf.convert_to_tensor(numpy.array([10])),
        tf.convert_to_tensor(numpy.array([100])),
    ]
    expected_predictions = [
        PredictionResult(ex, pred) for ex, pred in zip(
            batched_examples,
            [tf.math.multiply(n, 10) for n in batched_examples])
    ]

    inferences = inference_runner.run_inference(batched_examples, fake_model)
    for actual, expected in zip(inferences, expected_predictions):
      self.assertTrue(_compare_tensor_prediction_result(actual, expected))

  def test_predict_tensor_with_batch_size(self):
    model = _create_mult2_model()
    model_path = os.path.join(self.tmpdir, f'mult2_{uuid.uuid4()}.keras')
    tf.keras.models.save_model(model, model_path)
    # FnApiRunner guarantees large batches, which this pipeline assumes
    with TestPipeline('FnApiRunner') as pipeline:

      def fake_batching_inference_fn(
          model: tf.Module,
          batch: Union[Sequence[numpy.ndarray], Sequence[tf.Tensor]],
          inference_args: dict[str, Any],
          model_id: Optional[str] = None) -> Iterable[PredictionResult]:
        if len(batch) != 2:
          raise Exception(
              f'Expected batch of size 2, received batch of size {len(batch)}')
        batch = tf.stack(batch, axis=0)
        predictions = model(batch)
        return utils._convert_to_result(batch, predictions, model_id)

      model_handler = TFModelHandlerTensor(
          model_uri=model_path,
          inference_fn=fake_batching_inference_fn,
          load_model_args={'safe_mode': False},
          min_batch_size=2,
          max_batch_size=2)
      examples = [
          tf.convert_to_tensor(numpy.array([1.1, 2.2, 3.3], dtype='float32')),
          tf.convert_to_tensor(
              numpy.array([10.1, 20.2, 30.3], dtype='float32')),
          tf.convert_to_tensor(
              numpy.array([100.1, 200.2, 300.3], dtype='float32')),
          tf.convert_to_tensor(
              numpy.array([200.1, 300.2, 400.3], dtype='float32')),
      ]
      expected_predictions = [
          PredictionResult(ex, pred) for ex, pred in zip(
              examples, [tf.math.multiply(n, 2) for n in examples])
      ]

      pcoll = pipeline | 'start' >> beam.Create(examples)
      predictions = pcoll | RunInference(model_handler)
      assert_that(
          predictions,
          equal_to(
              expected_predictions,
              equals_fn=_compare_tensor_prediction_result))

  def test_predict_tensor_with_large_model(self):
    model = _create_mult2_model()
    model_path = os.path.join(self.tmpdir, f'mult2_{uuid.uuid4()}.keras')
    tf.keras.models.save_model(model, model_path)
    with TestPipeline() as pipeline:

      def fake_batching_inference_fn(
          model: tf.Module,
          batch: Union[Sequence[numpy.ndarray], Sequence[tf.Tensor]],
          inference_args: dict[str, Any],
          model_id: Optional[str] = None) -> Iterable[PredictionResult]:
        multi_process_shared_loaded = "multi_process_shared" in str(type(model))
        if not multi_process_shared_loaded:
          raise Exception(
              f'Loaded model of type {type(model)}, was ' +
              'expecting multi_process_shared_model')
        batch = tf.stack(batch, axis=0)
        predictions = model(batch)
        return utils._convert_to_result(batch, predictions, model_id)

      model_handler = TFModelHandlerTensor(
          model_uri=model_path,
          inference_fn=fake_batching_inference_fn,
          load_model_args={'safe_mode': False},
          large_model=True)
      examples = [
          tf.convert_to_tensor(numpy.array([1.1, 2.2, 3.3], dtype='float32')),
          tf.convert_to_tensor(
              numpy.array([10.1, 20.2, 30.3], dtype='float32')),
          tf.convert_to_tensor(
              numpy.array([100.1, 200.2, 300.3], dtype='float32')),
          tf.convert_to_tensor(
              numpy.array([200.1, 300.2, 400.3], dtype='float32')),
      ]
      expected_predictions = [
          PredictionResult(ex, pred) for ex, pred in zip(
              examples, [tf.math.multiply(n, 2) for n in examples])
      ]

      pcoll = pipeline | 'start' >> beam.Create(examples)
      predictions = pcoll | RunInference(model_handler)
      assert_that(
          predictions,
          equal_to(
              expected_predictions,
              equals_fn=_compare_tensor_prediction_result))

  def test_predict_numpy_with_batch_size(self):
    model = _create_mult2_model()
    model_path = os.path.join(self.tmpdir, f'mult2_{uuid.uuid4()}.keras')
    tf.keras.models.save_model(model, model_path)
    # TODO(https://github.com/apache/beam/issues/34549): This test relies on a
    # runner producing a single bundle or bundles of even size, neither of
    # which prism seems to do here
    with TestPipeline('FnApiRunner') as pipeline:

      def fake_batching_inference_fn(
          model: tf.Module,
          batch: Sequence[numpy.ndarray],
          inference_args: dict[str, Any],
          model_id: Optional[str] = None) -> Iterable[PredictionResult]:
        if len(batch) != 2:
          raise Exception(
              f'Expected batch of size 2, received batch of size {len(batch)}')
        vectorized_batch = numpy.stack(batch, axis=0)
        predictions = model.predict(vectorized_batch, **inference_args)
        return utils._convert_to_result(batch, predictions, model_id)

      model_handler = TFModelHandlerNumpy(
          model_uri=model_path,
          inference_fn=fake_batching_inference_fn,
          load_model_args={'safe_mode': False},
          min_batch_size=2,
          max_batch_size=2)
      examples = [
          numpy.array([1.1, 2.2, 3.3], dtype='float32'),
          numpy.array([10.1, 20.2, 30.3], dtype='float32'),
          numpy.array([100.1, 200.2, 300.3], dtype='float32'),
          numpy.array([200.1, 300.2, 400.3], dtype='float32'),
      ]
      expected_predictions = [
          PredictionResult(ex, pred) for ex, pred in zip(
              examples, [numpy.multiply(n, 2) for n in examples])
      ]

      pcoll = pipeline | 'start' >> beam.Create(examples)
      predictions = pcoll | RunInference(model_handler)
      assert_that(
          predictions,
          equal_to(
              expected_predictions,
              equals_fn=_compare_tensor_prediction_result))

  def test_predict_numpy_with_large_model(self):
    model = _create_mult2_model()
    model_path = os.path.join(self.tmpdir, f'mult2_{uuid.uuid4()}.keras')
    tf.keras.models.save_model(model, model_path)
    with TestPipeline() as pipeline:

      def fake_inference_fn(
          model: tf.Module,
          batch: Sequence[numpy.ndarray],
          inference_args: dict[str, Any],
          model_id: Optional[str] = None) -> Iterable[PredictionResult]:
        multi_process_shared_loaded = "multi_process_shared" in str(type(model))
        if not multi_process_shared_loaded:
          raise Exception(
              f'Loaded model of type {type(model)}, was ' +
              'expecting multi_process_shared_model')
        vectorized_batch = numpy.stack(batch, axis=0)
        predictions = model.predict(vectorized_batch, **inference_args)
        return utils._convert_to_result(batch, predictions, model_id)

      model_handler = TFModelHandlerNumpy(
          model_uri=model_path,
          load_model_args={'safe_mode': False},
          inference_fn=fake_inference_fn,
          large_model=True)
      examples = [
          numpy.array([1.1, 2.2, 3.3], dtype='float32'),
          numpy.array([10.1, 20.2, 30.3], dtype='float32'),
          numpy.array([100.1, 200.2, 300.3], dtype='float32'),
          numpy.array([200.1, 300.2, 400.3], dtype='float32'),
      ]
      expected_predictions = [
          PredictionResult(ex, pred) for ex, pred in zip(
              examples, [numpy.multiply(n, 2) for n in examples])
      ]

      pcoll = pipeline | 'start' >> beam.Create(examples)
      predictions = pcoll | RunInference(model_handler)
      assert_that(
          predictions,
          equal_to(
              expected_predictions,
              equals_fn=_compare_tensor_prediction_result))

  def test_predict_tensor_with_args(self):
    fake_model = FakeTFTensorModel()
    inference_runner = TFModelHandlerTensor(
        model_uri='unused', inference_fn=fake_inference_fn)
    batched_examples = [
        tf.convert_to_tensor(numpy.array([1])),
        tf.convert_to_tensor(numpy.array([10])),
        tf.convert_to_tensor(numpy.array([100])),
    ]
    expected_predictions = [
        PredictionResult(ex, pred) for ex, pred in zip(
            batched_examples, [
                tf.math.add(tf.math.multiply(n, 10), 10)
                for n in batched_examples
            ])
    ]

    inferences = inference_runner.run_inference(
        batched_examples, fake_model, inference_args={"add": True})
    for actual, expected in zip(inferences, expected_predictions):
      self.assertTrue(_compare_tensor_prediction_result(actual, expected))

  def test_predict_keyed_numpy(self):
    fake_model = FakeTFNumpyModel()
    inference_runner = KeyedModelHandler(
        TFModelHandlerNumpy(model_uri='unused', inference_fn=fake_inference_fn))
    batched_examples = [
        ('k1', numpy.array([1], dtype=numpy.int64)),
        ('k2', numpy.array([10], dtype=numpy.int64)),
        ('k3', numpy.array([100], dtype=numpy.int64)),
    ]
    expected_predictions = [
        (ex[0], PredictionResult(ex[1], pred)) for ex, pred in zip(
            batched_examples,
            [numpy.multiply(n[1], 10) for n in batched_examples])
    ]
    inferences = inference_runner.run_inference(batched_examples, fake_model)
    for actual, expected in zip(inferences, expected_predictions):
      self.assertTrue(_compare_prediction_result(actual[1], expected[1]))

  def test_predict_keyed_tensor(self):
    fake_model = FakeTFTensorModel()
    inference_runner = KeyedModelHandler(
        TFModelHandlerTensor(
            model_uri='unused', inference_fn=fake_inference_fn))
    batched_examples = [
        ('k1', tf.convert_to_tensor(numpy.array([1]))),
        ('k2', tf.convert_to_tensor(numpy.array([10]))),
        ('k3', tf.convert_to_tensor(numpy.array([100]))),
    ]
    expected_predictions = [
        (ex[0], PredictionResult(ex[1], pred)) for ex, pred in zip(
            batched_examples,
            [tf.math.multiply(n[1], 10) for n in batched_examples])
    ]
    inferences = inference_runner.run_inference(batched_examples, fake_model)
    for actual, expected in zip(inferences, expected_predictions):
      self.assertTrue(_compare_tensor_prediction_result(actual[1], expected[1]))

  def test_load_model_exception(self):
    with self.assertRaises(ValueError):
      tensorflow_inference._load_model(
          "https://tfhub.dev/google/imagenet/mobilenet_v1_075_192/quantops/classification/3",  # pylint: disable=line-too-long
          None,
          {})


@pytest.mark.uses_tf
class TFRunInferenceTestWithMocks(unittest.TestCase):
  def setUp(self):
    self._load_model = tensorflow_inference._load_model
    tensorflow_inference._load_model = unittest.mock.MagicMock()

  def tearDown(self):
    tensorflow_inference._load_model = self._load_model

  def test_load_model_args(self):
    load_model_args = {compile: False, 'custom_objects': {'optimizer': 1}}
    model_handler = TFModelHandlerNumpy(
        "dummy_model", load_model_args=load_model_args)
    model_handler.load_model()
    tensorflow_inference._load_model.assert_called_with(
        "dummy_model", "", load_model_args)

  def test_load_model_with_args_and_custom_weights(self):
    load_model_args = {compile: False, 'custom_objects': {'optimizer': 1}}
    model_handler = TFModelHandlerNumpy(
        "dummy_model",
        custom_weights="dummy_weights",
        load_model_args=load_model_args)
    model_handler.load_model()
    tensorflow_inference._load_model.assert_called_with(
        "dummy_model", "dummy_weights", load_model_args)

  def test_env_vars_set_correctly_tensor(self):
    handler_with_vars = TFModelHandlerTensor(
        env_vars={'FOO': 'bar'},
        model_uri='unused',
        inference_fn=fake_inference_fn)
    os.environ.pop('FOO', None)
    self.assertFalse('FOO' in os.environ)
    batched_examples = [
        tf.convert_to_tensor(numpy.array([1])),
        tf.convert_to_tensor(numpy.array([10])),
        tf.convert_to_tensor(numpy.array([100])),
    ]
    with TestPipeline() as pipeline:
      _ = (
          pipeline
          | 'start' >> beam.Create(batched_examples)
          | RunInference(handler_with_vars))
      pipeline.run()
      self.assertTrue('FOO' in os.environ)
      self.assertTrue((os.environ['FOO']) == 'bar')

  def test_env_vars_set_correctly_numpy(self):
    handler_with_vars = TFModelHandlerNumpy(
        env_vars={'FOO': 'bar'},
        model_uri="unused",
        inference_fn=fake_inference_fn)
    os.environ.pop('FOO', None)
    self.assertFalse('FOO' in os.environ)
    batched_examples = [numpy.array([1]), numpy.array([10]), numpy.array([100])]
    tensorflow_inference._load_model = unittest.mock.MagicMock()
    with TestPipeline() as pipeline:
      _ = (
          pipeline
          | 'start' >> beam.Create(batched_examples)
          | RunInference(handler_with_vars))
      pipeline.run()
      self.assertTrue('FOO' in os.environ)
      self.assertTrue((os.environ['FOO']) == 'bar')


if __name__ == '__main__':
  unittest.main()