Coverage for /pythoncovmergedfiles/medio/medio/usr/local/lib/python3.8/site-packages/tensorflow/python/keras/utils/dataset_creator.py: 44%

1# Copyright 2021 The TensorFlow Authors. All Rights Reserved. 

2# 

3# Licensed under the Apache License, Version 2.0 (the "License"); 

4# you may not use this file except in compliance with the License. 

5# You may obtain a copy of the License at 

6# 

7# http://www.apache.org/licenses/LICENSE-2.0 

8# 

9# Unless required by applicable law or agreed to in writing, software 

10# distributed under the License is distributed on an "AS IS" BASIS, 

11# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. 

12# See the License for the specific language governing permissions and 

13# limitations under the License. 

14# ============================================================================== 

15# pylint: disable=g-classes-have-attributes 

16"""Input dataset creator for `model.fit`.""" 

17 

18from tensorflow.python.distribute import distribute_lib 

19from tensorflow.python.types import data as data_types 

20from tensorflow.python.util.tf_export import keras_export 

21 

22 

23@keras_export('keras.utils.experimental.DatasetCreator', v1=[]) 

24class DatasetCreator(object): 

25 """Object that returns a `tf.data.Dataset` upon invoking. 

26 

27 `tf.keras.utils.experimental.DatasetCreator` is designated as a supported type 

28 for `x`, or the input, in `tf.keras.Model.fit`. Pass an instance of this class 

29 to `fit` when using a callable (with a `input_context` argument) that returns 

30 a `tf.data.Dataset`. 

31 

32 ```python 

33 model = tf.keras.Sequential([tf.keras.layers.Dense(10)]) 

34 model.compile(tf.keras.optimizers.SGD(), loss="mse") 

35 

36 def dataset_fn(input_context): 

37 global_batch_size = 64 

38 batch_size = input_context.get_per_replica_batch_size(global_batch_size) 

39 dataset = tf.data.Dataset.from_tensors(([1.], [1.])).repeat() 

40 dataset = dataset.shard( 

41 input_context.num_input_pipelines, input_context.input_pipeline_id) 

42 dataset = dataset.batch(batch_size) 

43 dataset = dataset.prefetch(2) 

44 return dataset 

45 

46 input_options = tf.distribute.InputOptions( 

47 experimental_fetch_to_device=True, 

48 experimental_per_replica_buffer_size=2) 

49 model.fit(tf.keras.utils.experimental.DatasetCreator( 

50 dataset_fn, input_options=input_options), epochs=10, steps_per_epoch=10) 

51 ``` 

52 

53 `Model.fit` usage with `DatasetCreator` is intended to work across all 

54 `tf.distribute.Strategy`s, as long as `Strategy.scope` is used at model 

55 creation: 

56 

57 ```python 

58 strategy = tf.distribute.experimental.ParameterServerStrategy( 

59 cluster_resolver) 

60 with strategy.scope(): 

61 model = tf.keras.Sequential([tf.keras.layers.Dense(10)]) 

62 model.compile(tf.keras.optimizers.SGD(), loss="mse") 

63 ... 

64 ``` 

65 

66 Note: When using `DatasetCreator`, `steps_per_epoch` argument in `Model.fit` 

67 must be provided as the cardinality of such input cannot be inferred. 

68 

69 Args: 

70 dataset_fn: A callable that takes a single argument of type 

71 `tf.distribute.InputContext`, which is used for batch size calculation and 

72 cross-worker input pipeline sharding (if neither is needed, the 

73 `InputContext` parameter can be ignored in the `dataset_fn`), and returns 

74 a `tf.data.Dataset`. 

75 input_options: Optional `tf.distribute.InputOptions`, used for specific 

76 options when used with distribution, for example, whether to prefetch 

77 dataset elements to accelerator device memory or host device memory, and 

78 prefetch buffer size in the replica device memory. No effect if not used 

79 with distributed training. See `tf.distribute.InputOptions` for more 

80 information. 

81 """ 

82 

83 def __init__(self, dataset_fn, input_options=None): 

84 if not callable(dataset_fn): 

85 raise TypeError('`dataset_fn` for `DatasetCreator` must be a `callable`.') 

86 if input_options and (not isinstance(input_options, 

87 distribute_lib.InputOptions)): 

88 raise TypeError('`input_options` for `DatasetCreator` must be a ' 

89 '`tf.distribute.InputOptions`.') 

90 

91 self.dataset_fn = dataset_fn 

92 self.input_options = input_options 

93 

94 def __call__(self, *args, **kwargs): 

95 # When a `DatasetCreator` is invoked, it forwards args/kwargs straight to 

96 # the callable. 

97 dataset = self.dataset_fn(*args, **kwargs) 

98 if not isinstance(dataset, data_types.DatasetV2): 

99 raise TypeError('The `callable` provided to `DatasetCreator` must return ' 

100 'a Dataset.') 

101 return dataset