Coverage for /pythoncovmergedfiles/medio/medio/usr/local/lib/python3.8/site-packages/tensorflow/python/distribute/numpy_dataset.py: 33%

1# Copyright 2019 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"""Code for creating a dataset out of a NumPy array.""" 

16 

17import numpy as np 

18 

19from tensorflow.python.data.ops import dataset_ops 

20from tensorflow.python.eager import context 

21from tensorflow.python.framework import dtypes 

22from tensorflow.python.framework import ops 

23from tensorflow.python.ops import array_ops 

24from tensorflow.python.ops import variable_scope 

25from tensorflow.python.ops import variable_v1 

26from tensorflow.python.util import nest 

27 

28 

29def init_var_from_numpy(input_var, numpy_input, session): 

30 """Initialize `input_var` to `numpy_input` using `session` in graph mode.""" 

31 with ops.init_scope(): 

32 if context.executing_eagerly(): 

33 input_var.assign(numpy_input) 

34 return 

35 

36 assert session is not None 

37 session.run(input_var.initializer) 

38 

39 start_placeholder = array_ops.placeholder(dtypes.int64, ()) 

40 end_placeholder = array_ops.placeholder(dtypes.int64, ()) 

41 slice_placeholder = array_ops.placeholder(input_var.dtype) 

42 assign_slice_op = input_var[start_placeholder:end_placeholder].assign( 

43 slice_placeholder) 

44 

45 # If each batch element is > 64 MB, then we copy each batch element 

46 # individually. Otherwise, the slices will be < 128 MB. There might be 

47 # padding which might mean that the slices are 128 MB even if the size of 

48 # the tensor allocated is less than 128 MB. This formula gives slices with 

49 # size: ceil(64 MB / byte size per batch element) bytes. Using ceil() 

50 # guarantees we get a number >= 1. 

51 

52 # Calculate the size of each batch element. 

53 byte_size_per_batch_element = ( 

54 np.prod(numpy_input.shape[1:]) * input_var.dtype.size) 

55 

56 # Calculate number of elements we want to copy per slice. 

57 batch_size_per_slice = int( 

58 np.ceil((64 << 20) / byte_size_per_batch_element)) 

59 

60 # Copy slices of the above size starting at 0, except the last slice will be 

61 # smaller. 

62 start = 0 

63 limit = numpy_input.shape[0] 

64 while start < limit: 

65 end = min(start + batch_size_per_slice, limit) 

66 session.run(assign_slice_op, feed_dict={ 

67 start_placeholder: start, 

68 end_placeholder: end, 

69 slice_placeholder: numpy_input[start:end]}) 

70 start = end 

71 

72 

73def one_host_numpy_dataset(numpy_input, colocate_with, session): 

74 """Create a dataset on `colocate_with` from `numpy_input`.""" 

75 

76 def create_colocated_variable(next_creator, **kwargs): 

77 kwargs["colocate_with"] = colocate_with 

78 return next_creator(**kwargs) 

79 

80 numpy_flat = nest.flatten(numpy_input) 

81 with variable_scope.variable_creator_scope(create_colocated_variable): 

82 vars_flat = tuple(variable_v1.VariableV1(array_ops.zeros(i.shape, i.dtype), 

83 trainable=False) 

84 for i in numpy_flat) 

85 for v, i in zip(vars_flat, numpy_flat): 

86 init_var_from_numpy(v, i, session) 

87 vars_nested = nest.pack_sequence_as(numpy_input, vars_flat) 

88 return dataset_ops.Dataset.from_tensor_slices(vars_nested) 

89 

90 

91class SingleDevice(object): 

92 """Used with `colocate_with` to create a non-mirrored variable.""" 

93 

94 def __init__(self, device): 

95 self.device = device