Coverage for /pythoncovmergedfiles/medio/medio/usr/local/lib/python3.8/site-packages/tensorflow/python/ops/cudnn_rnn_grad.py: 50%

1# Copyright 2015 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"""Gradients for CuudnnRNN operators.""" 

16from tensorflow.python.framework import ops 

17from tensorflow.python.ops import gen_cudnn_rnn_ops 

18 

19 

20@ops.RegisterGradient("CudnnRNN") 

21def _cudnn_rnn_backward(op, *grads): 

22 """Gradients for the CudnnRNN op.""" 

23 if not op.get_attr("is_training"): 

24 raise ValueError( 

25 "To use CudnnRNN in gradients, is_training must be set to True.") 

26 return gen_cudnn_rnn_ops.cudnn_rnn_backprop( 

27 input=op.inputs[0], 

28 input_h=op.inputs[1], 

29 input_c=op.inputs[2], 

30 params=op.inputs[3], 

31 output=op.outputs[0], 

32 output_h=op.outputs[1], 

33 output_c=op.outputs[2], 

34 output_backprop=grads[0], 

35 output_h_backprop=grads[1], 

36 output_c_backprop=grads[2], 

37 reserve_space=op.outputs[3], 

38 dropout=op.get_attr("dropout"), 

39 seed=op.get_attr("seed"), 

40 seed2=op.get_attr("seed2"), 

41 rnn_mode=op.get_attr("rnn_mode"), 

42 input_mode=op.get_attr("input_mode"), 

43 direction=op.get_attr("direction")) 

44 

45 

46@ops.RegisterGradient("CudnnRNNV2") 

47def _cudnn_rnn_backward_v2(op, *grad): 

48 if not op.get_attr("is_training"): 

49 raise ValueError( 

50 "To use CudnnRNNV2 in gradients, is_training must be set to True.") 

51 return gen_cudnn_rnn_ops.cudnn_rnn_backprop_v2( 

52 input=op.inputs[0], 

53 input_h=op.inputs[1], 

54 input_c=op.inputs[2], 

55 params=op.inputs[3], 

56 output=op.outputs[0], 

57 output_h=op.outputs[1], 

58 output_c=op.outputs[2], 

59 output_backprop=grad[0], 

60 output_h_backprop=grad[1], 

61 output_c_backprop=grad[2], 

62 reserve_space=op.outputs[3], 

63 host_reserved=op.outputs[4], 

64 dropout=op.get_attr("dropout"), 

65 seed=op.get_attr("seed"), 

66 seed2=op.get_attr("seed2"), 

67 rnn_mode=op.get_attr("rnn_mode"), 

68 input_mode=op.get_attr("input_mode"), 

69 direction=op.get_attr("direction")) 

70 

71 

72@ops.RegisterGradient("CudnnRNNV3") 

73def _cudnn_rnn_backwardv3(op, *grads): 

74 """Gradients for the CudnnRNNV3 op.""" 

75 if not op.get_attr("is_training"): 

76 raise ValueError( 

77 "To use CudnnRNNV3 in gradients, is_training must be set to True.") 

78 return gen_cudnn_rnn_ops.cudnn_rnn_backprop_v3( 

79 input=op.inputs[0], 

80 input_h=op.inputs[1], 

81 input_c=op.inputs[2], 

82 params=op.inputs[3], 

83 sequence_lengths=op.inputs[4], 

84 output=op.outputs[0], 

85 output_h=op.outputs[1], 

86 output_c=op.outputs[2], 

87 output_backprop=grads[0], 

88 output_h_backprop=grads[1], 

89 output_c_backprop=grads[2], 

90 reserve_space=op.outputs[3], 

91 host_reserved=op.outputs[4], 

92 dropout=op.get_attr("dropout"), 

93 seed=op.get_attr("seed"), 

94 seed2=op.get_attr("seed2"), 

95 time_major=op.get_attr("time_major"), 

96 num_proj=op.get_attr("num_proj"), 

97 rnn_mode=op.get_attr("rnn_mode"), 

98 input_mode=op.get_attr("input_mode"), 

99 direction=op.get_attr("direction")) + (None,)