1 /* Copyright 2016 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 16 #include "tensorflow/core/framework/op.h" 17 #include "tensorflow/core/framework/shape_inference.h" 18 19 namespace tensorflow { 20 21 using shape_inference::DimensionHandle; 22 using shape_inference::InferenceContext; 23 using shape_inference::ShapeHandle; 24 25 // CTC is Connectionist Temporal Classification. See util/ctc/ for details. 26 27 REGISTER_OP("CTCLoss") 28 .Input("inputs: float") 29 .Input("labels_indices: int64") 30 .Input("labels_values: int32") 31 .Input("sequence_length: int32") 32 .Attr("preprocess_collapse_repeated: bool = false") 33 .Attr("ctc_merge_repeated: bool = true") 34 .Attr("ignore_longer_outputs_than_inputs: bool = false") 35 .Output("loss: float") 36 .Output("gradient: float") __anon78c05cd60102(InferenceContext* c) 37 .SetShapeFn([](InferenceContext* c) { 38 ShapeHandle inputs; 39 ShapeHandle labels_indices; 40 ShapeHandle labels_values; 41 ShapeHandle sequence_length; 42 43 TF_RETURN_IF_ERROR(c->WithRank(c->input(0), 3, &inputs)); 44 TF_RETURN_IF_ERROR(c->WithRank(c->input(1), 2, &labels_indices)); 45 TF_RETURN_IF_ERROR(c->WithRank(c->input(2), 1, &labels_values)); 46 TF_RETURN_IF_ERROR(c->WithRank(c->input(3), 1, &sequence_length)); 47 48 DimensionHandle unused; 49 TF_RETURN_IF_ERROR(c->Merge(c->Dim(labels_indices, 0), 50 c->Dim(labels_values, 0), &unused)); 51 52 // Get batch size from inputs and sequence_length, and update inputs 53 // with the merged batch_size since it is returned. 54 DimensionHandle batch_size; 55 TF_RETURN_IF_ERROR( 56 c->Merge(c->Dim(inputs, 1), c->Dim(sequence_length, 0), &batch_size)); 57 TF_RETURN_IF_ERROR(c->ReplaceDim(inputs, 1, batch_size, &inputs)); 58 59 c->set_output(0, c->Vector(batch_size)); 60 c->set_output(1, inputs); 61 return Status::OK(); 62 }); 63 64 REGISTER_OP("CTCGreedyDecoder") 65 .Input("inputs: float") 66 .Input("sequence_length: int32") 67 .Attr("merge_repeated: bool = false") 68 .Output("decoded_indices: int64") 69 .Output("decoded_values: int64") 70 .Output("decoded_shape: int64") 71 .Output("log_probability: float") __anon78c05cd60202(InferenceContext* c) 72 .SetShapeFn([](InferenceContext* c) { 73 ShapeHandle inputs; 74 ShapeHandle sequence_length; 75 76 TF_RETURN_IF_ERROR(c->WithRank(c->input(0), 3, &inputs)); 77 TF_RETURN_IF_ERROR(c->WithRank(c->input(1), 1, &sequence_length)); 78 79 // Get batch size from inputs and sequence_length. 80 DimensionHandle batch_size; 81 TF_RETURN_IF_ERROR( 82 c->Merge(c->Dim(inputs, 1), c->Dim(sequence_length, 0), &batch_size)); 83 84 DimensionHandle total_decoded_outputs = c->UnknownDim(); 85 c->set_output(0, c->Matrix(total_decoded_outputs, 2)); 86 c->set_output(1, c->Vector(total_decoded_outputs)); 87 c->set_output(2, c->Vector(2)); 88 c->set_output(3, c->Matrix(batch_size, 1)); 89 return Status::OK(); 90 }); 91 92 REGISTER_OP("CTCBeamSearchDecoder") 93 .Input("inputs: float") 94 .Input("sequence_length: int32") 95 .Attr("beam_width: int >= 1") 96 .Attr("top_paths: int >= 1") 97 .Attr("merge_repeated: bool = true") 98 .Output("decoded_indices: top_paths * int64") 99 .Output("decoded_values: top_paths * int64") 100 .Output("decoded_shape: top_paths * int64") 101 .Output("log_probability: float") __anon78c05cd60302(InferenceContext* c) 102 .SetShapeFn([](InferenceContext* c) { 103 ShapeHandle inputs; 104 ShapeHandle sequence_length; 105 106 TF_RETURN_IF_ERROR(c->WithRank(c->input(0), 3, &inputs)); 107 TF_RETURN_IF_ERROR(c->WithRank(c->input(1), 1, &sequence_length)); 108 109 // Get batch size from inputs and sequence_length. 110 DimensionHandle batch_size; 111 TF_RETURN_IF_ERROR( 112 c->Merge(c->Dim(inputs, 1), c->Dim(sequence_length, 0), &batch_size)); 113 114 int32 top_paths; 115 TF_RETURN_IF_ERROR(c->GetAttr("top_paths", &top_paths)); 116 117 // Outputs. 118 int out_idx = 0; 119 for (int i = 0; i < top_paths; ++i) { // decoded_indices 120 c->set_output(out_idx++, c->Matrix(InferenceContext::kUnknownDim, 2)); 121 } 122 for (int i = 0; i < top_paths; ++i) { // decoded_values 123 c->set_output(out_idx++, c->Vector(InferenceContext::kUnknownDim)); 124 } 125 ShapeHandle shape_v = c->Vector(2); 126 for (int i = 0; i < top_paths; ++i) { // decoded_shape 127 c->set_output(out_idx++, shape_v); 128 } 129 c->set_output(out_idx++, c->Matrix(batch_size, top_paths)); 130 return Status::OK(); 131 }); 132 133 } // namespace tensorflow 134