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 16 // See docs in ../ops/nn_ops.cc. 17 18 #define EIGEN_USE_THREADS 19 20 #include "third_party/eigen3/unsupported/Eigen/CXX11/Tensor" 21 #include "tensorflow/core/framework/numeric_op.h" 22 #include "tensorflow/core/framework/op_kernel.h" 23 #include "tensorflow/core/framework/tensor.h" 24 #include "tensorflow/core/framework/tensor_shape.h" 25 #include "tensorflow/core/kernels/ops_util.h" 26 #include "tensorflow/core/kernels/pooling_ops_common.h" 27 #include "tensorflow/core/lib/core/errors.h" 28 #include "tensorflow/core/platform/logging.h" 29 #include "tensorflow/core/util/padding.h" 30 #include "tensorflow/core/util/tensor_format.h" 31 32 namespace tensorflow { 33 34 typedef Eigen::ThreadPoolDevice CPUDevice; 35 36 template <typename Device, typename T> 37 class QuantizedAvgPoolingOp : public OpKernel { 38 public: QuantizedAvgPoolingOp(OpKernelConstruction * context)39 explicit QuantizedAvgPoolingOp(OpKernelConstruction* context) 40 : OpKernel(context) { 41 OP_REQUIRES_OK(context, context->GetAttr("ksize", &ksize_)); 42 OP_REQUIRES(context, ksize_.size() == 4, 43 errors::InvalidArgument("Sliding window ksize field must " 44 "specify 4 dimensions")); 45 OP_REQUIRES_OK(context, context->GetAttr("strides", &stride_)); 46 OP_REQUIRES(context, stride_.size() == 4, 47 errors::InvalidArgument("Sliding window strides field must " 48 "specify 4 dimensions")); 49 OP_REQUIRES_OK(context, context->GetAttr("padding", &padding_)); 50 OP_REQUIRES(context, ksize_[0] == 1 && stride_[0] == 1, 51 errors::Unimplemented( 52 "Pooling is not yet supported on the batch dimension.")); 53 } 54 Compute(OpKernelContext * context)55 void Compute(OpKernelContext* context) override { 56 const Tensor& tensor_in = context->input(0); 57 PoolParameters params{context, ksize_, stride_, 58 padding_, FORMAT_NHWC, tensor_in.shape()}; 59 if (!context->status().ok()) { 60 return; 61 } 62 63 const float min_input = context->input(1).flat<float>()(0); 64 const float max_input = context->input(2).flat<float>()(0); 65 66 OP_REQUIRES(context, params.depth_window == 1, 67 errors::Unimplemented("Non-spatial pooling is not " 68 "yet supported. Volunteers? :)")); 69 70 OP_REQUIRES(context, tensor_in.dims() == 4, 71 errors::InvalidArgument("tensor_in must be 4-dimensional")); 72 73 Tensor* output = nullptr; 74 OP_REQUIRES_OK(context, context->allocate_output( 75 0, params.forward_output_shape(), &output)); 76 const int32 highest = static_cast<int32>(Eigen::NumTraits<T>::highest()); 77 const int32 lowest = static_cast<int32>(Eigen::NumTraits<T>::lowest()); 78 79 // TODO(vrv): Switch this to the Eigen::Tensor version of 80 // SpatialAvgPooling once that version is running quickly. 81 Tensor int32_output(DT_INT32, params.forward_output_shape()); 82 // Cast input to int32 tensor and call SpatialAvgPool. 83 Tensor int32_input(DT_INT32, tensor_in.shape()); 84 int32_input.flat<int32>() = tensor_in.flat<T>().template cast<int32>(); 85 SpatialAvgPool<Device, int32>(context, &int32_output, int32_input, params, 86 padding_); 87 88 // Clamp the int32 output back into quantized space. 89 output->flat<T>() = int32_output.flat<int32>() 90 .cwiseMax(lowest) 91 .cwiseMin(highest) 92 .template cast<T>(); 93 94 Tensor* output_min = nullptr; 95 OP_REQUIRES_OK(context, context->allocate_output(1, {}, &output_min)); 96 output_min->flat<float>()(0) = min_input; 97 Tensor* output_max = nullptr; 98 OP_REQUIRES_OK(context, context->allocate_output(2, {}, &output_max)); 99 output_max->flat<float>()(0) = max_input; 100 } 101 102 private: 103 std::vector<int32> ksize_; 104 std::vector<int32> stride_; 105 Padding padding_; 106 }; 107 108 template <typename Device, typename T> 109 class QuantizedMaxPoolingOp : public MaxPoolingOp<Device, T> { 110 public: QuantizedMaxPoolingOp(OpKernelConstruction * context)111 explicit QuantizedMaxPoolingOp(OpKernelConstruction* context) 112 : MaxPoolingOp<Device, T>(context) {} 113 Compute(OpKernelContext * context)114 void Compute(OpKernelContext* context) override { 115 const float min_input = context->input(1).flat<float>()(0); 116 const float max_input = context->input(2).flat<float>()(0); 117 MaxPoolingOp<Device, T>::Compute(context); 118 Tensor* output_min = nullptr; 119 OP_REQUIRES_OK(context, context->allocate_output(1, {}, &output_min)); 120 output_min->flat<float>()(0) = min_input; 121 Tensor* output_max = nullptr; 122 OP_REQUIRES_OK(context, context->allocate_output(2, {}, &output_max)); 123 output_max->flat<float>()(0) = max_input; 124 } 125 }; 126 127 REGISTER_KERNEL_BUILDER( 128 Name("QuantizedAvgPool").Device(DEVICE_CPU).TypeConstraint<quint8>("T"), 129 QuantizedAvgPoolingOp<CPUDevice, quint8>); 130 131 REGISTER_KERNEL_BUILDER( 132 Name("QuantizedMaxPool").Device(DEVICE_CPU).TypeConstraint<quint8>("T"), 133 QuantizedMaxPoolingOp<CPUDevice, quint8>); 134 135 } // namespace tensorflow 136