| 50 | } // namespace |
| 51 | |
| 52 | PoolParameters::PoolParameters(OpKernelContext* context, |
| 53 | const std::vector<int32>& ksize, |
| 54 | const std::vector<int32>& stride, |
| 55 | Padding padding, TensorFormat data_format, |
| 56 | const TensorShape& tensor_in_shape) { |
| 57 | // For maxpooling, tensor_in should have 2 spatial dimensions. |
| 58 | // Note: the total number of dimensions could be 4 for NHWC, NCHW, |
| 59 | // or 5 for NCHW_VECT_C. |
| 60 | OP_REQUIRES(context, |
| 61 | GetTensorSpatialDims(tensor_in_shape.dims(), data_format) == 2, |
| 62 | errors::InvalidArgument( |
| 63 | "tensor_in_shape must have 2 spatial dimensions. ", |
| 64 | tensor_in_shape.dims(), " ", data_format)); |
| 65 | |
| 66 | this->data_format = data_format; |
| 67 | depth = GetTensorDim(tensor_in_shape, data_format, 'C') * |
| 68 | (data_format == FORMAT_NCHW_VECT_C ? 4 : 1); |
| 69 | tensor_in_cols = GetTensorDim(tensor_in_shape, data_format, 'W'); |
| 70 | tensor_in_rows = GetTensorDim(tensor_in_shape, data_format, 'H'); |
| 71 | tensor_in_batch = GetTensorDim(tensor_in_shape, data_format, 'N'); |
| 72 | window_rows = GetTensorDim(ksize, data_format, 'H'); |
| 73 | window_cols = GetTensorDim(ksize, data_format, 'W'); |
| 74 | depth_window = GetTensorDim(ksize, data_format, 'C'); |
| 75 | row_stride = GetTensorDim(stride, data_format, 'H'); |
| 76 | col_stride = GetTensorDim(stride, data_format, 'W'); |
| 77 | depth_stride = GetTensorDim(stride, data_format, 'C'); |
| 78 | |
| 79 | // We only support 2D pooling across width/height and depthwise |
| 80 | // pooling, not a combination. |
| 81 | OP_REQUIRES(context, |
| 82 | (depth_window == 1 || (window_rows == 1 && window_cols == 1)), |
| 83 | errors::Unimplemented( |
| 84 | "MaxPooling supports exactly one of pooling across depth " |
| 85 | "or pooling across width/height.")); |
| 86 | |
| 87 | if (depth_window == 1) { |
| 88 | OP_REQUIRES_OK( |
| 89 | context, GetWindowedOutputSize(tensor_in_rows, window_rows, row_stride, |
| 90 | padding, &out_height, &pad_rows)); |
| 91 | OP_REQUIRES_OK( |
| 92 | context, GetWindowedOutputSize(tensor_in_cols, window_cols, col_stride, |
| 93 | padding, &out_width, &pad_cols)); |
| 94 | pad_depth = 0; |
| 95 | out_depth = depth; |
| 96 | } else { |
| 97 | OP_REQUIRES(context, depth_window > 0, |
| 98 | errors::InvalidArgument("depth_window must not be 0")); |
| 99 | // Our current version of depthwise max pooling does not support |
| 100 | // any padding, and expects the depth_window to equal the |
| 101 | // depth_stride (no overlapping). |
| 102 | OP_REQUIRES( |
| 103 | context, depth % depth_window == 0, |
| 104 | errors::Unimplemented("Depthwise max pooling requires the depth " |
| 105 | "window to evenly divide the input depth")); |
| 106 | OP_REQUIRES( |
| 107 | context, depth_stride == depth_window, |
| 108 | errors::Unimplemented("Depthwise max pooling requires the depth " |
| 109 | "window to equal the depth stride")); |
nothing calls this directly
no test coverage detected