diff options
Diffstat (limited to 'utils/Utils.h')
| -rw-r--r-- | utils/Utils.h | 105 |
1 files changed, 82 insertions, 23 deletions
diff --git a/utils/Utils.h b/utils/Utils.h index ced6b147f1..adb0e54f54 100644 --- a/utils/Utils.h +++ b/utils/Utils.h @@ -281,21 +281,24 @@ class NPYLoader public: /** Default constructor */ NPYLoader() - : _fs(), _shape(), _fortran_order(false), _typestring() + : _fs(), _shape(), _fortran_order(false), _typestring(), _file_layout(DataLayout::NCHW) { } /** Open a NPY file and reads its metadata * * @param[in] npy_filename File to open + * @param[in] file_layout (Optional) Layout in which the weights are stored in the file. */ - void open(const std::string &npy_filename) + void open(const std::string &npy_filename, DataLayout file_layout = DataLayout::NCHW) { ARM_COMPUTE_ERROR_ON(is_open()); try { - _fs.exceptions(std::ifstream::failbit | std::ifstream::badbit); _fs.open(npy_filename, std::ios::in | std::ios::binary); + ARM_COMPUTE_EXIT_ON_MSG(!_fs.good(), "Failed to load binary data from %s", npy_filename.c_str()); + _fs.exceptions(std::ifstream::failbit | std::ifstream::badbit); + _file_layout = file_layout; std::tie(_shape, _fortran_order, _typestring) = parse_npy_header(_fs); } @@ -332,7 +335,12 @@ public: shape.set_num_dimensions(_shape.size()); for(size_t i = 0; i < _shape.size(); ++i) { - shape.set(i, _shape.at(i)); + size_t src = i; + if(_fortran_order) + { + src = _shape.size() - 1 - i; + } + shape.set(i, _shape.at(src)); } arm_compute::TensorInfo tensor_info(shape, 1, dt); @@ -369,9 +377,25 @@ public: std::string expect_typestr = get_typestring(tensor.info()->data_type()); ARM_COMPUTE_ERROR_ON_MSG(_typestring != expect_typestr, "Typestrings mismatch"); + bool are_layouts_different = (_file_layout != tensor.info()->data_layout()); + // Correct dimensions (Needs to match TensorShape dimension corrections) + if(_shape.size() != tensor.info()->tensor_shape().num_dimensions()) + { + for(int i = static_cast<int>(_shape.size()) - 1; i > 0; --i) + { + if(_shape[i] == 1) + { + _shape.pop_back(); + } + else + { + break; + } + } + } // Validate tensor shape ARM_COMPUTE_ERROR_ON_MSG(_shape.size() != tensor.info()->tensor_shape().num_dimensions(), "Tensor ranks mismatch"); - if(_fortran_order) + if(!_fortran_order) { for(size_t i = 0; i < _shape.size(); ++i) { @@ -391,20 +415,63 @@ public: case arm_compute::DataType::F32: { // Read data - if(tensor.info()->padding().empty()) + if(!are_layouts_different && !_fortran_order && tensor.info()->padding().empty()) { // If tensor has no padding read directly from stream. _fs.read(reinterpret_cast<char *>(tensor.buffer()), tensor.info()->total_size()); } else { - // If tensor has padding accessing tensor elements through execution window. - Window window; - window.use_tensor_dimensions(tensor.info()->tensor_shape()); + // If tensor has padding or is in fortran order accessing tensor elements through execution window. + Window window; + TensorShape permuted_shape = tensor.info()->tensor_shape(); + const unsigned int num_dims = _shape.size(); + arm_compute::PermutationVector perm; + if(_fortran_order) + { + for(unsigned int dim = 0; dim < num_dims; dim++) + { + permuted_shape.set(dim, _shape[dim]); + perm.set(dim, num_dims - dim - 1); + } + } + else + { + for(unsigned int dim = 0; dim < num_dims; dim++) + { + perm.set(dim, dim); + } + } + if(are_layouts_different) + { + // Permute only if num_dimensions greater than 2 + if(num_dims > 2) + { + if(_file_layout == DataLayout::NHWC) // i.e destination is NCHW --> permute(1,2,0) + { + size_t perm_0 = perm[0]; + perm[0] = perm[1]; + perm[1] = perm[2]; + perm[2] = perm_0; + } + else + { + // destination layout is NHWC --> permute (2,0,1) + size_t perm_0 = perm[0]; + perm[0] = perm[2]; + perm[2] = perm[1]; + perm[1] = perm_0; + } + } + } + + window.use_tensor_dimensions(permuted_shape); execute_window_loop(window, [&](const Coordinates & id) { - _fs.read(reinterpret_cast<char *>(tensor.ptr_to_element(id)), tensor.info()->element_size()); + Coordinates dst(id); + arm_compute::permute(dst, perm); + _fs.read(reinterpret_cast<char *>(tensor.ptr_to_element(dst)), tensor.info()->element_size()); }); } @@ -428,6 +495,7 @@ private: std::vector<unsigned long> _shape; bool _fortran_order; std::string _typestring; + DataLayout _file_layout; }; /** Template helper function to save a tensor image to a PPM file. @@ -524,26 +592,18 @@ template <typename T> void save_to_npy(T &tensor, const std::string &npy_filename, bool fortran_order) { ARM_COMPUTE_ERROR_ON_DATA_TYPE_NOT_IN(&tensor, arm_compute::DataType::F32); - ARM_COMPUTE_ERROR_ON(tensor.info()->num_dimensions() > 2); std::ofstream fs; - try { fs.exceptions(std::ofstream::failbit | std::ofstream::badbit | std::ofstream::eofbit); fs.open(npy_filename, std::ios::out | std::ios::binary); - const unsigned int width = tensor.info()->tensor_shape()[0]; - const unsigned int height = tensor.info()->tensor_shape()[1]; - std::vector<npy::ndarray_len_t> shape(2); + std::vector<npy::ndarray_len_t> shape(tensor.info()->num_dimensions()); - if(!fortran_order) - { - shape[0] = height, shape[1] = width; - } - else + for(unsigned int i = 0; i < tensor.info()->num_dimensions(); ++i) { - shape[0] = width, shape[1] = height; + shape[i] = tensor.info()->tensor_shape()[i]; } // Map buffer if creating a CLTensor @@ -561,8 +621,7 @@ void save_to_npy(T &tensor, const std::string &npy_filename, bool fortran_order) npy::write_header(stream, typestring, fortran_order, shape); arm_compute::Window window; - window.set(arm_compute::Window::DimX, arm_compute::Window::Dimension(0, width, 1)); - window.set(arm_compute::Window::DimY, arm_compute::Window::Dimension(0, height, 1)); + window.use_tensor_dimensions(tensor.info()->tensor_shape()); arm_compute::Iterator in(&tensor, window); |