nnfw_api_wrapper.h

Go to the documentation of this file.

/*
 * Copyright (c) 2023 Samsung Electronics Co., Ltd. All Rights Reserved
 *
 * Licensed under the Apache License, Version 2.0 (the "License");
 * you may not use this file except in compliance with the License.
 * You may obtain a copy of the License at
 *
 *      http://www.apache.org/licenses/LICENSE-2.0
 *
 * Unless required by applicable law or agreed to in writing, software
 * distributed under the License is distributed on an "AS IS" BASIS,
 * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
 * See the License for the specific language governing permissions and
 * limitations under the License.
 */
 
#ifndef __ONERT_API_PYTHON_NNFW_API_WRAPPER_H__
#define __ONERT_API_PYTHON_NNFW_API_WRAPPER_H__
 
#include <string>
 
#include "nnfw.h"
#include "nnfw_experimental.h"
#include "nnfw_internal.h"
 
#include <pybind11/stl.h>
#include <pybind11/numpy.h>
 
namespace onert
{
namespace api
{
namespace python
{
 
namespace py = pybind11;
 
struct tensorinfo
{
  std::string dtype;
  int32_t rank;
  int32_t dims[NNFW_MAX_RANK];
};
 
void ensure_status(NNFW_STATUS status);
 
NNFW_LAYOUT getLayout(const char *layout = "");
 
NNFW_TYPE getType(const char *type = "");
 
const char *getStringType(NNFW_TYPE type);
 
uint64_t num_elems(const nnfw_tensorinfo *tensor_info);
 
py::list get_dims(const tensorinfo &tensor_info);
 
void set_dims(tensorinfo &tensor_info, const py::list &array);
 
class NNFW_SESSION
{
private:
  nnfw_session *session;
 
public:
  NNFW_SESSION(const char *package_file_path, const char *backends);
  ~NNFW_SESSION();
 
  void close_session();
  void set_input_tensorinfo(uint32_t index, const tensorinfo *tensor_info);
  void prepare();
  void run();
  void run_async();
  void wait();
  template <typename T> void set_input(uint32_t index, py::array_t<T> &buffer)
  {
    nnfw_tensorinfo tensor_info;
    nnfw_input_tensorinfo(this->session, index, &tensor_info);
    NNFW_TYPE type = tensor_info.dtype;
    uint32_t input_elements = num_elems(&tensor_info);
    size_t length = sizeof(T) * input_elements;
 
    ensure_status(nnfw_set_input(session, index, type, buffer.request().ptr, length));
  }
  template <typename T> void set_output(uint32_t index, py::array_t<T> &buffer)
  {
    nnfw_tensorinfo tensor_info;
    nnfw_output_tensorinfo(this->session, index, &tensor_info);
    NNFW_TYPE type = tensor_info.dtype;
    uint32_t output_elements = num_elems(&tensor_info);
    size_t length = sizeof(T) * output_elements;
 
    ensure_status(nnfw_set_output(session, index, type, buffer.request().ptr, length));
  }
  uint32_t input_size();
  uint32_t output_size();
  // process the input layout by receiving a string from Python instead of NNFW_LAYOUT
  void set_input_layout(uint32_t index, const char *layout);
  // process the output layout by receiving a string from Python instead of NNFW_LAYOUT
  void set_output_layout(uint32_t index, const char *layout);
  tensorinfo input_tensorinfo(uint32_t index);
  tensorinfo output_tensorinfo(uint32_t index);
 
  // Internal APIs
  py::array get_output(uint32_t index);
 
  // Experimental APIs for inference
  void set_prepare_config(NNFW_PREPARE_CONFIG config);
 
  // Experimental APIs for training
  nnfw_train_info train_get_traininfo();
  void train_set_traininfo(const nnfw_train_info *info);
 
  template <typename T> void train_set_input(uint32_t index, py::array_t<T> &buffer)
  {
    nnfw_tensorinfo tensor_info;
    nnfw_input_tensorinfo(this->session, index, &tensor_info);
 
    py::buffer_info buf_info = buffer.request();
    const auto buf_shape = buf_info.shape;
    assert(tensor_info.rank == static_cast<int32_t>(buf_shape.size()) && buf_shape.size() > 0);
    tensor_info.dims[0] = static_cast<int32_t>(buf_shape.at(0));
 
    ensure_status(nnfw_train_set_input(this->session, index, buffer.request().ptr, &tensor_info));
  }
  template <typename T> void train_set_expected(uint32_t index, py::array_t<T> &buffer)
  {
    nnfw_tensorinfo tensor_info;
    nnfw_output_tensorinfo(this->session, index, &tensor_info);
 
    py::buffer_info buf_info = buffer.request();
    const auto buf_shape = buf_info.shape;
    assert(tensor_info.rank == static_cast<int32_t>(buf_shape.size()) && buf_shape.size() > 0);
    tensor_info.dims[0] = static_cast<int32_t>(buf_shape.at(0));
 
    ensure_status(
      nnfw_train_set_expected(this->session, index, buffer.request().ptr, &tensor_info));
  }
  template <typename T> void train_set_output(uint32_t index, py::array_t<T> &buffer)
  {
    nnfw_tensorinfo tensor_info;
    nnfw_output_tensorinfo(this->session, index, &tensor_info);
    NNFW_TYPE type = tensor_info.dtype;
    uint32_t output_elements = num_elems(&tensor_info);
    size_t length = sizeof(T) * output_elements;
 
    ensure_status(nnfw_train_set_output(session, index, type, buffer.request().ptr, length));
  }
 
  void train_prepare();
  void train(bool update_weights);
  float train_get_loss(uint32_t index);
 
  void train_export_circle(const py::str &path);
  void train_import_checkpoint(const py::str &path);
  void train_export_checkpoint(const py::str &path);
 
  // TODO Add other apis
};
 
} // namespace python
} // namespace api
} // namespace onert
 
#endif // __ONERT_API_PYTHON_NNFW_API_WRAPPER_H__