mlmodel.hpp

// Copyright 2023 Viam Inc.
//
// 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.
 
#pragma once
 
#include <boost/mpl/joint_view.hpp>
#include <boost/mpl/list.hpp>
#include <boost/mpl/transform_view.hpp>
#include <boost/variant/variant.hpp>
#include <xtensor/xadapt.hpp>
 
#include <viam/sdk/common/utils.hpp>
#include <viam/sdk/services/service.hpp>
 
namespace viam {
namespace sdk {
 
class MLModelService : public Service {
   private:
    template <typename T>
    struct make_tensor_view_ {
        using shape_t = std::vector<std::size_t>;
 
        using xt_no_ownership_t = decltype(xt::no_ownership());
 
        using type = decltype(xt::adapt(std::declval<const T*>(),
                                        std::declval<std::size_t>(),
                                        std::declval<xt_no_ownership_t>(),
                                        std::declval<shape_t>()));
    };
 
   public:
    API api() const override;
 
    template <typename T>
    using tensor_view = typename make_tensor_view_<T>::type;
 
    template <typename T>
    static tensor_view<T> make_tensor_view(const T* data,
                                           std::size_t size,
                                           typename tensor_view<T>::shape_type shape) {
        return xt::adapt(std::move(data), std::move(size), xt::no_ownership(), std::move(shape));
    }
 
    // Now that we have a factory for our tensor view types, use mpl
    // to produce a variant over tensor views over the primitive types
    // we care about, which are the signed and unsigned fixed width
    // integral types and the two floating point types.
    using signed_integral_base_types =
        boost::mpl::list<std::int8_t, std::int16_t, std::int32_t, std::int64_t>;
 
    using unsigned_integral_base_types =
        boost::mpl::transform_view<signed_integral_base_types,
                                   std::make_unsigned<boost::mpl::placeholders::_1>>;
 
    using integral_base_types =
        boost::mpl::joint_view<signed_integral_base_types, unsigned_integral_base_types>;
 
    using fp_base_types = boost::mpl::list<float, double>;
 
    using base_types = boost::mpl::joint_view<integral_base_types, fp_base_types>;
 
    using tensor_view_types =
        boost::mpl::transform_view<base_types, make_tensor_view_<boost::mpl::placeholders::_1>>;
 
    // Union the tensor views for the various base types.
    using tensor_views = boost::make_variant_over<tensor_view_types>::type;
 
    // Our parameters to and from the model come as named tensor_views.
    using named_tensor_views = std::unordered_map<std::string, tensor_views>;
 
    inline std::shared_ptr<named_tensor_views> infer(const named_tensor_views& inputs) {
        return infer(inputs, {});
    }
 
    virtual std::shared_ptr<named_tensor_views> infer(const named_tensor_views& inputs,
                                                      const ProtoStruct& extra) = 0;
 
    struct tensor_info {
        struct file {
            std::string name;
            std::string description;
 
            enum : std::uint8_t {
                k_label_type_tensor_value = 0,
                k_label_type_tensor_axis = 1,
            } label_type;
        };
 
        std::string name;
        std::string description;
 
        enum class data_types : std::uint8_t {
            k_int8 = 0,
            k_uint8 = 1,
            k_int16 = 2,
            k_uint16 = 3,
            k_int32 = 4,
            k_uint32 = 5,
            k_int64 = 6,
            k_uint64 = 7,
            k_float32 = 8,
            k_float64 = 9,
        } data_type;
 
        std::vector<int> shape;
        std::vector<file> associated_files;
 
        ProtoStruct extra;
 
        static boost::optional<data_types> string_to_data_type(const std::string& str);
        static const char* data_type_to_string(data_types data_type);
 
        static data_types tensor_views_to_data_type(const tensor_views& view);
    };
 
    struct metadata {
        std::string name;
        std::string type;
        std::string description;
        std::vector<tensor_info> inputs;
        std::vector<tensor_info> outputs;
    };
 
    inline struct metadata metadata() {
        return metadata({});
    }
 
    virtual struct metadata metadata(const ProtoStruct& extra) = 0;
 
   protected:
    explicit MLModelService(std::string name);
};
 
template <>
struct API::traits<MLModelService> {
    static API api();
};
 
}  // namespace sdk
}  // namespace viam