polymorphic_layer_base.h

Go to the documentation of this file.

 /*
 * Layer.h
 *
 *  Created on: Mar 1, 2018
 *    Author: joseph
 */

#ifndef BLACKCAT_NEURAL_NETWORKS_POLYMORPHIC_LAYER_H_
#define LAYER_H_

#include "../layer_loader.h"
#include "layer_traits.h"
#include <string>

#include <sys/types.h>
#include <sys/stat.h>
#include <fstream>
#include <ostream>

namespace bc {
namespace nn {

template<class LearningRateValueType>
struct network_runtime_traits
{
    using value_type = LearningRateValueType;
    value_type m_learning_rate = .03;
    int m_batch_size = 1;
};

template<class SystemTag, class ValueType>
using layer_default_allocator =
        bc::allocators::Polymorphic_Allocator<SystemTag, ValueType>;

template<
    class Dimension,
    class ValueType,
    class SystemTag,
    class Allocator=layer_default_allocator<ValueType, SystemTag>,
    class OutputDimension=Dimension,
    class OutputValueType=ValueType,
    class OutputSystemTag=SystemTag,
    class OutputAllocator=Allocator>
struct Polymorphic_Layer_Base
{
    using allocator_type = Allocator;
    using value_type = ValueType;
    using system_tag = SystemTag;

    using input_tensor_dimension = Dimension;
    using batched_input_tensor_dimension = bc::traits::Integer<Dimension::value+1>;

    using output_tensor_dimension = OutputDimension;
    using batched_output_tensor_dimension = bc::traits::Integer<OutputDimension::value+1>;

    using output_value_type = OutputValueType;
    using output_system_tag = OutputSystemTag;
    using output_allocator_type = OutputAllocator;

    using next_layer_type = Polymorphic_Layer_Base<
            output_tensor_dimension,
            output_value_type,
            output_system_tag>;

    using this_layer_type = Polymorphic_Layer_Base<
            input_tensor_dimension,
            value_type,
            system_tag>;

    using input_tensor_type  = bc::Tensor< input_tensor_dimension::value, value_type, allocator_type>;
    using output_tensor_type = bc::Tensor<output_tensor_dimension::value, value_type, output_allocator_type>;

    using batched_input_tensor_type  = bc::Tensor<input_tensor_dimension::value  + 1, value_type, allocator_type>;
    using batched_output_tensor_type = bc::Tensor<output_tensor_dimension::value + 1, value_type, output_allocator_type>;

    using this_layer_pointer_type = std::shared_ptr<this_layer_type>;
    using next_layer_pointer_type = std::shared_ptr<next_layer_type>;

    using input_shape_type = bc::Dim<input_tensor_dimension::value>;
    using output_shape_type = bc::Dim<output_tensor_dimension::value>;

private:

    const std::string m_classname;
    std::shared_ptr<network_runtime_traits<value_type>> m_network_vars
        = std::shared_ptr<network_runtime_traits<value_type>>(
            new network_runtime_traits<value_type>());

protected:

    std::shared_ptr<allocator_type> m_allocator;
    std::weak_ptr<this_layer_type> m_input_layer;
    std::shared_ptr<next_layer_type> m_output_layer;

    input_shape_type m_input_shape;
    output_shape_type m_output_shape; //ptr??

public:

    Polymorphic_Layer_Base(std::string classname):
        m_classname(parse_classname(classname)) {}

    batched_output_tensor_type y;

    virtual ~Polymorphic_Layer_Base()=default;

    template<int ADL=0>
    std::string get_string_architecture() const {
        return classname() + ':'
            + "\n\tinput_shape: " + m_input_shape.to_string();
    }

    virtual batched_output_tensor_type forward_propagation(
            const batched_input_tensor_type& inputs) = 0;

    virtual batched_input_tensor_type back_propagation(
            const batched_output_tensor_type& delta) = 0;

    batched_output_tensor_type fp(
            const batched_input_tensor_type& inputs) {
        return forward_propagation(inputs);
    }

    batched_input_tensor_type bp(
            const batched_output_tensor_type& delta) {
        return back_propagation(delta);
    }

    virtual void update_weights() {}
    //post constructor initialization
    virtual void init() = 0;

    void set_batch_size(int bs)
    {
        set_batch_size_hook(bs);
        m_network_vars->m_batch_size = bs;
        y = batched_output_tensor_type(this->batched_output_shape());
    }

    void set_learning_rate(double lr)
    {
        set_learning_rate_hook(lr);
        m_network_vars->m_learning_rate = lr;
    }

    std::shared_ptr<this_layer_type> prev() {
        return m_input_layer.lock();
    }

    std::shared_ptr<this_layer_type> prev() const {
        return m_input_layer.lock();
    }

    std::shared_ptr<next_layer_type> next() {
        return m_input_layer;
    }

    std::shared_ptr<next_layer_type> next() const {
        return m_output_layer;
    }

    void set_prev(this_layer_pointer_type prev_layer) {
        m_input_layer = prev_layer;
    }

    void set_next(next_layer_pointer_type next_layer) {
        m_output_layer = next_layer;
    }

protected:

    virtual void set_batch_size_hook(int bs) {}
    virtual void set_learning_rate_hook(double lr) {}

public:

    auto batch_size() const { return m_network_vars->m_batch_size; }

    auto learning_rate() const         { return m_network_vars->m_learning_rate; }
    auto batched_learning_rate() const { return learning_rate() / batch_size();  }

    auto input_shape() const { return m_input_shape; }
    auto output_shape() const { return m_output_shape; }

    auto batched_input_shape()  const { return m_input_shape.concat(batch_size());  }
    auto batched_output_shape() const { return m_output_shape.concat(batch_size()); }

    std::string classname() const { return m_classname; }
    virtual void save(Layer_Loader&) const = 0;
    virtual void save_from_cache(Layer_Loader&, Cache&) const {}
    virtual void load(Layer_Loader&) = 0;
    virtual void load_to_cache(Layer_Loader&, Cache&) {}

    void copy_training_data_to_single_predict(Cache&, int batch_index) {}

    //push me
    static std::string parse_classname(std::string classname) {
        auto classname_ns = std::find(classname.rbegin(), classname.rend(), ':');
        classname.erase(classname.rend().base(), classname_ns.base());
        return classname;
    }
};

template<class... ArgsA, class... ArgsB>
void link(std::shared_ptr<Layer_Base<ArgsA...>> prev, std::shared_ptr<Layer_Base<ArgsB...>> next)
{
    prev->set_next(next);
    next->set_prev(prev);
}

}
}



#endif /* LAYER_H_ */