search_space.h

Go to the documentation of this file.

/* Brush
copyright 2020 William La Cava
license: GNU/GPL v3
*/
#ifndef SEARCHSPACE_H 
#define SEARCHSPACE_H
//internal includes
#include "../init.h"
#include "../program/node.h"
#include "../program/nodetype.h"
#include "../program/tree_node.h"
// #include "program/program.h"
#include "../util/error.h"
#include "../util/utils.h"
#include "../util/rnd.h"
#include "../params.h"
#include <utility>
#include <optional>
#include <iostream>
 
/* Defines the search space of Brush. 
 *  The search spaces consists of nodes and their accompanying probability
 *  distribution. 
 *  Nodes can be accessed by type, signature, or a combination.  
 *  You may also sample the search space by return type. 
 *  Sampling is done in proportion to the weight associated with 
 *  each node. By default, sampling is done uniform randomly.
*/
using namespace Brush::Data;
using namespace Brush::Util; 
using Brush::Node;
using Brush::DataType;
using std::type_index; 
 
 
namespace Brush
{
// node generation routines
/* template<typename T> */
/* tuple<set<Node>,set<type_index>> generate_nodes(vector<string>& op_names); */
/* tuple<set<Node>,set<type_index>> generate_split_nodes(vector<string>& op_names); */
 
// forward declarations
using TreeIter = tree<Node>::pre_order_iterator;
// template<typename T> struct Program;
// enum class ProgramType: uint32_t;
// template<typename T> struct ProgramTypeEnum; 
 
vector<Node> generate_terminals(const Dataset& d, const bool weights_init);

 
extern std::unordered_map<std::size_t, std::string> ArgsName; 

struct SearchSpace
{
    using ArgsHash = std::size_t; 
 
    template<typename T>
    using Map = unordered_map<DataType,           // return type
                    unordered_map<ArgsHash,         // hash of arg types
                        unordered_map<NodeType,   // node type 
                            T>>>;        // the data!
    
    Map<Node> node_map;

    Map<float> node_map_weights; 

    unordered_map<DataType, vector<Node>> terminal_map;

    unordered_map<DataType, vector<float>> terminal_weights;

    vector<DataType> terminal_types;
 
    // serialization
#ifndef DOXYGEN_SKIP
 
    NLOHMANN_DEFINE_TYPE_INTRUSIVE(SearchSpace, 
        node_map,
        node_map_weights,
        terminal_map,
        terminal_weights,
        terminal_types
    ) 
 
#endif
    
    template<typename PT>
    PT make_program(const Parameters& params, int max_d=0, int max_size=0);

    RegressorProgram make_regressor(int max_d = 0, int max_size = 0, const Parameters& params=Parameters());

    ClassifierProgram make_classifier(int max_d = 0, int max_size = 0,  const Parameters& params=Parameters());

    MulticlassClassifierProgram make_multiclass_classifier(int max_d = 0, int max_size = 0,  const Parameters& params=Parameters());

    RepresenterProgram make_representer(int max_d = 0, int max_size = 0,  const Parameters& params=Parameters());
 
    SearchSpace() = default;

    SearchSpace(const Dataset& d, const unordered_map<string,float>& user_ops = {}, bool weights_init = true){
        init(d,user_ops,weights_init);
    }

    void init(const Dataset& d, const unordered_map<string,float>& user_ops = {}, bool weights_init = true);

    bool check(DataType R) const {
        if (node_map.find(R) == node_map.end()){
            auto msg = fmt::format("{} not in node_map\n",R);
            HANDLE_ERROR_THROW(msg); 
        }
        return true;
    }

    bool check(DataType R, size_t sig_hash) const
    {
        if (check(R)){
            if (node_map.at(R).find(sig_hash) == node_map.at(R).end()){
                auto msg = fmt::format("{} not in node_map.at({})\n", sig_hash, R);
                HANDLE_ERROR_THROW(msg); 
            }
        }
        return true;
    }

    bool check(DataType R, size_t sig_hash, NodeType type) const
    {
        if (check(R,sig_hash)){
            if (node_map.at(R).at(sig_hash).find(type) == node_map.at(R).at(sig_hash).end()){
 
                auto msg = fmt::format("{} not in node_map[{}][{}]\n",type, sig_hash, R);
                HANDLE_ERROR_THROW(msg); 
            }}
        return true;
    }

    template<typename Iter>
    bool has_solution_space(Iter start, Iter end) const {
        return !std::all_of(start, end, [](const auto& w) { return w<=0.0; });
    }
 
    template<typename F> Node get(const string& name);

    Node get(NodeType type, DataType R, size_t sig_hash)
    {
        check(R, sig_hash, type);
        return node_map.at(R).at(sig_hash).at(type);
    };

    template<typename S>
    Node get(NodeType type, DataType R, S sig){ return get(type, R, sig.hash()); };

    vector<float> get_weights() const
    {
        vector<float> v;
        for (auto& [ret, arg_w_map]: node_map_weights) 
        {
            v.push_back(0);
            for (const auto& [arg, name_map] : arg_w_map)
            {
                for (const auto& [name, w]: name_map)
                {
                    v.back() += w; 
                }
            }
        }
        return v;
    };

    vector<float> get_weights(DataType ret) const
    {
        vector<float> v;
        for (const auto& [arg, name_map] : node_map_weights.at(ret))
        {
            v.push_back(0);
            for (const auto& [name, w]: name_map)
            {
                v.back() += w; 
            }
 
        }
        return v;
    };

    vector<float> get_weights(DataType ret, ArgsHash sig_hash) const
    {
        vector<float> v;
        for (const auto& [name, w]: node_map_weights.at(ret).at(sig_hash))
            v.push_back(w); 
 
        return v;
    };

    std::optional<Node> sample_terminal(bool force_return=false) const
    {
        //TODO: match terminal args_type (probably '{}' or something?)
        //  make a separate terminal_map
 
        // We'll make terminal types to have its weights proportional to the
        // DataTypes Weights they hold
        vector<float> data_type_weights(terminal_weights.size());
        if (force_return)
        {
            std::fill(data_type_weights.begin(), data_type_weights.end(), 1.0f); 
        }
        else
        {
            std::transform(
                terminal_weights.begin(),
                terminal_weights.end(),
                data_type_weights.begin(),
                [](const auto& tw){ 
                    return std::reduce(tw.second.begin(), tw.second.end()); }
            );
            
            if (!has_solution_space(data_type_weights.begin(), 
                                    data_type_weights.end()))
                return std::nullopt;
        }
 
        // If we got this far, then it is garanteed that we'll return something
        // The match take into account datatypes with non-zero weights
        auto match = *r.select_randomly(
            terminal_map.begin(),
            terminal_map.end(),
            data_type_weights.begin(),
            data_type_weights.end()
        );
 
        // theres always a constant of each data type
        vector<float> match_weights(match.second.size());
        if (force_return)
        {
            std::fill(match_weights.begin(), match_weights.end(), 1.0f); 
        }
        else
        {
            std::transform(
                terminal_weights.at(match.first).begin(),
                terminal_weights.at(match.first).end(),
                match_weights.begin(),
                [](const auto& w){ return w; });
            
            if (!has_solution_space(match_weights.begin(), 
                                    match_weights.end()))
                return std::nullopt;
        }
 
        return *r.select_randomly(match.second.begin(),  match.second.end(), 
                                  match_weights.begin(), match_weights.end());
    };

    std::optional<Node> sample_terminal(DataType R, bool force_return=false) const
    {
        // should I keep doing this check?
        // if (terminal_map.find(R) == terminal_map.end()){
        //     auto msg = fmt::format("{} not in terminal_map\n",R);
        //     HANDLE_ERROR_THROW(msg); 
        // }
 
        // If there's at least one constant for every data type, its always possible to force sample_terminal to return something
 
        // TODO: try to combine with above function
        vector<float> match_weights(terminal_weights.at(R).size());
        if (force_return)
        {
            std::fill(match_weights.begin(), match_weights.end(), 1.0f); 
        }
        else
        {
            std::transform(
                terminal_weights.at(R).begin(),
                terminal_weights.at(R).end(),
                match_weights.begin(),
                [](const auto& w){  return w; }
            );
 
            if ( (terminal_map.find(R) == terminal_map.end())
            ||   (!has_solution_space(match_weights.begin(), 
                                      match_weights.end())) )
            return std::nullopt;
        }
    
        return *r.select_randomly(terminal_map.at(R).begin(), 
                                  terminal_map.at(R).end(),
                                  match_weights.begin(),
                                  match_weights.end());
    };

    std::optional<Node> sample_op(DataType ret) const
    {
        // check(ret);
        if (node_map.find(ret) == node_map.end())
            return std::nullopt;
 
        //TODO: match terminal args_type (probably '{}' or something?)
        auto ret_match = node_map.at(ret);
 
        vector<float> args_w = get_weights(ret);
 
        if (!has_solution_space(args_w.begin(), args_w.end()))
            return std::nullopt;
 
        auto arg_match = *r.select_randomly(ret_match.begin(), 
                                            ret_match.end(), 
                                            args_w.begin(), 
                                            args_w.end());
 
        vector<float> name_w = get_weights(ret, arg_match.first);
 
        if (!has_solution_space(name_w.begin(), name_w.end()))
            return std::nullopt;
 
        return (*r.select_randomly(arg_match.second.begin(), 
                                   arg_match.second.end(), 
                                   name_w.begin(), 
                                   name_w.end())).second;
    };

    std::optional<Node> sample_op(NodeType type, DataType R)
    {
        // check(R);
        if (node_map.find(R) == node_map.end())
            return std::nullopt;
 
        auto ret_match = node_map.at(R);
        
        vector<Node> matches; 
        vector<float> weights; 
        for (const auto& kv: ret_match)
        {
            auto arg_hash = kv.first;
            auto node_type_map = kv.second;
            if (node_type_map.find(type) != node_type_map.end())
            {
                matches.push_back(node_type_map.at(type));
                weights.push_back(node_map_weights.at(R).at(arg_hash).at(type));
            }
        }
 
        if ( (weights.size()==0)
        ||   (!has_solution_space(weights.begin(), 
                                  weights.end())) )
            return std::nullopt;
        
        return (*r.select_randomly(matches.begin(),
                                   matches.end(),
                                   weights.begin(),
                                   weights.end()));
    };

    std::optional<Node> sample_op_with_arg(DataType ret, DataType arg, 
                              bool terminal_compatible=true,
                              int max_args=0) const
    {
        // thoughts (TODO):
        //  this could be templated by return type and arg. although the lookup in the map should be
        //  fairly fast. 
        //TODO: these needs to be overhauled 
        // fmt::print("sample_op_with_arg");
        check(ret);
 
        auto args_map = node_map.at(ret);
        vector<Node>  matches;
        vector<float> weights;
 
        for (const auto& [args_type, name_map]: args_map) {
            for (const auto& [name, node]: name_map) {
                auto node_arg_types = node.get_arg_types();
 
                // has no size limit (max_arg_count==0) or the number of
                // arguments woudn't exceed the maximum number of arguments
                auto within_size_limit = !(max_args) || (node.get_arg_count() <= max_args);
                
                if ( in(node_arg_types, arg) && within_size_limit) {
                    // if checking terminal compatibility, make sure there's
                    // a compatible terminal for the node's other arguments
                    if (terminal_compatible) {
                        bool compatible = true;
                        for (const auto& arg_type: node_arg_types) { 
                            if (arg_type != arg) {
                                if ( ! in(terminal_types, arg_type) ) {
                                    compatible = false;
                                    break;
                                }
                            }
                        }
                        if (! compatible)
                            continue;
                    }
                    // if we made it this far, include the node as a match!
                    matches.push_back(node);
                    weights.push_back(node_map_weights.at(ret).at(args_type).at(name));
                }
            }
        }
 
        if ( (weights.size()==0)
        ||   (!has_solution_space(weights.begin(), 
                                  weights.end())) )
            return std::nullopt;
 
        return (*r.select_randomly(matches.begin(), matches.end(), 
                                   weights.begin(), weights.end()));
    };
    
    std::optional<Node> get_node_like(Node node) const
    {
        if (Is<NodeType::Terminal, NodeType::Constant, NodeType::MeanLabel>(node.node_type)){
            return sample_terminal(node.ret_type);
        }
 
        auto matches = node_map.at(node.ret_type).at(node.args_type());
        auto match_weights = get_weights(node.ret_type, node.args_type());
 
        if ( (match_weights.size()==0)
        ||   (!has_solution_space(match_weights.begin(), 
                                  match_weights.end())) )
            return std::nullopt;
 
        return (*r.select_randomly(matches.begin(), 
                                   matches.end(), 
                                   match_weights.begin(), 
                                   match_weights.end())
               ).second;
    };

    std::optional<tree<Node>> sample_subtree(Node root, int max_d, int max_size) const;

    void print() const; 
 
    private:
        tree<Node>& PTC2(tree<Node>& Tree, tree<Node>::iterator root, int max_d, int max_size) const;
 
        template<NodeType NT, typename S>
        requires (!is_in_v<NT, NodeType::Terminal, NodeType::Constant, NodeType::MeanLabel>)
        static constexpr std::optional<Node> CreateNode(
            const auto& unique_data_types, 
            bool use_all, 
            bool weighted
        )
        {
            // prune the operators out that don't have argument types that
            // overlap with feature data types
            for (auto arg: S::get_arg_types()){
                if (! in(unique_data_types,arg) ){
                    return {}; 
                }
            }    
            ArgsName[S::hash()] = fmt::format("{}", S::get_arg_types());
            return Node(NT, S{}, weighted);
        }
       
        template<NodeType NT, typename S>
        constexpr void AddNode(
            const unordered_map<string,float>& user_ops, 
            const vector<DataType>& unique_data_types
        )
        {
            bool use_all = user_ops.size() == 0;
            auto name = NodeTypeName[NT];
 
            bool weighted = false;
            if (Is<NodeType::OffsetSum>(NT)) // this has to have weights on by default
                weighted = true;
    
            auto n_maybe = CreateNode<NT,S>(unique_data_types, use_all, weighted);
 
            if (n_maybe){
                auto n = n_maybe.value();
                node_map[n.ret_type][n.args_type()][n.node_type] = n;
                // sampling probability map
                float w = use_all? 1.0 : user_ops.at(name);
                node_map_weights[n.ret_type][n.args_type()][n.node_type] = w;
            }
        }
 
        template <NodeType NT, typename Sigs, std::size_t... Is>
        constexpr void AddNodes(const unordered_map<string, float> &user_ops,
                                const vector<DataType> &unique_data_types,
                                std::index_sequence<Is...>)
        {
            (AddNode<NT, std::tuple_element_t<Is, Sigs>>(user_ops, unique_data_types), ...);
        }
 
        template<NodeType NT>
        void MakeNodes(const unordered_map<string,float>& user_ops, 
                       const vector<DataType>& unique_data_types
                      ) 
        {
            if (Is<NodeType::Terminal, NodeType::Constant, NodeType::MeanLabel>(NT))
                return;
            bool use_all = user_ops.size() == 0;
            auto name = NodeTypeName.at(NT);
 
            // skip operators not defined by user
            if (!use_all & user_ops.find(name) == user_ops.end())
                return;
 
            using signatures = Signatures<NT>::type;
            constexpr auto size = std::tuple_size<signatures>::value;
            AddNodes<NT, signatures>(
                user_ops,
                unique_data_types,
                std::make_index_sequence<size>()
            ); 
        }
 
        template<std::size_t... Is>
        void GenerateNodeMap(const unordered_map<string,float>& user_ops, 
                const vector<DataType>& unique_data_types, 
                std::index_sequence<Is...> 
                )
        {
            auto nt = [](auto i) { return static_cast<NodeType>(1UL << i); };
            (MakeNodes<nt(Is)>(user_ops, unique_data_types), ...);
        }
}; // SearchSpace

template<typename T>
T RandomDequeue(std::vector<T>& Q)
{
    int loc = r.rnd_int(0, Q.size()-1);
    std::swap(Q[loc], Q[Q.size()-1]);
    T val = Q.back();
    Q.pop_back();
    return val;
};
 
template<typename P>
P SearchSpace::make_program(const Parameters& params, int max_d, int max_size)
{
    // this is what makes `make_program` create uniformly distributed
    // individuals to feed initial population
    if (max_d < 1)
        max_d = r.rnd_int(1, params.max_depth);
    if (max_size < 1) 
        max_size = r.rnd_int(1, params.max_size);
 
    DataType root_type = DataTypeEnum<typename P::TreeType>::value;
    ProgramType program_type = P::program_type;
    // ProgramType program_type = ProgramTypeEnum<PT>::value;
    
    // Tree is pre-filled with some fixed nodes depending on program type
    auto Tree = tree<Node>();
 
    // building the tree for each program case. Then, we give the spot to PTC2,
    // and it will fill the rest of the tree
    tree<Node>::iterator spot;
 
    // building the root node for each program case
    if (P::program_type == ProgramType::BinaryClassifier)
    {
        Node node_logit = get(NodeType::Logistic, DataType::ArrayF, Signature<ArrayXf(ArrayXf)>());
        node_logit.set_prob_change(0.0);
        node_logit.fixed=true;
        auto spot_logit = Tree.insert(Tree.begin(), node_logit);
 
        if (true) { // Logistic(Add(Constant, <>)). 
            Node node_offset = get(NodeType::OffsetSum, DataType::ArrayF, Signature<ArrayXf(ArrayXf)>());
            node_offset.set_prob_change(0.0);
            node_offset.fixed=true;
 
            auto spot_offset = Tree.append_child(spot_logit);
            
            spot = Tree.replace(spot_offset, node_offset);
        }
        else { // If false, then model will be Logistic(<>)
            spot = spot_logit;
        }
    }
    else if (P::program_type == ProgramType::MulticlassClassifier)
    {
        Node node_softmax = get(NodeType::Softmax, DataType::MatrixF, Signature<ArrayXXf(ArrayXXf)>());
        node_softmax.set_prob_change(0.0);
        node_softmax.fixed=true;
        
        spot = Tree.insert(Tree.begin(), node_softmax);
    }
    else // regression or representer --- sampling any candidate op or terminal
    {
        Node root;
 
        std::optional<Node> opt=std::nullopt;
 
        if (max_size>1 && max_d>1)
            opt = sample_op(root_type);
 
        if (!opt) // if failed, then we dont have any operator to use as root...
            opt = sample_terminal(root_type, true);
 
        root = opt.value();
    
        spot = Tree.insert(Tree.begin(), root);
    }
 
    // max_d-1 because we always pick the root before calling ptc2
    PTC2(Tree, spot, max_d-1, max_size); // change inplace
 
    return P(*this, Tree);
};
 
extern SearchSpace SS;
 
} // Brush
 
// format overload 
template <> struct fmt::formatter<Brush::SearchSpace>: formatter<string_view> {
  template <typename FormatContext>
  auto format(const Brush::SearchSpace& SS, FormatContext& ctx) const {
    string output = "Search Space\n===\n";
    output += fmt::format("terminal_map: {}\n", SS.terminal_map);
    output += fmt::format("terminal_weights: {}\n", SS.terminal_weights);
    for (const auto& [ret_type, v] : SS.node_map) {
        for (const auto& [args_type, v2] : v) {
            for (const auto& [node_type, node] : v2) {
                output += fmt::format("node_map[{}][{}][{}] = {}, weight = {}\n", 
                        ret_type,
                        ArgsName[args_type],
                        node_type,
                        node,
                        SS.node_map_weights.at(ret_type).at(args_type).at(node_type)
                        );
            }
        }
    }
    output += "===";
    return formatter<string_view>::format(output, ctx);
  }
};
#endif