-module(gleam_synapses@model@net_elems@neuron).
-compile(no_auto_import).

-export([init/3, output/2, back_propagated/4, serialized/1, deserialized/1, json_encoded/1, json_decoder/0, generator/1]).
-export_type([neuron/0, neuron_serialized/0]).

-type neuron() :: {neuron,
                   gleam_synapses@model@net_elems@activation:activation(),
                   gleam_zlists@interop:z_list(float())}.

-type neuron_serialized() :: {neuron_serialized, binary(), list(float())}.

-spec init(
    integer(),
    gleam_synapses@model@net_elems@activation:activation(),
    fun(() -> float())
) -> neuron().
init(Input_size, Activation_f, Weight_init_f) ->
    Weights = gleam_zlists:map(
        gleam_zlists:take(gleam_zlists:indices(), Input_size + 1),
        fun(_) -> Weight_init_f() end
    ),
    {neuron, Activation_f, Weights}.

-spec output(neuron(), gleam_zlists@interop:z_list(float())) -> float().
output(Neuron, Input_val) ->
    Activation_input = gleam_synapses@model@mathematics:dot_product(
        gleam_zlists:cons(Input_val, 1.0),
        erlang:element(3, Neuron)
    ),
    (erlang:element(3, erlang:element(2, Neuron)))(
        gleam_synapses@model@net_elems@activation:restricted_input(
            erlang:element(2, Neuron),
            Activation_input
        )
    ).

-spec back_propagated(
    neuron(),
    float(),
    gleam_zlists@interop:z_list(float()),
    {float(), float()}
) -> {gleam_zlists@interop:z_list(float()), neuron()}.
back_propagated(Neuron, Learning_rate, Input_val, Output_with_error) ->
    {Output_val, Error} = Output_with_error,
    Output_inverse = (erlang:element(5, erlang:element(2, Neuron)))(Output_val),
    Common = Error * (erlang:element(4, erlang:element(2, Neuron)))(
        Output_inverse
    ),
    In_errors = gleam_zlists:map(Input_val, fun(X) -> X * Common end),
    New_weights = gleam_zlists:map(
        gleam_zlists:zip(
            gleam_zlists:cons(Input_val, 1.0),
            erlang:element(3, Neuron)
        ),
        fun(X@1) -> {A, B} = X@1,
            B - ((Learning_rate * Common) * A) end
    ),
    New_neuron = {neuron, erlang:element(2, Neuron), New_weights},
    {In_errors, New_neuron}.

-spec serialized(neuron()) -> neuron_serialized().
serialized(Neuron) ->
    {neuron_serialized,
     gleam_synapses@model@net_elems@activation:serialized(
         erlang:element(2, Neuron)
     ),
     gleam_zlists:to_list(erlang:element(3, Neuron))}.

-spec deserialized(neuron_serialized()) -> neuron().
deserialized(Neuron_serialized) ->
    {neuron,
     gleam_synapses@model@net_elems@activation:deserialized(
         erlang:element(2, Neuron_serialized)
     ),
     gleam_zlists:of_list(erlang:element(3, Neuron_serialized))}.

-spec json_encoded(neuron_serialized()) -> gleam@jsone:json_value().
json_encoded(Neuron_serialized) ->
    gleam@jsone:object(
        [{<<"activationF"/utf8>>,
          gleam_synapses@model@net_elems@activation:json_encoded(
              erlang:element(2, Neuron_serialized)
          )},
         {<<"weights"/utf8>>,
          gleam@jsone:array(
              erlang:element(3, Neuron_serialized),
              fun gleam@jsone:float/1
          )}]
    ).

-spec json_decoder() -> decode:decoder(neuron_serialized()).
json_decoder() ->
    decode:map2(
        fun(A, B) -> {neuron_serialized, A, B} end,
        decode:field(
            <<"activationF"/utf8>>,
            gleam_synapses@model@net_elems@activation:json_decoder()
        ),
        decode:field(<<"weights"/utf8>>, decode:list(decode:float()))
    ).

-spec generator(integer()) -> minigen:generator(neuron()).
generator(Input_size) ->
    Weights_generator = minigen:map(
        minigen:list(
            minigen:map(minigen:float(), fun(X) -> 1.0 - (2.0 * X) end),
            Input_size
            + 1
        ),
        fun gleam_zlists:of_list/1
    ),
    minigen:map2(
        gleam_synapses@model@net_elems@activation:generator(),
        Weights_generator,
        fun(A, B) -> {neuron, A, B} end
    ).