path: root/nerv/layer/affine.lua

--- Parameter and layer classes related to linear transform.

local MatrixParam = nerv.class('nerv.MatrixParam', 'nerv.Param')
local LinearTransParam = nerv.class('nerv.LinearTransParam', 'nerv.MatrixParam')
local BiasParam = nerv.class('nerv.BiasParam', 'nerv.MatrixParam')
local AffineLayer = nerv.class('nerv.AffineLayer', 'nerv.Layer')

--- A parameter that consists of a single matrix
-- @type nerv.MatrixParam

function MatrixParam:check(checker)
    -- check trans matrix type
    checker(self.trans)
end

--- Read from a file handle.
-- @param handle the file handle
function MatrixParam:read(handle)
    self.trans = self.gconf.mmat_type.load(handle)
end

function MatrixParam:write(handle)
    self.trans:save(handle)
end

function MatrixParam:train_init()
    self.correction = self.trans:create()
    self.correction:fill(0)
end

function MatrixParam:copy(copier)
    local target = nerv.MatrixParam(self.id, self.gconf)
    target.trans = copier(self.trans)
    return target
end

function MatrixParam:_update_by_gradient(gradient, alpha, beta)
    local gconf = self.gconf
    -- momentum gain
    local mmt_gain = 1.0 / (1.0 - gconf.momentum)
    local n = gconf.batch_size * mmt_gain
    -- perform update
    if gconf.momentum > 0 then
        self.correction:add(self.correction, gradient, gconf.momentum, 1.0)
        self.trans:add(self.trans, self.correction, alpha, -gconf.lrate / n * beta)
    else
        self.trans:add(self.trans, gradient, alpha, -gconf.lrate / n * beta)
    end
end

function MatrixParam:_update_by_err_input(err, input, alpha, beta)
    local gconf = self.gconf
    -- momentum gain
    local mmt_gain = 1.0 / (1.0 - gconf.momentum)
    local n = gconf.batch_size * mmt_gain
    -- perform update
    if gconf.momentum > 0 then
        self.correction:mul(input, err, 1.0, gconf.momentum, 'T', 'N')
        self.trans:add(self.trans, self.correction, alpha, -gconf.lrate / n * beta)
    else
        self.trans:mul(input, err, -gconf.lrate / n * beta, alpha, 'T', 'N')
    end
end

function MatrixParam:update_by_gradient(gradient)
    self:_update_by_gradient(gradient, 1.0, 1.0)
end

function MatrixParam:update_by_err_input(err, input)
    self:_update_by_err_input(err, input, 1.0, 1.0)
end

function LinearTransParam:update_by_err_input(err, input)
    local gconf = self.gconf
    local l2 = 1 - gconf.lrate * gconf.wcost
    self:_update_by_err_input(err, input, l2, l2)
end

--- A fully-connected linear transform layer.