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')
function MatrixParam:read(handle)
self.trans = self.gconf.cumat_type.new_from_host(
self.gconf.mmat_type.load(handle))
end
function MatrixParam:write(handle)
self.trans:new_to_host():save(handle)
end
function MatrixParam:train_init()
self.correction = self.trans:create()
self.correction:fill(0)
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 = self.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 = self.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 l2 = 1 - gconf.lrate * gconf.wcost
self:_update_by_err_input(err, input, l2, l2)
end
function AffineLayer:__init(id, global_conf, layer_conf)
self.id = id
self.ltp = layer_conf.ltp
self.bp = layer_conf.bp
self.dim_in = layer_conf.dim_in
self.dim_out = layer_conf.dim_out
self.gconf = global_conf
self:check_dim_len(1, 1) -- exactly one input and one output
-- self.direct_update = layer_conf.direct_update or global_conf.direct_update
end
function AffineLayer:init(batch_size)
if self.ltp.trans:ncol() ~= self.bp.trans:ncol() then
nerv.error("mismatching dimensions of linear transform and bias paramter")
end
if self.dim_in[1] ~= self.ltp.trans:nrow() then
nerv.error("mismatching dimensions of linear transform parameter and input")
end
if self.dim_out[1] ~= self.ltp.trans:ncol() then
nerv.error("mismatching dimensions of linear transform parameter and output")
end
self.ltp_grad = self.ltp.trans:create()
self.ltp:train_init()
self.bp:train_init()
end
function AffineLayer:batch_resize(batch_size)
-- do nothing
end
function AffineLayer:update(bp_err, input, output)
self.ltp:update_by_err_input(bp_err[1], input[1])
self.bp:update_by_gradient(bp_err[1]:colsum())
end
function AffineLayer:propagate(input, output)
-- apply linear transform
output[1]:mul(input[1], self.ltp.trans, 1.0, 0.0, 'N', 'N')
-- add bias
output[1]:add_row(self.bp.trans, 1.0)
end
function AffineLayer:back_propagate(bp_err, next_bp_err, input, output)
next_bp_err[1]:mul(bp_err[1], self.ltp.trans, 1.0, 0.0, 'N', 'T')
end
function AffineLayer:get_params()
return nerv.ParamRepo({self.ltp, self.bp})
end