diff options
Diffstat (limited to 'nerv/layer')
| -rw-r--r-- | nerv/layer/affine_recurrent.lua | 52 |
1 files changed, 19 insertions, 33 deletions
diff --git a/nerv/layer/affine_recurrent.lua b/nerv/layer/affine_recurrent.lua index d537f4a..fd6f38f 100644 --- a/nerv/layer/affine_recurrent.lua +++ b/nerv/layer/affine_recurrent.lua @@ -9,31 +9,37 @@ function Recurrent:__init(id, global_conf, layer_conf) self.dim_in = layer_conf.dim_in self.dim_out = layer_conf.dim_out self.gconf = global_conf + self.log_pre = self.id .. "[LOG]" self.bp = self:find_param("bp", layer_conf, global_conf, nerv.BiasParam, {1, self.dim_out[1]}) --layer_conf.bp - self.ltp_hh = self:find_param("ltp_hh", layer_conf, global_conf, nerv.LinearTransParam, {self.dim_in[2], self.dim_out[1]}) --layer_conf.ltp_hh --from hidden to hidden + self.ltp_hh = self:find_param("ltphh", layer_conf, global_conf, nerv.LinearTransParam, {self.dim_in[2], self.dim_out[1]}) --layer_conf.ltp_hh --from hidden to hidden + self.ltp_ih = self:find_param("ltpih", layer_conf, global_conf, nerv.LinearTransParam, {self.dim_in[1], self.dim_out[1]}) --layer_conf.ltp_hh --from hidden to hidden self:check_dim_len(2, 1) self.direct_update = layer_conf.direct_update self.clip = layer_conf.clip --clip error in back_propagate + if self.clip ~= nil then + nerv.info("%s creating, will clip the error by %f", self.log_pre, self.clip) + end end --Check parameter function Recurrent:init(batch_size) - if (self.ltp_hh.trans:ncol() ~= self.bp.trans:ncol()) then + if self.ltp_hh.trans:ncol() ~= self.bp.trans:ncol() or + self.ltp_ih.trans:ncol() ~= self.bp.trans:ncol() then nerv.error("mismatching dimensions of ltp and bp") end - if (self.dim_in[1] ~= self.ltp_hh.trans:nrow() or - self.dim_in[2] ~= self.ltp_hh.trans:nrow()) then + if self.dim_in[1] ~= self.ltp_ih.trans:nrow() or + self.dim_in[2] ~= self.ltp_hh.trans:nrow() then nerv.error("mismatching dimensions of ltp and input") end if (self.dim_out[1] ~= self.bp.trans:ncol()) then nerv.error("mismatching dimensions of bp and output") end - self.ltp_hh_grad = self.ltp_hh.trans:create() self.ltp_hh:train_init() + self.ltp_ih:train_init() self.bp:train_init() end @@ -42,39 +48,19 @@ function Recurrent:batch_resize(batch_size) end function Recurrent:update(bp_err, input, output) - if self.direct_update == true then - local ltp_hh = self.ltp_hh.trans - local bp = self.bp.trans - local gconf = self.gconf - if (gconf.momentum > 0) then - -- momentum gain - local mmt_gain = 1.0 / (1.0 - gconf.momentum) - local n = input[1]:nrow() * mmt_gain - -- update corrections (accumulated errors) - self.ltp_hh.correction:mul(input[2], bp_err[1], 1.0, gconf.momentum, 'T', 'N') - self.bp.correction:add(self.bp.correction, bp_err[1]:colsum(), gconf.momentum, 1.0) - -- perform update and weight decay - ltp_hh:add(ltp_hh, self.ltp_hh.correction, 1.0 - gconf.lrate * gconf.wcost / gconf.batch_size, - gconf.lrate / n) - bp:add(bp, self.bp.correction, 1.0 - gconf.lrate * gconf.wcost / gconf.batch_size, - gconf.lrate / n) - else - ltp_hh:mul(input[2], bp_err[1], - gconf.lrate / gconf.batch_size, 1.0 - gconf.wcost * gconf.lrate / gconf.batch_size, 'T', 'N') - bp:add(bp, bp_err[1]:colsum(), 1.0 - gconf.lrate * gconf.wcost / gconf.batch_size, - gconf.lrate / gconf.batch_size) - end - else - --self.ltp_hh_grad:mul(input[2], bp_err[1], 1.0, 0.0, 'T', 'N') - self.ltp_hh:update_by_err_input(bp_err[1], input[2]) - self.bp:update_by_gradient(bp_err[1]:colsum()) - end + self.ltp_ih:update_by_err_input(bp_err[1], input[1]) + self.ltp_hh:update_by_err_input(bp_err[1], input[2]) + self.bp:update_by_gradient(bp_err[1]:colsum()) end function Recurrent:propagate(input, output) - output[1]:copy_fromd(input[1]) + output[1]:mul(input[1], self.ltp_ih.trans, 1.0, 0.0, 'N', 'N') output[1]:mul(input[2], self.ltp_hh.trans, 1.0, 1.0, 'N', 'N') output[1]:add_row(self.bp.trans, 1.0) end function Recurrent:back_propagate(bp_err, next_bp_err, input, output) - next_bp_err[1]:copy_fromd(bp_err[1]) + next_bp_err[1]:mul(bp_err[1], self.ltp_ih.trans, 1.0, 0.0, 'N', 'T') next_bp_err[2]:mul(bp_err[1], self.ltp_hh.trans, 1.0, 0.0, 'N', 'T') --[[ for i = 0, next_bp_err[2]:nrow() - 1 do @@ -84,11 +70,11 @@ function Recurrent:back_propagate(bp_err, next_bp_err, input, output) end end ]]-- - if (self.clip ~= nil) then - next_bp_err[2]:clip(- self.clip, self.clip) + if self.clip ~= nil then + next_bp_err[2]:clip(-self.clip, self.clip) end end function Recurrent:get_params() - return nerv.ParamRepo({self.ltp_hh, self.bp}) + return nerv.ParamRepo({self.ltp_ih, self.ltp_hh, self.bp}) end |