diff options
Diffstat (limited to 'nerv/layer')
| -rw-r--r-- | nerv/layer/affine.lua | 70 | ||||
| -rw-r--r-- | nerv/layer/affine_recurrent.lua | 29 |
2 files changed, 68 insertions, 31 deletions
diff --git a/nerv/layer/affine.lua b/nerv/layer/affine.lua index 015ec3f..c24af16 100644 --- a/nerv/layer/affine.lua +++ b/nerv/layer/affine.lua @@ -19,20 +19,47 @@ end function MatrixParam:update(gradient) local gconf = self.gconf - self.correction:add(self.correction, gradient, gconf.momentum, 1.0) - -- momentum gain - local mmt_gain = 1.0 / (1.0 - gconf.momentum); - local n = self.gconf.batch_size * mmt_gain - -- perform update - self.trans:add(self.trans, self.correction, 1.0, -gconf.lrate / n) + if (gconf.momentum > 0) then + self.correction:add(self.correction, gradient, gconf.momentum, 1.0) + -- momentum gain + local mmt_gain = 1.0 / (1.0 - gconf.momentum); + local n = self.gconf.batch_size * mmt_gain + -- perform update + self.trans:add(self.trans, self.correction, 1.0-gconf.lrate*gconf.wcost/gconf.batch_size, -gconf.lrate/n) + else + self.trans:add(self.trans, gradient, 1.0-gconf.lrate*gconf.wcost/gconf.batch_size, -gconf.lrate/gconf.batch_size) + end end +function MatrixParam:updateEI(err, input) + local gconf = self.gconf + if (gconf.momentum > 0) then + self.correction:mul(input, err, 1.0, gconf.momentum, 'T', 'N') + -- momentum gain + local mmt_gain = 1.0 / (1.0 - gconf.momentum); + local n = self.gconf.batch_size * mmt_gain + -- perform update + self.trans:add(self.trans, self.correction, 1.0-gconf.lrate*gconf.wcost/gconf.batch_size, -gconf.lrate/n) + else + self.trans:mul(input, err, -gconf.lrate/gconf.batch_size, 1.0-gconf.lrate*gconf.wcost/gconf.batch_size, 'T', 'N') + end +end + +--[[ --these updates are the same function LinearTransParam:update(gradient) MatrixParam.update(self, gradient) - local gconf = self.gconf + -- local gconf = self.gconf + -- weight decay(put into MatrixParam:update) + -- self.trans:add(self.trans, self.trans, 1.0, -gconf.lrate * gconf.wcost / gconf.batch_size) +end + +function BiasParam:update(gradient) + MatrixParam.update(self, gradient) + --local gconf = self.gconf -- weight decay - self.trans:add(self.trans, self.trans, 1.0, -gconf.lrate * gconf.wcost) + -- self.trans:add(self.trans, self.trans, 1.0, -gconf.lrate * gconf.wcost / gconf.batch_size) end +]]-- function AffineLayer:__init(id, global_conf, layer_conf) self.id = id @@ -65,18 +92,25 @@ function AffineLayer:batch_resize(batch_size) end function AffineLayer:update(bp_err, input, output) - if self.direct_update then - self.ltp.correction:mul(input[1], bp_err[1], 1.0, gconf.momentum, 'T', 'N') - -- momentum gain - local mmt_gain = 1.0 / (1.0 - gconf.momentum); - local n = self.gconf.batch_size * mmt_gain - -- perform update - self.ltp.trans:add(self.ltp.trans, self.ltp.correction, 1.0, -gconf.lrate / n) + if (self.direct_update == true) then + local gconf = self.gconf + if (gconf.momentum > 0) then + self.ltp.correction:mul(input[1], bp_err[1], 1.0, gconf.momentum, 'T', 'N') + self.bp.correction:add(self.bp.correction, bp_err[1]:colsum(), gconf.momentum, 1) + -- momentum gain + local mmt_gain = 1.0 / (1.0 - gconf.momentum); + local n = self.gconf.batch_size * mmt_gain + -- perform update + self.ltp.trans:add(self.ltp.trans, self.ltp.correction, 1.0-gconf.lrate*gconf.wcost/gconf.batch_size, -gconf.lrate / n) + self.bp.trans:add(self.bp.trans, self.bp.correction, 1.0-gconf.lrate*gconf.wcost/gconf.batch_size, -gconf.lrate / n) + else + self.ltp.trans:mul(input[1], bp_err[1], -gconf.lrate / gconf.batch_size, 1.0-gconf.lrate*gconf.wcost/gconf.batch_size, 'T', 'N') + self.bp.trans:add(self.bp.trans, bp_err[1]:colsum(), 1.0-gconf.lrate*gconf.wcost/gconf.batch_size, -gconf.lrate / gconf.batch_size) + end else - self.ltp_grad:mul(input[1], bp_err[1], 1.0, 0.0, 'T', 'N') - self.ltp:update(self.ltp_grad) + self.ltp:updateEI(bp_err[1], input[1]) + self.bp:update(bp_err[1]:colsum()) end - self.bp:update(bp_err[1]:colsum()) end function AffineLayer:propagate(input, output) diff --git a/nerv/layer/affine_recurrent.lua b/nerv/layer/affine_recurrent.lua index 92d98e2..b465e95 100644 --- a/nerv/layer/affine_recurrent.lua +++ b/nerv/layer/affine_recurrent.lua @@ -46,20 +46,23 @@ function Recurrent:update(bp_err, input, output) local ltp_hh = self.ltp_hh.trans local bp = self.bp.trans local gconf = self.gconf - -- 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(bc, bp_err[1]:colsum(), gconf.momentum, 1.0) - -- perform update - ltp_hh:add(ltp_hh, self.ltp_hh.correction, 1.0, -gconf.lrate / n) - bp:add(bp, self.bp.correction, 1.0, -gconf.lrate / n) - -- weight decay - ltp_hh:add(ltp_hh, ltp_hh, 1.0, -gconf.lrate * gconf.wcost) + 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(self.ltp_hh_grad) + --self.ltp_hh_grad:mul(input[2], bp_err[1], 1.0, 0.0, 'T', 'N') + self.ltp_hh:updateEI(bp_err[1], input[2]) self.bp:update(bp_err[1]:colsum()) end end |