aboutsummaryrefslogtreecommitdiff
path: root/layer/affine.lua
diff options
context:
space:
mode:
authorcloudygoose <[email protected]>2015-06-21 10:50:46 +0800
committercloudygoose <[email protected]>2015-06-21 10:50:46 +0800
commitad2c4013dec71667d6bbd5e9a249df2085208507 (patch)
tree474c536ff4a7d757bb61915a1bfc7a073f40d09f /layer/affine.lua
parent839d938df0d83ec311c5d1299923c667adff6a87 (diff)
parentf3f4e74eb4dbb8829e5ee136ba4b0c0a7938b551 (diff)
Merge upstream 'parameter update big-change'.
Merge remote-tracking branch 'upstream/master'
Diffstat (limited to 'layer/affine.lua')
-rw-r--r--layer/affine.lua75
1 files changed, 45 insertions, 30 deletions
diff --git a/layer/affine.lua b/layer/affine.lua
index 2cd7acb..00cbcfb 100644
--- a/layer/affine.lua
+++ b/layer/affine.lua
@@ -3,13 +3,35 @@ 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(pcdata)
+function MatrixParam:read(handle)
self.trans = self.gconf.cumat_type.new_from_host(
- nerv.MMatrixFloat.load(pcdata))
+ nerv.MMatrixFloat.load(handle))
end
-function MatrixParam:write(pfhandle)
- self.trans:new_to_host():save(pfhandle)
+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(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)
+end
+
+function LinearTransParam: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)
end
function AffineLayer:__init(id, global_conf, layer_conf)
@@ -20,9 +42,10 @@ function AffineLayer:__init(id, global_conf, layer_conf)
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
end
-function AffineLayer:init()
+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
@@ -32,32 +55,24 @@ function AffineLayer:init()
if self.dim_out[1] ~= self.ltp.trans:ncol() then
nerv.error("mismatching dimensions of linear transform parameter and output")
end
-
- -- linear transform correction
- self.ltc = self.ltp.trans:create()
- self.ltc:fill(0)
- -- bias correction
- self.bc = self.bp.trans:create()
- self.bc:fill(0)
+ self.ltp_grad = self.ltp.trans:create()
+ self.ltp:train_init()
+ self.bp:train_init()
end
function AffineLayer:update(bp_err, input, output)
- local ltp = self.ltp.trans
- local bp = self.bp.trans
- local ltc = self.ltc
- local bc = self.bc
- 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)
- ltc:mul(input[1], bp_err[1], 1.0, gconf.momentum, 'T', 'N')
- bc:add(bc, bp_err[1]:colsum(), gconf.momentum, 1.0)
- -- perform update
- ltp:add(ltp, ltc, 1.0, -gconf.lrate / n)
- bp:add(bp, bc, 1.0, -gconf.lrate / n)
- -- weight decay
- ltp:add(ltp, ltp, 1.0, -gconf.lrate * gconf.wcost)
+ 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)
+ else
+ self.ltp_grad:mul(input[1], bp_err[1], 1.0, 0.0, 'T', 'N')
+ self.ltp:update(self.ltp_grad)
+ end
+ self.bp:update(bp_err[1]:colsum())
end
function AffineLayer:propagate(input, output)
@@ -67,10 +82,10 @@ function AffineLayer:propagate(input, output)
output[1]:add_row(self.bp.trans, 1.0)
end
-function AffineLayer:back_propagate(next_bp_err, bp_err, input, output)
+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 {self.ltp, self.bp}
+ return nerv.ParamRepo({self.ltp, self.bp})
end