Merge upstream 'parameter update big-change'.

Merge remote-tracking branch 'upstream/master'

8 files changed, 97 insertions, 73 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
diff --git a/layer/bias.lua b/layer/bias.lua
index 8cd326b..c99274d 100644
--- a/layer/bias.lua
+++ b/layer/bias.lua
@@ -24,5 +24,5 @@ function BiasLayer:propagate(input, output)
 end
 
 function BiasLayer:get_params()
-    return {self.bias}
+    return nerv.ParamRepo({self.bias})
 end
diff --git a/layer/combiner.lua b/layer/combiner.lua
index 75e47e2..7bd7617 100644
--- a/layer/combiner.lua
+++ b/layer/combiner.lua
@@ -7,9 +7,15 @@ function CombinerLayer:__init(id, global_conf, layer_conf)
     self.dim_out = layer_conf.dim_out
     self.gconf = global_conf
     self:check_dim_len(#self.lambda, -1)
+    if #self.dim_in < 1 then
+        nerv.error("no input specified")
+    end
+    if #self.dim_out < 1 then
+        nerv.error("no output specified")
+    end
 end
 
-function CombinerLayer:init()
+function CombinerLayer:init(batch_size)
     local dim = self.dim_in[1]
     for i = 2, #self.dim_in do
         if self.dim_in[i] ~= dim then
@@ -21,6 +27,7 @@ function CombinerLayer:init()
             nerv.error("mismatching dimensions of inputs/outputs")
         end
     end
+    self.sum = self.gconf.cumat_type(batch_size, dim)
 end
 
 function CombinerLayer:update(bp_err, input, output)
@@ -32,24 +39,21 @@ function CombinerLayer:propagate(input, output)
         output[1]:add(output[1], input[i], 1.0, self.lambda[i])
     end
     for i = 2, #self.dim_out do
-        output[i]:copy_fromd(output[1]) 
+        output[i]:copy_fromd(output[1])
     end
 end
 
-function CombinerLayer:back_propagate(next_bp_err, bp_err, input, output)
-    local sum = bp_err[1]:create()
-    sum:fill(0)
-    for i = 1, #self.dim_out do
+function CombinerLayer:back_propagate(bp_err, next_bp_err, input, output)
+    local sum = self.sum
+    sum:copy_fromd(bp_err[1])
+    for i = 2, #self.dim_out do
         sum:add(sum, bp_err[i], 1.0, 1.0)
     end
     for i = 1, #self.dim_in do
-        local scale = nerv.CuMatrixFloat(sum:nrow(), 1)
-        scale:fill(self.lambda[i])
-        next_bp_err[i]:copy_fromd(sum)
-        next_bp_err[i]:scale_rows_by_col(scale)
+        next_bp_err[i]:add(next_bp_err[i], sum, 0.0, self.lambda[i])
     end
 end
 
 function CombinerLayer:get_params()
-    return {}
+    return nerv.ParamRepo({})
 end
diff --git a/layer/init.lua b/layer/init.lua
index 169427d..e39af94 100644
--- a/layer/init.lua
+++ b/layer/init.lua
@@ -15,11 +15,15 @@ function Param:set_info(info)
     self.info = info
 end
 
-function Param:read(pfhandle)
+function Param:read(handle)
     nerv.error_method_not_implemented()
 end
 
-function Param:write(pfhandle)
+function Param:write(handle)
+    nerv.error_method_not_implemented()
+end
+
+function Param:update(gradient)
     nerv.error_method_not_implemented()
 end
 
@@ -29,7 +33,7 @@ function Layer:__init(id, global_conf, layer_conf)
     nerv.error_method_not_implemented()
 end
 
-function Layer:init()
+function Layer:init(batch_size)
     nerv.error_method_not_implemented()
 end
 
@@ -41,7 +45,7 @@ function Layer:propagate(input, output)
     nerv.error_method_not_implemented()
 end
 
-function Layer:back_propagate(next_bp_err, bp_err, input, output)
+function Layer:back_propagate(bp_err, next_bp_err, input, output)
     nerv.error_method_not_implemented()
 end
 
diff --git a/layer/mse.lua b/layer/mse.lua
index da5b24d..9a97add 100644
--- a/layer/mse.lua
+++ b/layer/mse.lua
@@ -8,12 +8,16 @@ function MSELayer:__init(id, global_conf, layer_conf)
     self:check_dim_len(2, -1)
 end
 
-function MSELayer:init()
+function MSELayer:init(batch_size)
     if self.dim_in[1] ~= self.dim_in[2] then
         nerv.error("mismatching dimensions of previous network output and labels")
     end
+    self.scale = 1 / self.dim_in[1]
     self.total_mse = 0.0
     self.total_frames = 0
+    self.mse = self.gconf.cumat_type(batch_size, self.dim_in[1])
+    self.mse_sum = self.gconf.cumat_type(batch_size, 1)
+    self.diff = self.mse:create()
 end
 
 function MSELayer:update(bp_err, input, output)
@@ -21,32 +25,28 @@ function MSELayer:update(bp_err, input, output)
 end
 
 function MSELayer:propagate(input, output)
-    local mse = input[1]:create()
+    local mse = self.mse
+    local mse_sum = self.mse_sum
     mse:add(input[1], input[2], 1.0, -1.0)
-    self.diff = mse:create()
     self.diff:copy_fromd(mse)
     mse:mul_elem(mse, mse)
-    mse = mse:rowsum(mse)
-    local scale = nerv.CuMatrixFloat(mse:nrow(), 1)
-    scale:fill(1 / input[1]:ncol())
-    mse:scale_rows_by_col(scale)
+    mse_sum:add(mse_sum, mse:rowsum(mse), 0.0, self.scale)
     if output[1] ~= nil then
-        output[1]:copy_fromd(mse)
+        output[1]:copy_fromd(mse_sum)
     end
-    self.total_mse = self.total_mse + mse:colsum()[0]
-    self.total_frames = self.total_frames + mse:nrow()
+    self.total_mse = self.total_mse + mse_sum:colsum()[0]
+    self.total_frames = self.total_frames + mse_sum:nrow()
 end
 
 -- NOTE: must call propagate before back_propagate
-function MSELayer:back_propagate(next_bp_err, bp_err, input, output)
+function MSELayer:back_propagate(bp_err, next_bp_err, input, output)
     local nbe = next_bp_err[1]
-    nbe:copy_fromd(self.diff)
-    self.diff = nil
+    nbe:add(nbe, self.diff, 0.0, 2 * self.scale)
     if bp_err[1] ~= nil then
         nbe:scale_rows_by_col(bp_err[1])
     end
 end
 
 function MSELayer:get_params()
-    return {}
+    return nerv.ParamRepo({})
 end
diff --git a/layer/sigmoid.lua b/layer/sigmoid.lua
index dd10fb9..dfd09eb 100644
--- a/layer/sigmoid.lua
+++ b/layer/sigmoid.lua
@@ -22,10 +22,10 @@ function SigmoidLayer:propagate(input, output)
     output[1]:sigmoid(input[1])
 end
 
-function SigmoidLayer:back_propagate(next_bp_err, bp_err, input, output)
+function SigmoidLayer:back_propagate(bp_err, next_bp_err, input, output)
     next_bp_err[1]:sigmoid_grad(bp_err[1], output[1])
 end
 
 function SigmoidLayer:get_params()
-    return {}
+    return nerv.ParamRepo({})
 end
diff --git a/layer/softmax_ce.lua b/layer/softmax_ce.lua
index 7888540..daf891e 100644
--- a/layer/softmax_ce.lua
+++ b/layer/softmax_ce.lua
@@ -12,13 +12,15 @@ function SoftmaxCELayer:__init(id, global_conf, layer_conf)
     self:check_dim_len(2, -1) -- two inputs: nn output and label
 end
 
-function SoftmaxCELayer:init()
+function SoftmaxCELayer:init(batch_size)
     if not self.compressed and (self.dim_in[1] ~= self.dim_in[2]) then
         nerv.error("mismatching dimensions of previous network output and labels")
     end
     self.total_ce = 0.0
     self.total_correct = 0
     self.total_frames = 0
+    self.softmax = self.gconf.cumat_type(batch_size, self.dim_in[1])
+    self.ce = self.softmax:create()
 end
 
 function SoftmaxCELayer:update(bp_err, input, output)
@@ -26,12 +28,11 @@ function SoftmaxCELayer:update(bp_err, input, output)
 end
 
 function SoftmaxCELayer:propagate(input, output)
-    local soutput = input[1]:create()  -- temporary value for calc softmax
-    self.soutput = soutput
-    local classified = soutput:softmax(input[1])
-    local ce = soutput:create()
-    ce:log_elem(soutput)
+    local softmax = self.softmax
+    local ce = self.ce
+    local classified = softmax:softmax(input[1])
     local label = input[2]
+    ce:log_elem(softmax)
     if self.compressed then
         label = label:decompress(input[1]:ncol())
     end
@@ -42,26 +43,26 @@ function SoftmaxCELayer:propagate(input, output)
     end
     -- add total ce
     self.total_ce = self.total_ce - ce:colsum()[0]
-    self.total_frames = self.total_frames + soutput:nrow()
+    self.total_frames = self.total_frames + softmax:nrow()
     -- TODO: add colsame for uncompressed label
     if self.compressed then
         self.total_correct = self.total_correct + classified:colsame(input[2])[0]
     end
 end
 
-function SoftmaxCELayer:back_propagate(next_bp_err, bp_err, input, output)
+function SoftmaxCELayer:back_propagate(bp_err, next_bp_err, input, output)
     -- softmax output - label
     local label = input[2]
     if self.compressed then
         label = label:decompress(input[1]:ncol())
     end
     local nbe = next_bp_err[1]
-    nbe:add(self.soutput, label, 1.0, -1.0)
+    nbe:add(self.softmax, label, 1.0, -1.0)
     if bp_err[1] ~= nil then
         nbe:scale_rows_by_col(bp_err[1])
     end
 end
 
 function SoftmaxCELayer:get_params()
-    return {}
+    return nerv.ParamRepo({})
 end
diff --git a/layer/window.lua b/layer/window.lua
index 3a093f4..4e9a3b1 100644
--- a/layer/window.lua
+++ b/layer/window.lua
@@ -24,5 +24,5 @@ function WindowLayer:propagate(input, output)
 end
 
 function WindowLayer:get_params()
-    return {self.window}
+    return nerv.ParamRepo({self.window})
 end