fix bug of momentum & update mse layer

4 files changed, 49 insertions, 42 deletions

diff --git a/nerv/layer/affine.lua b/nerv/layer/affine.lua
index 38743aa..a05ae17 100644
--- a/nerv/layer/affine.lua
+++ b/nerv/layer/affine.lua
@@ -25,7 +25,9 @@ end
 
 function MatrixParam:train_init()
     self.correction = self.trans:create()
+    self.correction_acc = self.correction:create()
     self.correction:fill(0)
+    self.correction_acc:fill(0)
 end
 
 function MatrixParam:copy(copier)
@@ -34,46 +36,37 @@ function MatrixParam:copy(copier)
     return target
 end
 
-function MatrixParam:_update_by_gradient(gradient, alpha, beta)
+function MatrixParam:_update(alpha, beta)
     local gconf = self.gconf
     -- momentum gain
     local mmt_gain = 1.0 / (1.0 - gconf.momentum)
     local n = gconf.batch_size * mmt_gain
     -- perform update
     if gconf.momentum > 0 then
-        self.correction:add(self.correction, gradient, gconf.momentum, 1.0)
+        self.correction:add(self.correction, self.correction_acc, 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)
+        self.trans:add(self.trans, self.correction_acc, alpha, -gconf.lrate / n * beta)
     end
+    self.correction_acc:fill(0)
 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 = 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
+function MatrixParam:back_propagate_by_gradient(gradient)
+    self.correction_acc:add(self.correction_acc, gradient, 1.0, 1.0)
 end
 
-function MatrixParam:update_by_gradient(gradient)
-    self:_update_by_gradient(gradient, 1.0, 1.0)
+function MatrixParam:back_propagate_by_err_input(err, input)
+    self.correction_acc:mul(input, err, 1.0, 1.0, 'T', 'N')
 end
 
-function MatrixParam:update_by_err_input(err, input)
-    self:_update_by_err_input(err, input, 1.0, 1.0)
+function MatrixParam:update_by_gradient()
+    self:_update(1.0, 1.0)
 end
 
-function LinearTransParam:update_by_err_input(err, input)
+function MatrixParam:update_by_err_input()
     local gconf = self.gconf
     local l2 = 1 - gconf.lrate * gconf.wcost
-    self:_update_by_err_input(err, input, l2, l2)
+    self:_update(l2, l2)
 end
 
 --- A fully-connected linear transform layer.
@@ -121,11 +114,11 @@ function AffineLayer:batch_resize(batch_size)
     -- do nothing
 end
 
-function AffineLayer:update(bp_err, input, output)
+function AffineLayer:update()
     for i = 1, #self.dim_in do
-        self["ltp" .. i]:update_by_err_input(bp_err[1], input[i])
+        self["ltp" .. i]:update_by_err_input()
     end
-    self.bp:update_by_gradient(bp_err[1]:colsum())
+    self.bp:update_by_gradient()
 end
 
 function AffineLayer:propagate(input, output)
@@ -141,7 +134,9 @@ end
 function AffineLayer:back_propagate(bp_err, next_bp_err, input, output)
     for i = 1, #self.dim_in do
         next_bp_err[i]:mul(bp_err[1], self["ltp" .. i].trans, 1.0, 0.0, 'N', 'T')
+        self["ltp" .. i]:back_propagate_by_err_input(bp_err[1], input[i])
     end
+    self.bp:back_propagate_by_gradient(bp_err[1]:colsum())
 end
 
 function AffineLayer:get_params()
diff --git a/nerv/layer/lstm_gate.lua b/nerv/layer/lstm_gate.lua
index e690721..9d79b04 100644
--- a/nerv/layer/lstm_gate.lua
+++ b/nerv/layer/lstm_gate.lua
@@ -60,18 +60,19 @@ function LSTMGateLayer:back_propagate(bp_err, next_bp_err, input, output)
     self.err_bakm:sigmoid_grad(bp_err[1], output[1])
     for i = 1, #self.dim_in do
         next_bp_err[i]:mul(self.err_bakm, self["ltp" .. i].trans, 1.0, 0.0, 'N', 'T')
+        self["ltp" .. i]:back_propagate_by_err_input(self.err_bakm, input[i])
     end
+    self.bp:back_propagate_by_gradient(self.err_bakm:colsum())
 end
 
-function LSTMGateLayer:update(bp_err, input, output)
-    self.err_bakm:sigmoid_grad(bp_err[1], output[1])
+function LSTMGateLayer:update()
     for i = 1, #self.dim_in do
-        self["ltp" .. i]:update_by_err_input(self.err_bakm, input[i])
+        self["ltp" .. i]:update_by_err_input()
         if self.param_type[i] == 'D' then
             self["ltp" .. i].trans:diagonalize()
         end
     end
-    self.bp:update_by_gradient(self.err_bakm:colsum())
+    self.bp:update_by_gradient()
 end
 
 function LSTMGateLayer:get_params()
diff --git a/nerv/layer/mse.lua b/nerv/layer/mse.lua
index 458d086..c1ea596 100644
--- a/nerv/layer/mse.lua
+++ b/nerv/layer/mse.lua
@@ -9,23 +9,28 @@ function MSELayer:bind_params()
     -- do nothing
 end
 
-function MSELayer:init(batch_size)
+function MSELayer:init(batch_size, chunk_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.scale = 1.0 / self.dim_in[1]
     self.total_mse = 0.0
     self.total_frames = 0
     self.mse = self.mat_type(batch_size, self.dim_in[1])
     self.mse_sum = self.mat_type(batch_size, 1)
-    self.diff = self.mse:create()
+    self.diff = {}
+    for t = 1, chunk_size do
+        self.diff[t] = self.mse:create()
+    end
 end
 
 function MSELayer:batch_resize(batch_size)
     if self.mse:nrow() ~= batch_resize then
         self.mse = self.mat_type(batch_size, self.dim_in[1])
         self.mse_sum = self.mat_type(batch_size, 1)
-        self.diff = self.mse:create()
+        for t = 1, chunk_size do
+            self.diff[t] = self.mse:create()
+        end
     end
 end
 
@@ -33,24 +38,32 @@ function MSELayer:update(bp_err, input, output)
     -- no params, therefore do nothing
 end
 
-function MSELayer:propagate(input, output)
+function MSELayer:propagate(input, output, t)
+    if t == nil then
+        t = 1
+    end
     local mse = self.mse
     local mse_sum = self.mse_sum
+    local diff = self.diff[t]
     mse:add(input[1], input[2], 1.0, -1.0)
-    self.diff:copy_from(mse)
+    mse:set_values_by_mask(self.gconf.mask[t], 0)
+    diff:copy_from(mse)
     mse:mul_elem(mse, mse)
-    mse_sum:add(mse_sum, mse:rowsum(mse), 0.0, self.scale)
+    mse_sum:add(mse_sum, mse:rowsum(), 0.0, self.scale * 0.5)
     if output[1] ~= nil then
         output[1]:copy_from(mse_sum)
     end
     self.total_mse = self.total_mse + mse_sum:colsum()[0][0]
-    self.total_frames = self.total_frames + mse_sum:nrow()
+    self.total_frames = self.total_frames + self.gconf.mask[t]:colsum()[0][0]
 end
 
 -- NOTE: must call propagate before back_propagate
-function MSELayer:back_propagate(bp_err, next_bp_err, input, output)
+function MSELayer:back_propagate(bp_err, next_bp_err, input, output, t)
+    if t == nil then
+        t = 1
+    end
     local nbe = next_bp_err[1]
-    nbe:add(nbe, self.diff, 0.0, 2 * self.scale)
+    nbe:add(nbe, self.diff[t], 0.0, self.scale)
     if bp_err[1] ~= nil then
         nbe:scale_rows_by_col(bp_err[1])
     end
diff --git a/nerv/nn/network.lua b/nerv/nn/network.lua
index bb03be4..cf6a4d3 100644
--- a/nerv/nn/network.lua
+++ b/nerv/nn/network.lua
@@ -615,10 +615,8 @@ function network:back_propagate()
 end
 
 function network:update()
-    for t = 1, self.max_length do
-        for i = 1, #self.layers do
-            self.layers[i]:update(self.err_input[t][i], self.input[t][i], self.output[t][i], t)
-        end
+    for i = 1, #self.layers do
+        self.layers[i]:update()
     end
 end