aboutsummaryrefslogtreecommitdiff
path: root/nerv/layer
diff options
context:
space:
mode:
Diffstat (limited to 'nerv/layer')
-rw-r--r--nerv/layer/affine.lua110
-rw-r--r--nerv/layer/affine_recurrent.lua4
-rw-r--r--nerv/layer/elem_mul.lua38
-rw-r--r--nerv/layer/gate_fff.lua71
-rw-r--r--nerv/layer/init.lua23
-rw-r--r--nerv/layer/tanh.lua35
6 files changed, 225 insertions, 56 deletions
diff --git a/nerv/layer/affine.lua b/nerv/layer/affine.lua
index 3ba9408..566e9bc 100644
--- a/nerv/layer/affine.lua
+++ b/nerv/layer/affine.lua
@@ -5,7 +5,7 @@ local AffineLayer = nerv.class('nerv.AffineLayer', 'nerv.Layer')
function MatrixParam:read(handle)
self.trans = self.gconf.cumat_type.new_from_host(
- nerv.MMatrixFloat.load(handle))
+ self.gconf.mmat_type.load(handle))
end
function MatrixParam:write(handle)
@@ -17,74 +17,82 @@ function MatrixParam:train_init()
self.correction:fill(0)
end
-function MatrixParam:update_by_gradient(gradient)
+function MatrixParam:_update_by_gradient(gradient, 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)
- -- 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)
+ self.trans:add(self.trans, self.correction, alpha, -gconf.lrate / n * beta)
else
- self.trans:add(self.trans, gradient, 1.0 - gconf.lrate * gconf.wcost / gconf.batch_size, - gconf.lrate / gconf.batch_size)
+ self.trans:add(self.trans, gradient, alpha, -gconf.lrate / n * beta)
end
end
-function MatrixParam:update_by_err_input(err, input)
+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')
- -- 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)
+ self.trans:add(self.trans, self.correction, alpha, -gconf.lrate / n * beta)
else
- self.trans:mul(input, err, - gconf.lrate / gconf.batch_size, 1.0 - gconf.lrate * gconf.wcost / gconf.batch_size, 'T', 'N')
+ self.trans:mul(input, err, -gconf.lrate / n * beta, alpha, 'T', 'N')
end
end
---[[ --these updates are the same
-function LinearTransParam:update(gradient)
- MatrixParam.update(self, gradient)
- -- 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)
+function MatrixParam:update_by_gradient(gradient)
+ self:_update_by_gradient(gradient, 1.0, 1.0)
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 / gconf.batch_size)
+function MatrixParam:update_by_err_input(err, input)
+ self:_update_by_err_input(err, input, 1.0, 1.0)
+end
+
+function LinearTransParam:update_by_err_input(err, input)
+ local gconf = self.gconf
+ local l2 = 1 - gconf.lrate * gconf.wcost
+ self:_update_by_err_input(err, input, l2, l2)
end
-]]--
function AffineLayer:__init(id, global_conf, layer_conf)
self.id = id
- self.ltp = layer_conf.ltp
- self.bp = layer_conf.bp
self.dim_in = layer_conf.dim_in
self.dim_out = layer_conf.dim_out
+ self.ltp = self:find_param("ltp", layer_conf, global_conf, nerv.LinearTransParam, {self.dim_in[1], self.dim_out[1]}) --layer_conf.ltp
+ for i = 2, #self.dim_in do
+ self["ltp" .. i] = self:find_param("ltp" .. i, layer_conf, global_conf, nerv.LinearTransParam, {self.dim_in[i], self.dim_out[1]})
+ end
+ self.bp = self:find_param("bp", layer_conf, global_conf, nerv.BiasParam, {1, self.dim_out[1]}) --layer_conf.bp
self.gconf = global_conf
- self:check_dim_len(1, 1) -- exactly one input and one output
- self.direct_update = layer_conf.direct_update or global_conf.direct_update
+ self:check_dim_len(-1, 1) -- exactly one output, allow multiple inputs
end
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
+ self.bp:train_init()
if self.dim_in[1] ~= self.ltp.trans:nrow() then
nerv.error("mismatching dimensions of linear transform parameter and input")
end
if self.dim_out[1] ~= self.ltp.trans:ncol() then
nerv.error("mismatching dimensions of linear transform parameter and output")
end
- self.ltp_grad = self.ltp.trans:create()
self.ltp:train_init()
- self.bp:train_init()
+ for i = 2, #self.dim_in do
+ if self.dim_in[i] ~= self["ltp" .. i].trans:nrow() then
+ nerv.error("mismatching dimensions of linear transform parameter and input")
+ end
+ if self.dim_out[1] ~= self["ltp" .. i].trans:ncol() then
+ nerv.error("mismatching dimensions of linear transform parameter and output")
+ end
+ self["ltp" .. i]:train_init()
+ end
end
function AffineLayer:batch_resize(batch_size)
@@ -92,38 +100,32 @@ function AffineLayer:batch_resize(batch_size)
end
function AffineLayer:update(bp_err, input, output)
- 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:update_by_err_input(bp_err[1], input[1])
- self.bp:update_by_gradient(bp_err[1]:colsum())
+ self.ltp:update_by_err_input(bp_err[1], input[1])
+ for i = 2, #self.dim_in do
+ self["ltp" .. i]:update_by_err_input(bp_err[1], input[i])
end
+ self.bp:update_by_gradient(bp_err[1]:colsum())
end
function AffineLayer:propagate(input, output)
- -- apply linear transform
output[1]:mul(input[1], self.ltp.trans, 1.0, 0.0, 'N', 'N')
- -- add bias
+ for i = 2, #self.dim_in do
+ output[1]:mul(input[i], self["ltp" .. i].trans, 1.0, 1.0, 'N', 'N')
+ end
output[1]:add_row(self.bp.trans, 1.0)
end
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')
+ for i = 2, #self.dim_in do
+ next_bp_err[i]:mul(bp_err[1], self["ltp" .. i].trans, 1.0, 0.0, 'N', 'T')
+ end
end
function AffineLayer:get_params()
- return nerv.ParamRepo({self.ltp, self.bp})
+ local pr = nerv.ParamRepo({self.ltp, self.bp})
+ for i = 2, #self.dim_in do
+ pr:add(self["ltp" .. i].id, self["ltp" .. i])
+ end
+ return pr
end
diff --git a/nerv/layer/affine_recurrent.lua b/nerv/layer/affine_recurrent.lua
index da189e0..d537f4a 100644
--- a/nerv/layer/affine_recurrent.lua
+++ b/nerv/layer/affine_recurrent.lua
@@ -10,8 +10,8 @@ function Recurrent:__init(id, global_conf, layer_conf)
self.dim_out = layer_conf.dim_out
self.gconf = global_conf
- self.bp = layer_conf.bp
- self.ltp_hh = layer_conf.ltp_hh --from hidden to hidden
+ 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:check_dim_len(2, 1)
self.direct_update = layer_conf.direct_update
diff --git a/nerv/layer/elem_mul.lua b/nerv/layer/elem_mul.lua
new file mode 100644
index 0000000..c809d3e
--- /dev/null
+++ b/nerv/layer/elem_mul.lua
@@ -0,0 +1,38 @@
+local ElemMulLayer = nerv.class('nerv.ElemMulLayer', 'nerv.Layer')
+
+function ElemMulLayer:__init(id, global_conf, layer_conf)
+ self.id = id
+ self.dim_in = layer_conf.dim_in
+ self.dim_out = layer_conf.dim_out
+ self.gconf = global_conf
+
+ self:check_dim_len(2, 1) -- Element-multiply input[1] and input[2]
+end
+
+function ElemMulLayer:init(batch_size)
+ if self.dim_in[1] ~= self.dim_in[2] or
+ self.dim_in[1] ~= self.dim_out[1] then
+ nerv.error("dim_in and dim_out mismatch for ElemMulLayer")
+ end
+end
+
+function ElemMulLayer:batch_resize(batch_size)
+ --do nothing
+end
+
+function ElemMulLayer:propagate(input, output)
+ output[1]:mul_elem(input[1], input[2])
+end
+
+function ElemMulLayer:back_propagate(bp_err, next_bp_err, input, output)
+ next_bp_err[1]:mul_elem(bp_err[1], input[2])
+ next_bp_err[2]:mul_elem(bp_err[1], input[1])
+end
+
+function ElemMulLayer:update(bp_err, input, output)
+ --do nothing
+end
+
+function ElemMulLayer:get_params()
+ return nerv.ParamRepo({})
+end
diff --git a/nerv/layer/gate_fff.lua b/nerv/layer/gate_fff.lua
new file mode 100644
index 0000000..751dde1
--- /dev/null
+++ b/nerv/layer/gate_fff.lua
@@ -0,0 +1,71 @@
+local GateFFFLayer = nerv.class('nerv.GateFFFLayer', 'nerv.Layer')
+
+function GateFFFLayer:__init(id, global_conf, layer_conf)
+ self.id = id
+ self.dim_in = layer_conf.dim_in
+ self.dim_out = layer_conf.dim_out
+ self.gconf = global_conf
+
+ self.ltp1 = self:find_param("ltp1", layer_conf, global_conf, nerv.LinearTransParam, {self.dim_in[1], self.dim_out[1]}) --layer_conf.ltp
+ self.ltp2 = self:find_param("ltp2", layer_conf, global_conf, nerv.LinearTransParam, {self.dim_in[2], self.dim_out[1]}) --layer_conf.ltp
+ self.ltp3 = self:find_param("ltp3", layer_conf, global_conf, nerv.LinearTransParam, {self.dim_in[3], self.dim_out[1]}) --layer_conf.ltp
+ self.bp = self:find_param("bp", layer_conf, global_conf, nerv.BiasParam, {1, self.dim_out[1]})--layer_conf.bp
+
+ self:check_dim_len(3, 1) -- exactly one input and one output
+end
+
+function GateFFFLayer:init(batch_size)
+ if self.ltp1.trans:ncol() ~= self.bp.trans:ncol() or
+ self.ltp2.trans:ncol() ~= self.bp.trans:ncol() or
+ self.ltp3.trans:ncol() ~= self.bp.trans:ncol() then
+ nerv.error("mismatching dimensions of linear transform and bias paramter")
+ end
+ if self.dim_in[1] ~= self.ltp1.trans:nrow() or
+ self.dim_in[2] ~= self.ltp2.trans:nrow() or
+ self.dim_in[3] ~= self.ltp3.trans:nrow() then
+ nerv.error("mismatching dimensions of linear transform parameter and input")
+ end
+ if self.dim_out[1] ~= self.ltp1.trans:ncol() then
+ nerv.error("mismatching dimensions of linear transform parameter and output")
+ end
+ self.ltp1:train_init()
+ self.ltp2:train_init()
+ self.ltp3:train_init()
+ self.bp:train_init()
+ self.err_bakm = self.gconf.cumat_type(batch_size, self.dim_out[1])
+end
+
+function GateFFFLayer:batch_resize(batch_size)
+ if self.err_m:nrow() ~= batch_size then
+ self.err_bakm = self.gconf.cumat_type(batch_size, self.dim_out[1])
+ end
+end
+
+function GateFFFLayer:propagate(input, output)
+ -- apply linear transform
+ output[1]:mul(input[1], self.ltp1.trans, 1.0, 0.0, 'N', 'N')
+ output[1]:mul(input[2], self.ltp2.trans, 1.0, 1.0, 'N', 'N')
+ output[1]:mul(input[3], self.ltp3.trans, 1.0, 1.0, 'N', 'N')
+ -- add bias
+ output[1]:add_row(self.bp.trans, 1.0)
+ output[1]:sigmoid(output[1])
+end
+
+function GateFFFLayer:back_propagate(bp_err, next_bp_err, input, output)
+ self.err_bakm:sigmoid_grad(bp_err[1], output[1])
+ next_bp_err[1]:mul(self.err_bakm, self.ltp1.trans, 1.0, 0.0, 'N', 'T')
+ next_bp_err[2]:mul(self.err_bakm, self.ltp2.trans, 1.0, 0.0, 'N', 'T')
+ next_bp_err[3]:mul(self.err_bakm, self.ltp3.trans, 1.0, 0.0, 'N', 'T')
+end
+
+function GateFFFLayer:update(bp_err, input, output)
+ self.err_bakm:sigmoid_grad(bp_err[1], output[1])
+ self.ltp1:update_by_err_input(self.err_bakm, input[1])
+ self.ltp2:update_by_err_input(self.err_bakm, input[2])
+ self.ltp3:update_by_err_input(self.err_bakm, input[3])
+ self.bp:update_by_gradient(self.err_bakm:colsum())
+end
+
+function GateFFFLayer:get_params()
+ return nerv.ParamRepo({self.ltp1, self.ltp2, self.ltp3, self.bp})
+end
diff --git a/nerv/layer/init.lua b/nerv/layer/init.lua
index 6861b0e..23606e1 100644
--- a/nerv/layer/init.lua
+++ b/nerv/layer/init.lua
@@ -70,8 +70,29 @@ function Layer:get_dim()
return self.dim_in, self.dim_out
end
+function Layer:find_param(pid, l_conf, gconf, p_type, p_dim)
+ if l_conf[pid] ~= nil then
+ nerv.info("Param [%s] of layer [%s] found in layer_conf.", pid, self.id)
+ return l_conf[pid]
+ end
+ local pid_g = self.id .. '_' .. pid --global identifier
+ local pr = l_conf.pr
+ local p
+ if pr ~= nil and pr:has_param(pid_g) == true then
+ nerv.info("Param [%s] of layer [%s] found in layer_conf.paramRepo.", pid, self.id)
+ p = pr:get_param(pid_g)
+ return p
+ end
+ nerv.info("Param [%s] of layer [%s] is not found in layer_conf or layer_conf.paramRepo, switch to auto-generate.", pid, self.id)
+ p = p_type(pid_g, gconf)
+ p.trans = gconf.cumat_type(unpack(p_dim))
+ p.trans:generate(gconf.param_random)
+ return p
+end
+
nerv.include('affine.lua')
nerv.include('sigmoid.lua')
+nerv.include('tanh.lua')
nerv.include('softmax_ce.lua')
nerv.include('bias.lua')
nerv.include('window.lua')
@@ -79,3 +100,5 @@ nerv.include('mse.lua')
nerv.include('combiner.lua')
nerv.include('affine_recurrent.lua')
nerv.include('softmax.lua')
+nerv.include('elem_mul.lua')
+nerv.include('gate_fff.lua')
diff --git a/nerv/layer/tanh.lua b/nerv/layer/tanh.lua
new file mode 100644
index 0000000..e1c32f2
--- /dev/null
+++ b/nerv/layer/tanh.lua
@@ -0,0 +1,35 @@
+local TanhLayer = nerv.class("nerv.TanhLayer", "nerv.Layer")
+
+function TanhLayer:__init(id, global_conf, layer_conf)
+ self.id = id
+ self.gconf = global_conf
+ self.dim_in = layer_conf.dim_in
+ self.dim_out = layer_conf.dim_out
+ self:check_dim_len(1, 1)
+end
+
+function TanhLayer:init()
+ if self.dim_in[1] ~= self.dim_out[1] then
+ nerv.error("mismatching dimensions of input and output")
+ end
+end
+
+function TanhLayer:batch_resize(batch_size)
+ -- do nothing
+end
+
+function TanhLayer:update(bp_err, input, output)
+ -- no params, therefore do nothing
+end
+
+function TanhLayer:propagate(input, output)
+ output[1]:tanh(input[1])
+end
+
+function TanhLayer:back_propagate(bp_err, next_bp_err, input, output)
+ next_bp_err[1]:tanh_grad(bp_err[1], output[1])
+end
+
+function TanhLayer:get_params()
+ return nerv.ParamRepo({})
+end
Powered by cgit, Themed by Ted Yin.