aboutsummaryrefslogtreecommitdiff
path: root/nerv/layer
diff options
context:
space:
mode:
Diffstat (limited to 'nerv/layer')
-rw-r--r--nerv/layer/dropout.lua11
-rw-r--r--nerv/layer/graph.lua2
-rw-r--r--nerv/layer/lstm.lua191
-rw-r--r--nerv/layer/lstm_gate.lua7
-rw-r--r--nerv/layer/rnn.lua20
5 files changed, 91 insertions, 140 deletions
diff --git a/nerv/layer/dropout.lua b/nerv/layer/dropout.lua
index 1a379c9..de0fb64 100644
--- a/nerv/layer/dropout.lua
+++ b/nerv/layer/dropout.lua
@@ -2,8 +2,7 @@ local DropoutLayer = nerv.class("nerv.DropoutLayer", "nerv.Layer")
function DropoutLayer:__init(id, global_conf, layer_conf)
nerv.Layer.__init(self, id, global_conf, layer_conf)
- self.rate = layer_conf.dropout_rate or global_conf.dropout_rate
- if self.rate == nil then
+ if self.gconf.dropout_rate == nil then
nerv.warning("[DropoutLayer:propagate] dropout rate is not set")
end
self:check_dim_len(1, 1) -- two inputs: nn output and label
@@ -41,12 +40,12 @@ function DropoutLayer:propagate(input, output, t)
if t == nil then
t = 1
end
- if self.rate then
+ if self.gconf.dropout_rate then
self.mask[t]:rand_uniform()
-- since we will lose a portion of the actvations, we multiply the
-- activations by 1 / (1 - rate) to compensate
- self.mask[t]:thres_mask(self.mask[t], self.rate,
- 0, 1 / (1.0 - self.rate))
+ self.mask[t]:thres_mask(self.mask[t], self.gconf.dropout_rate,
+ 0, 1 / (1.0 - self.gconf.dropout_rate))
output[1]:mul_elem(input[1], self.mask[t])
else
output[1]:copy_fromd(input[1])
@@ -61,7 +60,7 @@ function DropoutLayer:back_propagate(bp_err, next_bp_err, input, output, t)
if t == nil then
t = 1
end
- if self.rate then
+ if self.gconf.dropout_rate then
next_bp_err[1]:mul_elem(bp_err[1], self.mask[t])
else
next_bp_err[1]:copy_fromd(bp_err[1])
diff --git a/nerv/layer/graph.lua b/nerv/layer/graph.lua
index 5f42fca..68d5f51 100644
--- a/nerv/layer/graph.lua
+++ b/nerv/layer/graph.lua
@@ -112,7 +112,7 @@ function GraphLayer:graph_init(layer_repo, connections)
end
for i = 1, #ref.dim_out do
if ref.outputs[i] == nil then
- nerv.error('dangling output port %d os layer %s', i, id)
+ nerv.error('dangling output port %d of layer %s', i, id)
end
end
end
diff --git a/nerv/layer/lstm.lua b/nerv/layer/lstm.lua
index 641d5dc..56f674a 100644
--- a/nerv/layer/lstm.lua
+++ b/nerv/layer/lstm.lua
@@ -1,144 +1,85 @@
-local LSTMLayer = nerv.class('nerv.LSTMLayer', 'nerv.Layer')
+local LSTMLayer = nerv.class('nerv.LSTMLayer', 'nerv.GraphLayer')
function LSTMLayer:__init(id, global_conf, layer_conf)
- -- input1:x
- -- input2:h
- -- input3:c
nerv.Layer.__init(self, id, global_conf, layer_conf)
- -- prepare a DAGLayer to hold the lstm structure
+ self:check_dim_len(1, 1)
+
+ local din = layer_conf.dim_in[1]
+ local dout = layer_conf.dim_out[1]
+
local pr = layer_conf.pr
if pr == nil then
pr = nerv.ParamRepo({}, self.loc_type)
end
-
- local function ap(str)
- return self.id .. '.' .. str
- end
- local din1, din2, din3 = self.dim_in[1], self.dim_in[2], self.dim_in[3]
- local dout1, dout2, dout3 = self.dim_out[1], self.dim_out[2], self.dim_out[3]
- local layers = {
- ["nerv.CombinerLayer"] = {
- [ap("inputXDup")] = {dim_in = {din1},
- dim_out = {din1, din1, din1, din1},
- lambda = {1}},
- [ap("inputHDup")] = {dim_in = {din2},
- dim_out = {din2, din2, din2, din2},
- lambda = {1}},
-
- [ap("inputCDup")] = {dim_in = {din3},
- dim_out = {din3, din3, din3},
- lambda = {1}},
-
- [ap("mainCDup")] = {dim_in = {din3, din3},
- dim_out = {din3, din3, din3},
- lambda = {1, 1}},
+ local layers = {
+ ['nerv.CombinerLayer'] = {
+ mainCombine = {dim_in = {dout, dout}, dim_out = {dout}, lambda = {1, 1}},
},
- ["nerv.AffineLayer"] = {
- [ap("mainAffineL")] = {dim_in = {din1, din2},
- dim_out = {dout1},
- pr = pr},
+ ['nerv.DuplicateLayer'] = {
+ inputDup = {dim_in = {din}, dim_out = {din, din, din, din}},
+ outputDup = {dim_in = {dout}, dim_out = {dout, dout, dout, dout, dout}},
+ cellDup = {dim_in = {dout}, dim_out = {dout, dout, dout, dout, dout}},
},
- ["nerv.TanhLayer"] = {
- [ap("mainTanhL")] = {dim_in = {dout1}, dim_out = {dout1}},
- [ap("outputTanhL")] = {dim_in = {dout1}, dim_out = {dout1}},
+ ['nerv.AffineLayer'] = {
+ mainAffine = {dim_in = {din, dout}, dim_out = {dout}, pr = pr},
},
- ["nerv.LSTMGateLayer"] = {
- [ap("forgetGateL")] = {dim_in = {din1, din2, din3},
- dim_out = {din3}, pr = pr},
- [ap("inputGateL")] = {dim_in = {din1, din2, din3},
- dim_out = {din3}, pr = pr},
- [ap("outputGateL")] = {dim_in = {din1, din2, din3},
- dim_out = {din3}, pr = pr},
-
+ ['nerv.TanhLayer'] = {
+ mainTanh = {dim_in = {dout}, dim_out = {dout}},
+ outputTanh = {dim_in = {dout}, dim_out = {dout}},
},
- ["nerv.ElemMulLayer"] = {
- [ap("inputGMulL")] = {dim_in = {din3, din3},
- dim_out = {din3}},
- [ap("forgetGMulL")] = {dim_in = {din3, din3},
- dim_out = {din3}},
- [ap("outputGMulL")] = {dim_in = {din3, din3},
- dim_out = {din3}},
+ ['nerv.LSTMGateLayer'] = {
+ forgetGate = {dim_in = {din, dout, dout}, dim_out = {dout}, param_type = {'N', 'N', 'D'}, pr = pr},
+ inputGate = {dim_in = {din, dout, dout}, dim_out = {dout}, param_type = {'N', 'N', 'D'}, pr = pr},
+ outputGate = {dim_in = {din, dout, dout}, dim_out = {dout}, param_type = {'N', 'N', 'D'}, pr = pr},
+ },
+ ['nerv.ElemMulLayer'] = {
+ inputGateMul = {dim_in = {dout, dout}, dim_out = {dout}},
+ forgetGateMul = {dim_in = {dout, dout}, dim_out = {dout}},
+ outputGateMul = {dim_in = {dout, dout}, dim_out = {dout}},
},
}
- self.lrepo = nerv.LayerRepo(layers, pr, global_conf)
-
local connections = {
- ["<input>[1]"] = ap("inputXDup[1]"),
- ["<input>[2]"] = ap("inputHDup[1]"),
- ["<input>[3]"] = ap("inputCDup[1]"),
-
- [ap("inputXDup[1]")] = ap("mainAffineL[1]"),
- [ap("inputHDup[1]")] = ap("mainAffineL[2]"),
- [ap("mainAffineL[1]")] = ap("mainTanhL[1]"),
-
- [ap("inputXDup[2]")] = ap("inputGateL[1]"),
- [ap("inputHDup[2]")] = ap("inputGateL[2]"),
- [ap("inputCDup[1]")] = ap("inputGateL[3]"),
-
- [ap("inputXDup[3]")] = ap("forgetGateL[1]"),
- [ap("inputHDup[3]")] = ap("forgetGateL[2]"),
- [ap("inputCDup[2]")] = ap("forgetGateL[3]"),
-
- [ap("mainTanhL[1]")] = ap("inputGMulL[1]"),
- [ap("inputGateL[1]")] = ap("inputGMulL[2]"),
-
- [ap("inputCDup[3]")] = ap("forgetGMulL[1]"),
- [ap("forgetGateL[1]")] = ap("forgetGMulL[2]"),
-
- [ap("inputGMulL[1]")] = ap("mainCDup[1]"),
- [ap("forgetGMulL[1]")] = ap("mainCDup[2]"),
-
- [ap("inputXDup[4]")] = ap("outputGateL[1]"),
- [ap("inputHDup[4]")] = ap("outputGateL[2]"),
- [ap("mainCDup[3]")] = ap("outputGateL[3]"),
-
- [ap("mainCDup[2]")] = "<output>[2]",
- [ap("mainCDup[1]")] = ap("outputTanhL[1]"),
-
- [ap("outputTanhL[1]")] = ap("outputGMulL[1]"),
- [ap("outputGateL[1]")] = ap("outputGMulL[2]"),
-
- [ap("outputGMulL[1]")] = "<output>[1]",
+ -- lstm input
+ {'<input>[1]', 'inputDup[1]', 0},
+
+ -- input gate
+ {'inputDup[1]', 'inputGate[1]', 0},
+ {'outputDup[1]', 'inputGate[2]', 1},
+ {'cellDup[1]', 'inputGate[3]', 1},
+
+ -- forget gate
+ {'inputDup[2]', 'forgetGate[1]', 0},
+ {'outputDup[2]', 'forgetGate[2]', 1},
+ {'cellDup[2]', 'forgetGate[3]', 1},
+
+ -- lstm cell
+ {'forgetGate[1]', 'forgetGateMul[1]', 0},
+ {'cellDup[3]', 'forgetGateMul[2]', 1},
+ {'inputDup[3]', 'mainAffine[1]', 0},
+ {'outputDup[3]', 'mainAffine[2]', 1},
+ {'mainAffine[1]', 'mainTanh[1]', 0},
+ {'inputGate[1]', 'inputGateMul[1]', 0},
+ {'mainTanh[1]', 'inputGateMul[2]', 0},
+ {'inputGateMul[1]', 'mainCombine[1]', 0},
+ {'forgetGateMul[1]', 'mainCombine[2]', 0},
+ {'mainCombine[1]', 'cellDup[1]', 0},
+
+ -- forget gate
+ {'inputDup[4]', 'outputGate[1]', 0},
+ {'outputDup[4]', 'outputGate[2]', 1},
+ {'cellDup[4]', 'outputGate[3]', 0},
+
+ -- lstm output
+ {'cellDup[5]', 'outputTanh[1]', 0},
+ {'outputGate[1]', 'outputGateMul[1]', 0},
+ {'outputTanh[1]', 'outputGateMul[2]', 0},
+ {'outputGateMul[1]', 'outputDup[1]', 0},
+ {'outputDup[5]', '<output>[1]', 0},
}
- self.dag = nerv.DAGLayer(self.id, global_conf,
- {dim_in = self.dim_in,
- dim_out = self.dim_out,
- sub_layers = self.lrepo,
- connections = connections})
-
- self:check_dim_len(3, 2) -- x, h, c and h, c
-end
-
-function LSTMLayer:bind_params()
- local pr = layer_conf.pr
- if pr == nil then
- pr = nerv.ParamRepo({}, self.loc_type)
- end
- self.lrepo:rebind(pr)
-end
-
-function LSTMLayer:init(batch_size, chunk_size)
- self.dag:init(batch_size, chunk_size)
-end
-
-function LSTMLayer:batch_resize(batch_size, chunk_size)
- self.dag:batch_resize(batch_size, chunk_size)
-end
-
-function LSTMLayer:update(bp_err, input, output, t)
- self.dag:update(bp_err, input, output, t)
-end
-
-function LSTMLayer:propagate(input, output, t)
- self.dag:propagate(input, output, t)
-end
-
-function LSTMLayer:back_propagate(bp_err, next_bp_err, input, output, t)
- self.dag:back_propagate(bp_err, next_bp_err, input, output, t)
-end
-function LSTMLayer:get_params()
- return self.dag:get_params()
+ self:add_prefix(layers, connections)
+ local layer_repo = nerv.LayerRepo(layers, pr, global_conf)
+ self:graph_init(layer_repo, connections)
end
diff --git a/nerv/layer/lstm_gate.lua b/nerv/layer/lstm_gate.lua
index 7a27bab..e690721 100644
--- a/nerv/layer/lstm_gate.lua
+++ b/nerv/layer/lstm_gate.lua
@@ -3,6 +3,7 @@ local LSTMGateLayer = nerv.class('nerv.LSTMGateLayer', 'nerv.Layer')
function LSTMGateLayer:__init(id, global_conf, layer_conf)
nerv.Layer.__init(self, id, global_conf, layer_conf)
+ self.param_type = layer_conf.param_type
self:check_dim_len(-1, 1) --accept multiple inputs
self:bind_params()
end
@@ -12,6 +13,9 @@ function LSTMGateLayer:bind_params()
self["ltp" .. i] = self:find_param("ltp" .. i, self.lconf, self.gconf,
nerv.LinearTransParam,
{self.dim_in[i], self.dim_out[1]})
+ if self.param_type[i] == 'D' then
+ self["ltp" .. i].trans:diagonalize()
+ end
end
self.bp = self:find_param("bp", self.lconf, self.gconf,
nerv.BiasParam, {1, self.dim_out[1]})
@@ -63,6 +67,9 @@ function LSTMGateLayer:update(bp_err, input, output)
self.err_bakm:sigmoid_grad(bp_err[1], output[1])
for i = 1, #self.dim_in do
self["ltp" .. i]:update_by_err_input(self.err_bakm, input[i])
+ if self.param_type[i] == 'D' then
+ self["ltp" .. i].trans:diagonalize()
+ end
end
self.bp:update_by_gradient(self.err_bakm:colsum())
end
diff --git a/nerv/layer/rnn.lua b/nerv/layer/rnn.lua
index e59cf5b..0b5ccaa 100644
--- a/nerv/layer/rnn.lua
+++ b/nerv/layer/rnn.lua
@@ -4,6 +4,10 @@ function RNNLayer:__init(id, global_conf, layer_conf)
nerv.Layer.__init(self, id, global_conf, layer_conf)
self:check_dim_len(1, 1)
+ if layer_conf.activation == nil then
+ layer_conf.activation = 'nerv.SigmoidLayer'
+ end
+
local din = layer_conf.dim_in[1]
local dout = layer_conf.dim_out[1]
@@ -16,20 +20,20 @@ function RNNLayer:__init(id, global_conf, layer_conf)
['nerv.AffineLayer'] = {
main = {dim_in = {din, dout}, dim_out = {dout}, pr = pr},
},
- ['nerv.SigmoidLayer'] = {
- sigmoid = {dim_in = {dout}, dim_out = {dout}},
+ [layers.activation] = {
+ activation = {dim_in = {dout}, dim_out = {dout}},
},
['nerv.DuplicateLayer'] = {
- dup = {dim_in = {dout}, dim_out = {dout, dout}},
- }
+ duplicate = {dim_in = {dout}, dim_out = {dout, dout}},
+ },
}
local connections = {
{'<input>[1]', 'main[1]', 0},
- {'main[1]', 'sigmoid[1]', 0},
- {'sigmoid[1]', 'dup[1]', 0},
- {'dup[1]', 'main[2]', 1},
- {'dup[2]', '<output>[1]', 0},
+ {'main[1]', 'activation[1]', 0},
+ {'activation[1]', 'duplicate[1]', 0},
+ {'duplicate[1]', 'main[2]', 1},
+ {'duplicate[2]', '<output>[1]', 0},
}
self:add_prefix(layers, connections)
Powered by cgit, Themed by Ted Yin.