Merge branch 'master' into 'master'

add projection layer and lstmp layer



See merge request !1

4 files changed, 158 insertions, 0 deletions

diff --git a/nerv/Makefile b/nerv/Makefile
index dde8fe7..f74a92f 100644
--- a/nerv/Makefile
+++ b/nerv/Makefile
@@ -40,6 +40,7 @@ OBJS := $(CORE_OBJS) $(NERV_OBJS) $(LUAT_OBJS)
 LIBS := $(INST_LIBDIR)/libnerv.so $(LIB_PATH)/libnervcore.so $(LIB_PATH)/libluaT.so
 LUA_LIBS := matrix/init.lua io/init.lua init.lua \
 			layer/init.lua layer/affine.lua layer/sigmoid.lua layer/tanh.lua layer/softmax_ce.lua layer/softmax.lua \
+			layer/lstmp.lua layer/projection.lua \
 			layer/window.lua layer/bias.lua layer/combiner.lua layer/mse.lua \
 			layer/elem_mul.lua layer/lstm.lua layer/lstm_gate.lua layer/dropout.lua layer/gru.lua \
 			layer/graph.lua layer/rnn.lua layer/duplicate.lua layer/identity.lua \
diff --git a/nerv/layer/init.lua b/nerv/layer/init.lua
index 3a6cbcd..c893df3 100644
--- a/nerv/layer/init.lua
+++ b/nerv/layer/init.lua
@@ -152,6 +152,8 @@ nerv.include('gru.lua')
 nerv.include('rnn.lua')
 nerv.include('duplicate.lua')
 nerv.include('identity.lua')
+nerv.include('projection.lua')
+nerv.include('lstmp.lua')
 
 -- The following lines are for backward compatibility, and will be removed in
 -- the future. The use of these names are deprecated.
diff --git a/nerv/layer/lstmp.lua b/nerv/layer/lstmp.lua
new file mode 100644
index 0000000..bbb2091
--- /dev/null
+++ b/nerv/layer/lstmp.lua
@@ -0,0 +1,91 @@
+local LSTMPLayer = nerv.class('nerv.LSTMPLayer', 'nerv.GraphLayer')
+
+function LSTMPLayer:__init(id, global_conf, layer_conf)
+    nerv.Layer.__init(self, id, global_conf, layer_conf)
+    self:check_dim_len(1, 1)
+
+    local din = layer_conf.dim_in[1]
+    local dcell = layer_conf.cell_dim
+    local dout = layer_conf.dim_out[1]
+
+    local pr = layer_conf.pr
+    if pr == nil then
+        pr = nerv.ParamRepo({}, self.loc_type)
+    end
+
+    local layers = {
+        ['nerv.CombinerLayer'] = {
+            mainCombine = {dim_in = {dcell, dcell}, dim_out = {dcell}, lambda = {1, 1}},
+        },
+        ['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 = {dcell}, dim_out = {dcell, dcell, dcell, dcell, dcell}},
+        },
+        ['nerv.AffineLayer'] = {
+            mainAffine = {dim_in = {din, dout}, dim_out = {dcell}, pr = pr},
+        },
+        ['nerv.TanhLayer'] = {
+            mainTanh = {dim_in = {dcell}, dim_out = {dcell}},
+            outputTanh = {dim_in = {dcell}, dim_out = {dcell}},
+        },
+        ['nerv.LSTMGateLayer'] = {
+            forgetGate = {dim_in = {din, dout, dcell}, dim_out = {dcell}, param_type = {'N', 'N', 'D'}, pr = pr},
+            inputGate = {dim_in = {din, dout, dcell}, dim_out = {dcell}, param_type = {'N', 'N', 'D'}, pr = pr},
+            outputGate = {dim_in = {din, dout, dcell}, dim_out = {dcell}, param_type = {'N', 'N', 'D'}, pr = pr},
+        },
+        ['nerv.ElemMulLayer'] = {
+            inputGateMul = {dim_in = {dcell, dcell}, dim_out = {dcell}},
+            forgetGateMul = {dim_in = {dcell, dcell}, dim_out = {dcell}},
+            outputGateMul = {dim_in = {dcell, dcell}, dim_out = {dcell}},
+        },
+        ['nerv.ProjectionLayer'] = {
+            projection = {dim_in = {dcell}, dim_out = {dout}, pr = pr},
+        },
+    }
+    
+    local connections = {
+        -- 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]', 'projection[1]', 0},
+        {'projection[1]', 'outputDup[1]', 0},
+        {'outputDup[5]', '<output>[1]', 0},
+    }
+
+    self:add_prefix(layers, connections)
+    local layer_repo = nerv.LayerRepo(layers, pr, global_conf)
+    self.lrepo = layer_repo
+    self:graph_init(layer_repo, connections)
+end
diff --git a/nerv/layer/projection.lua b/nerv/layer/projection.lua
new file mode 100644
index 0000000..d99401c
--- /dev/null
+++ b/nerv/layer/projection.lua
@@ -0,0 +1,64 @@
+local ProjectionLayer = nerv.class('nerv.ProjectionLayer', 'nerv.Layer')
+
+--- The constructor.
+function ProjectionLayer:__init(id, global_conf, layer_conf)
+    nerv.Layer.__init(self, id, global_conf, layer_conf)
+    self:check_dim_len(-1, 1) -- exactly one output, allow multiple inputs
+    self:bind_params()
+end
+
+function ProjectionLayer:bind_params()
+    for i = 1, #self.dim_in do
+        local pid = "ltp" .. i
+        local pid_list = i == 1 and {pid, "ltp"} or pid
+        self["ltp" .. i] = self:find_param(pid_list, self.lconf, self.gconf,
+                                            nerv.LinearTransParam,
+                                            {self.dim_in[i], self.dim_out[1]})
+    end
+    self.ltp = self.ltp1 -- alias of ltp1
+end
+
+function ProjectionLayer:init(batch_size)
+    for i = 1, #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 ProjectionLayer:batch_resize(batch_size)
+    -- do nothing
+end
+
+function ProjectionLayer:update()
+    for i = 1, #self.dim_in do
+        self["ltp" .. i]:update_by_err_input()
+    end
+end
+
+function ProjectionLayer:propagate(input, output)
+    -- apply linear transform
+    output[1]:mul(input[1], self.ltp1.trans, 1.0, 0.0, 'N', 'N')
+    for i = 2, #self.dim_in do
+        output[1]:mul(input[i], self["ltp" .. i].trans, 1.0, 1.0, 'N', 'N')
+    end
+end
+
+function ProjectionLayer: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
+end
+
+function ProjectionLayer:get_params()
+    local pr = nerv.ParamRepo({self.ltp1}, self.loc_type)
+    for i = 2, #self.dim_in do
+        pr:add(self["ltp" .. i].id, self["ltp" .. i])
+    end
+    return pr
+end