lstm & lstmp layer support multiple input

3 files changed, 72 insertions, 51 deletions

diff --git a/nerv/layer/lstm.lua b/nerv/layer/lstm.lua
index e568ee8..5d73ad2 100644
--- a/nerv/layer/lstm.lua
+++ b/nerv/layer/lstm.lua
@@ -2,9 +2,12 @@ local LSTMLayer = nerv.class('nerv.LSTMLayer', 'nerv.GraphLayer')
 
 function LSTMLayer:__init(id, global_conf, layer_conf)
     nerv.Layer.__init(self, id, global_conf, layer_conf)
-    self:check_dim_len(1, 1)
+    self:check_dim_len(-1, 1)
+    if #self.dim_in == 0 then
+        nerv.error('LSTM layer %s has no input', self.id)
+    end
 
-    local din = layer_conf.dim_in[1]
+    local din = layer_conf.dim_in
     local dout = layer_conf.dim_out[1]
 
     local pr = layer_conf.pr
@@ -17,15 +20,17 @@ function LSTMLayer:__init(id, global_conf, layer_conf)
             mainCombine = {dim_in = {dout, dout}, dim_out = {dout}, 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 = {dout}, dim_out = {dout, dout, dout, dout, dout}},
         },
         ['nerv.AffineLayer'] = {
-            mainAffine = {dim_in = {din, dout}, dim_out = {dout}, pr = pr},
-            forgetGate = {dim_in = {din, dout, dout}, dim_out = {dout}, param_type = {'N', 'N', 'D'}, pr = pr, activation = nerv.SigmoidLayer},
-            inputGate = {dim_in = {din, dout, dout}, dim_out = {dout}, param_type = {'N', 'N', 'D'}, pr = pr, activation = nerv.SigmoidLayer},
-            outputGate = {dim_in = {din, dout, dout}, dim_out = {dout}, param_type = {'N', 'N', 'D'}, pr = pr, activation = nerv.SigmoidLayer},
+            mainAffine = {dim_in = table.connect({dout}, din), dim_out = {dout}, pr = pr},
+            forgetGate = {dim_in = table.connect({dout, dout}, din), dim_out = {dout},
+                param_type = table.connect({'N', 'D'}, table.vector(#din, 'N')), pr = pr, activation = nerv.SigmoidLayer},
+            inputGate = {dim_in = table.connect({dout, dout}, din), dim_out = {dout},
+                param_type = table.connect({'N', 'D'}, table.vector(#din, 'N')), pr = pr, activation = nerv.SigmoidLayer},
+            outputGate = {dim_in = table.connect({dout, dout}, din), dim_out = {dout},
+                param_type = table.connect({'N', 'D'}, table.vector(#din, 'N')), pr = pr, activation = nerv.SigmoidLayer},
         },
         ['nerv.TanhLayer'] = {
             mainTanh = {dim_in = {dout}, dim_out = {dout}},
@@ -37,26 +42,29 @@ function LSTMLayer:__init(id, global_conf, layer_conf)
             outputGateMul = {dim_in = {dout, dout}, dim_out = {dout}},
         },
     }
+    for i = 1, #din do
+        layers['nerv.DuplicateLayer']['inputDup' .. i] = {dim_in = {din[i]}, dim_out = {din[i], din[i], din[i], din[i]}}
+    end
     
     local connections = {
         -- lstm input
-        {'<input>[1]', 'inputDup[1]', 0},
+        --{'<input>[1 .. n]', 'inputDup(1 .. n)[1]', 0},
 
         -- input gate
-        {'inputDup[1]', 'inputGate[1]', 0},
-        {'outputDup[1]', 'inputGate[2]', 1},
-        {'cellDup[1]', 'inputGate[3]', 1},
+        {'outputDup[1]', 'inputGate[1]', 1},
+        {'cellDup[1]', 'inputGate[2]', 1},
+        --{'inputDup(1 .. n)[1]', 'inputGate[3 .. n + 2]', 0},
 
         -- forget gate
-        {'inputDup[2]', 'forgetGate[1]', 0},
-        {'outputDup[2]', 'forgetGate[2]', 1},
-        {'cellDup[2]', 'forgetGate[3]', 1},
+        {'outputDup[2]', 'forgetGate[1]', 1},
+        {'cellDup[2]', 'forgetGate[2]', 1},
+        --{'inputDup(1 .. n)[2]', 'forgetGate[3 .. n + 2]', 0},
 
         -- lstm cell
         {'forgetGate[1]', 'forgetGateMul[1]', 0},
         {'cellDup[3]', 'forgetGateMul[2]', 1},
-        {'inputDup[3]', 'mainAffine[1]', 0},
-        {'outputDup[3]', 'mainAffine[2]', 1},
+        {'outputDup[3]', 'mainAffine[1]', 1},
+        --{'inputDup(1 .. n)[3]', 'mainAffine[2 .. n + 1]', 0},
         {'mainAffine[1]', 'mainTanh[1]', 0},
         {'inputGate[1]', 'inputGateMul[1]', 0},
         {'mainTanh[1]', 'inputGateMul[2]', 0},
@@ -65,9 +73,9 @@ function LSTMLayer:__init(id, global_conf, layer_conf)
         {'mainCombine[1]', 'cellDup[1]', 0},
 
         -- forget gate
-        {'inputDup[4]', 'outputGate[1]', 0},
-        {'outputDup[4]', 'outputGate[2]', 1},
-        {'cellDup[4]', 'outputGate[3]', 0},
+        {'outputDup[4]', 'outputGate[1]', 1},
+        {'cellDup[4]', 'outputGate[2]', 0},
+        --{'inputDup(1 .. n)[4]', 'outputGate[2 .. n + 1]', 0},
 
         -- lstm output
         {'cellDup[5]', 'outputTanh[1]', 0},
@@ -76,6 +84,13 @@ function LSTMLayer:__init(id, global_conf, layer_conf)
         {'outputGateMul[1]', 'outputDup[1]', 0},
         {'outputDup[5]', '<output>[1]', 0},
     }
+    for i = 1, #din do
+        table.insert(connections, {'<input>[' .. i .. ']', 'inputDup' .. i .. '[1]', 0})
+        table.insert(connections, {'inputDup' .. i .. '[1]', 'inputGate[' .. (i + 2) .. ']', 0})
+        table.insert(connections, {'inputDup' .. i .. '[2]', 'forgetGate[' .. (i + 2) .. ']', 0})
+        table.insert(connections, {'inputDup' .. i .. '[3]', 'mainAffine[' .. (i + 1) .. ']', 0})
+        table.insert(connections, {'inputDup' .. i .. '[4]', 'outputGate[' .. (i + 2) .. ']', 0})
+    end
 
     self:add_prefix(layers, connections)
     local layer_repo = nerv.LayerRepo(layers, pr, global_conf)
diff --git a/nerv/layer/lstmp.lua b/nerv/layer/lstmp.lua
index dc30797..49c9516 100644
--- a/nerv/layer/lstmp.lua
+++ b/nerv/layer/lstmp.lua
@@ -2,9 +2,12 @@ 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)
+    self:check_dim_len(-1, 1)
+    if #self.dim_in == 0 then
+        nerv.error('LSTMP layer %s has no input', self.id)
+    end
 
-    local din = layer_conf.dim_in[1]
+    local din = layer_conf.dim_in
     local dcell = layer_conf.cell_dim
     local dout = layer_conf.dim_out[1]
 
@@ -18,15 +21,17 @@ function LSTMPLayer:__init(id, global_conf, layer_conf)
             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},
-            forgetGate = {dim_in = {din, dout, dcell}, dim_out = {dcell}, param_type = {'N', 'N', 'D'}, pr = pr, activation = nerv.SigmoidLayer},
-            inputGate = {dim_in = {din, dout, dcell}, dim_out = {dcell}, param_type = {'N', 'N', 'D'}, pr = pr, activation = nerv.SigmoidLayer},
-            outputGate = {dim_in = {din, dout, dcell}, dim_out = {dcell}, param_type = {'N', 'N', 'D'}, pr = pr, activation = nerv.SigmoidLayer},
+            mainAffine = {dim_in = table.connect({dout}, din), dim_out = {dcell}, pr = pr},
+            forgetGate = {dim_in = table.connect({dout, dcell}, din), dim_out = {dcell},
+                param_type = table.connect({'N', 'D'}, table.vector(#din, 'N')), pr = pr, activation = nerv.SigmoidLayer},
+            inputGate = {dim_in = table.connect({dout, dcell}, din), dim_out = {dcell},
+                param_type = table.connect({'N', 'D'}, table.vector(#din, 'N')), pr = pr, activation = nerv.SigmoidLayer},
+            outputGate = {dim_in = table.connect({dout, dcell}, din), dim_out = {dcell},
+                param_type = table.connect({'N', 'D'}, table.vector(#din, 'N')), pr = pr, activation = nerv.SigmoidLayer},
             projection = {dim_in = {dcell}, dim_out = {dout}, pr = pr, no_bias = true},
         },
         ['nerv.TanhLayer'] = {
@@ -39,26 +44,29 @@ function LSTMPLayer:__init(id, global_conf, layer_conf)
             outputGateMul = {dim_in = {dcell, dcell}, dim_out = {dcell}},
         },
     }
-    
+    for i = 1, #din do
+        layers['nerv.DuplicateLayer']['inputDup' .. i] = {dim_in = {din[i]}, dim_out = {din[i], din[i], din[i], din[i]}}
+    end
+
     local connections = {
         -- lstm input
-        {'<input>[1]', 'inputDup[1]', 0},
+        --{'<input>[1 .. n]', 'inputDup(1 .. n)[1]', 0},
 
         -- input gate
-        {'inputDup[1]', 'inputGate[1]', 0},
-        {'outputDup[1]', 'inputGate[2]', 1},
-        {'cellDup[1]', 'inputGate[3]', 1},
+        {'outputDup[1]', 'inputGate[1]', 1},
+        {'cellDup[1]', 'inputGate[2]', 1},
+        --{'inputDup(1 .. n)[1]', 'inputGate[3 .. n + 2]', 0},
 
         -- forget gate
-        {'inputDup[2]', 'forgetGate[1]', 0},
-        {'outputDup[2]', 'forgetGate[2]', 1},
-        {'cellDup[2]', 'forgetGate[3]', 1},
+        {'outputDup[2]', 'forgetGate[1]', 1},
+        {'cellDup[2]', 'forgetGate[2]', 1},
+        --{'inputDup(1 .. n)[2]', 'forgetGate[3 .. n + 2]', 0},
 
         -- lstm cell
         {'forgetGate[1]', 'forgetGateMul[1]', 0},
         {'cellDup[3]', 'forgetGateMul[2]', 1},
-        {'inputDup[3]', 'mainAffine[1]', 0},
-        {'outputDup[3]', 'mainAffine[2]', 1},
+        {'outputDup[3]', 'mainAffine[1]', 1},
+        --{'inputDup(1 .. n)[3]', 'mainAffine[2 .. n + 1]', 0},
         {'mainAffine[1]', 'mainTanh[1]', 0},
         {'inputGate[1]', 'inputGateMul[1]', 0},
         {'mainTanh[1]', 'inputGateMul[2]', 0},
@@ -67,9 +75,9 @@ function LSTMPLayer:__init(id, global_conf, layer_conf)
         {'mainCombine[1]', 'cellDup[1]', 0},
 
         -- forget gate
-        {'inputDup[4]', 'outputGate[1]', 0},
-        {'outputDup[4]', 'outputGate[2]', 1},
-        {'cellDup[4]', 'outputGate[3]', 0},
+        {'outputDup[4]', 'outputGate[1]', 1},
+        {'cellDup[4]', 'outputGate[2]', 0},
+        --{'inputDup(1 .. n)[4]', 'outputGate[2 .. n + 1]', 0},
 
         -- lstm output
         {'cellDup[5]', 'outputTanh[1]', 0},
@@ -79,6 +87,13 @@ function LSTMPLayer:__init(id, global_conf, layer_conf)
         {'projection[1]', 'outputDup[1]', 0},
         {'outputDup[5]', '<output>[1]', 0},
     }
+    for i = 1, #din do
+        table.insert(connections, {'<input>[' .. i .. ']', 'inputDup' .. i .. '[1]', 0})
+        table.insert(connections, {'inputDup' .. i .. '[1]', 'inputGate[' .. (i + 2) .. ']', 0})
+        table.insert(connections, {'inputDup' .. i .. '[2]', 'forgetGate[' .. (i + 2) .. ']', 0})
+        table.insert(connections, {'inputDup' .. i .. '[3]', 'mainAffine[' .. (i + 1) .. ']', 0})
+        table.insert(connections, {'inputDup' .. i .. '[4]', 'outputGate[' .. (i + 2) .. ']', 0})
+    end
 
     self:add_prefix(layers, connections)
     local layer_repo = nerv.LayerRepo(layers, pr, global_conf)
diff --git a/nerv/layer/rnn.lua b/nerv/layer/rnn.lua
index fd6e753..63e0b55 100644
--- a/nerv/layer/rnn.lua
+++ b/nerv/layer/rnn.lua
@@ -4,12 +4,7 @@ function RNNLayer:__init(id, global_conf, layer_conf)
     nerv.Layer.__init(self, id, global_conf, layer_conf)
     self:check_dim_len(-1, 1)
     if #self.dim_in == 0 then
-        nerv.error('RNN Layer %s has no input', self.id)
-    end
-
-    self.activation = layer_conf.activation
-    if self.activation == nil then
-        self.activation = 'nerv.SigmoidLayer'
+        nerv.error('RNN layer %s has no input', self.id)
     end
 
     local din = layer_conf.dim_in
@@ -22,10 +17,7 @@ function RNNLayer:__init(id, global_conf, layer_conf)
 
     local layers = {
         ['nerv.AffineLayer'] = {
-            main = {dim_in = table.connect({dout}, din), dim_out = {dout}, pr = pr},
-        },
-        [self.activation] = {
-            activation = {dim_in = {dout}, dim_out = {dout}},
+            main = {dim_in = table.connect({dout}, din), dim_out = {dout}, pr = pr, activation = layer_conf.activation},
         },
         ['nerv.DuplicateLayer'] = {
             duplicate = {dim_in = {dout}, dim_out = {dout, dout}},
@@ -33,8 +25,7 @@ function RNNLayer:__init(id, global_conf, layer_conf)
     }
 
     local connections = {
-        {'main[1]', 'activation[1]', 0},
-        {'activation[1]', 'duplicate[1]', 0},
+        {'main[1]', 'duplicate[1]', 0},
         {'duplicate[1]', 'main[1]', 1},
         {'duplicate[2]', '<output>[1]', 0},
     }