1 files changed, 77 insertions, 0 deletions
diff --git a/nerv/layer/lstm_gate.lua b/nerv/layer/lstm_gate.lua
new file mode 100644
index 0000000..1963eba
--- /dev/null
+++ b/nerv/layer/lstm_gate.lua
@@ -0,0 +1,77 @@
+local LSTMGateLayer = nerv.class('nerv.LSTMGateLayer', 'nerv.Layer')
+-- NOTE: this is a full matrix gate
+
+function LSTMGateLayer:__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
+
+    for i = 1, #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]})
+  
+    self:check_dim_len(-1, 1) --accept multiple inputs
+end
+
+function LSTMGateLayer:init(batch_size)
+    for i = 1, #self.dim_in do
+        if self["ltp" .. i].trans:ncol() ~= self.bp.trans:ncol() then
+            nerv.error("mismatching dimensions of linear transform and bias paramter")
+        end
+        if self.dim_in[i] ~= self["ltp" .. i].trans:nrow() then
+            nerv.error("mismatching dimensions of linear transform parameter and input")
+        end
+        self["ltp"..i]:train_init()
+    end
+    
+    if self.dim_out[1] ~= self.ltp1.trans:ncol() then
+        nerv.error("mismatching dimensions of linear transform parameter and output")
+    end
+    self.bp:train_init()
+    self.err_bakm = self.gconf.cumat_type(batch_size, self.dim_out[1])
+end
+
+function LSTMGateLayer: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 LSTMGateLayer: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
+    -- add bias
+    output[1]:add_row(self.bp.trans, 1.0)
+    output[1]:sigmoid(output[1])
+end
+
+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')
+    end
+end
+
+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])
+    end
+    self.bp:update_by_gradient(self.err_bakm:colsum())
+end
+
+function LSTMGateLayer:get_params()
+    local pr = nerv.ParamRepo({self.bp})
+    for i = 1, #self.dim_in do
+        pr:add(self["ltp" .. i].id, self["ltp" .. i])
+    end
+    return pr
+end