add ParamRepo, LayerRepo, DAGLayer

11 files changed, 375 insertions, 10 deletions

diff --git a/Makefile b/Makefile
index 3325b4d..934235f 100644
--- a/Makefile
+++ b/Makefile
@@ -8,7 +8,8 @@ LIBS := libnerv.so
 LUA_LIBS := matrix/init.lua io/init.lua nerv.lua \
 			pl/utils.lua pl/compat.lua \
 			layer/init.lua layer/affine.lua layer/sigmoid.lua layer/softmax_ce.lua \
-			layer/window.lua layer/bias.lua
+			layer/window.lua layer/bias.lua \
+			nn/init.lua nn/layer_repo.lua nn/param_repo.lua nn/layer_dag.lua
 INCLUDE := -I build/luajit-2.0/include/luajit-2.0/ -DLUA_USE_APICHECK
 CUDA_BASE := /usr/local/cuda-6.5
 CUDA_INCLUDE := -I $(CUDA_BASE)/include/
@@ -18,7 +19,7 @@ CFLAGS := -Wall -Wextra
 OBJ_DIR := $(BUILD_DIR)/objs
 LUA_DIR := $(BUILD_DIR)/lua
 LIB_DIR := $(BUILD_DIR)/lib
-SUBDIR := matrix io layer examples pl
+SUBDIR := matrix io layer examples pl nn
 NVCC := $(CUDA_BASE)/bin/nvcc
 NVCC_FLAGS := -Xcompiler -fPIC,-Wall,-Wextra
 
diff --git a/examples/test_dnn_layers.lua b/examples/test_dnn_layers.lua
index 9be9d71..6e4d98d 100644
--- a/examples/test_dnn_layers.lua
+++ b/examples/test_dnn_layers.lua
@@ -50,15 +50,14 @@ for i = 0, 3 do
     sg:propagate(input2, output2)
     sm:propagate(input3, output3)
 
-
     -- back_propagate
     sm:back_propagate(err_output1, err_input1, input3, output3)
-    sm:update(err_input1, input3, output3)
-
     sg:back_propagate(err_output2, err_input2, input2, output2)
-    sg:update(err_input2, input2, output2)
-
     af:back_propagate(err_output3, err_input3, input1, output1)
+
+    -- update
+    sm:update(err_input1, input3, output3)
+    sg:update(err_input2, input2, output2)
     af:update(err_input3, input1, output1)
 
 
diff --git a/examples/test_nn_lib.lua b/examples/test_nn_lib.lua
new file mode 100644
index 0000000..fd7167a
--- /dev/null
+++ b/examples/test_nn_lib.lua
@@ -0,0 +1,63 @@
+require 'layer.affine'
+require 'layer.sigmoid'
+require 'layer.softmax_ce'
+
+gconf = {lrate = 0.8, wcost = 1e-6, momentum = 0.9,
+        mat_type = nerv.CuMatrixFloat,
+        batch_size = 10}
+
+param_repo = nerv.ParamRepo({"affine.param"})
+sublayer_repo = nerv.LayerRepo(
+    {
+        ["nerv.AffineLayer"] =
+        {
+            affine1 = {{ltp = "a", bp = "b"}, {dim_in = {429}, dim_out = {2048}}}
+        },
+        ["nerv.SigmoidLayer"] =
+        {
+            sigmoid1 = {{}, {dim_in = {2048}, dim_out = {2048}}}
+        },
+        ["nerv.SoftmaxCELayer"] =
+        {
+            softmax_ce1 = {{}, {dim_in = {2048, 2048}, dim_out = {}}}
+        }
+    }, param_repo, gconf)
+
+layer_repo = nerv.LayerRepo(
+    {
+        ["nerv.DAGLayer"] =
+        {
+            main = {{}, {
+                dim_in = {429, 2048}, dim_out = {},
+                sub_layers = sublayer_repo,
+                connections = {
+                    ["<input>[1]"] = "affine1[1]",
+                    ["affine1[1]"] = "sigmoid1[1]",
+                    ["sigmoid1[1]"] = "softmax_ce1[1]",
+                    ["<input>[2]"] = "softmax_ce1[2]"
+                }
+            }}
+        }
+    }, param_repo, gconf)
+
+df = nerv.ChunkFile("input.param", "r")
+label = nerv.CuMatrixFloat(10, 2048)
+label:fill(0)
+for i = 0, 9 do
+    label[i][i] = 1.0
+end
+
+input = {df:read_chunk("input", gconf).trans, label}
+output = {}
+err_input = {}
+err_output = {input[1]:create()}
+sm = sublayer_repo:get_layer("softmax_ce1")
+main = layer_repo:get_layer("main")
+main:init()
+for i = 0, 3 do
+    main:propagate(input, output)
+    main:back_propagate(err_output, err_input, input, output)
+    main:update(err_input, input, output)
+    nerv.utils.printf("cross entropy: %.8f\n", sm.total_ce)
+    nerv.utils.printf("frames: %.8f\n", sm.total_frames)
+end
diff --git a/io/init.lua b/io/init.lua
index 7c312f4..4a663a7 100644
--- a/io/init.lua
+++ b/io/init.lua
@@ -22,7 +22,7 @@ function nerv.ChunkFile:read_chunk(id, global_conf)
     if metadata == nil then
         nerv.error("chunk with id %s does not exist", id)
     end
-    local chunk_type = assert(loadstring("return " .. metadata.type))()
+    local chunk_type = nerv.get_type(metadata.type)
     local chunk = chunk_type(id, global_conf)
     chunk:set_info(metadata.info)
     chunk:read(self:get_chunkdata(id))
diff --git a/layer/init.lua b/layer/init.lua
index 4881cb7..c8c691b 100644
--- a/layer/init.lua
+++ b/layer/init.lua
@@ -46,8 +46,8 @@ function nerv.Layer:back_propagate(next_bp_err, bp_err, input, output)
 end
 
 function nerv.Layer:check_dim_len(len_in, len_out)
-    local expected_in = table.getn(self.dim_in)
-    local expected_out = table.getn(self.dim_out)
+    local expected_in = #self.dim_in
+    local expected_out = #self.dim_out
     if len_in > 0 and expected_in ~= len_in then
         nerv.error("layer %s expects %d inputs, %d given",
                     self.id, len_in, expected_in)
@@ -57,3 +57,13 @@ function nerv.Layer:check_dim_len(len_in, len_out)
                     self.id, len_out, expected_out)
     end
 end
+
+function nerv.Layer:get_dim()
+    return self.dim_in, self.dim_out
+end
+
+require 'layer.affine'
+require 'layer.sigmoid'
+require 'layer.softmax_ce'
+require 'layer.bias'
+require 'layer.window'
diff --git a/nerv.lua b/nerv.lua
index 00042a7..cb53f29 100644
--- a/nerv.lua
+++ b/nerv.lua
@@ -71,6 +71,11 @@ function table.tostring(tbl)
   return "{" .. table.concat(result, ",") .. "}"
 end
 
+function nerv.get_type(typename)
+    return assert(loadstring("return " .. typename))()
+end
+
 require 'matrix.init'
 require 'io.init'
 require 'layer.init'
+require 'nn.init'
diff --git a/nn/init.lua b/nn/init.lua
new file mode 100644
index 0000000..1bafa77
--- /dev/null
+++ b/nn/init.lua
@@ -0,0 +1,3 @@
+require 'nn.layer_repo'
+require 'nn.param_repo'
+require 'nn.layer_dag'
diff --git a/nn/layer_dag.lua b/nn/layer_dag.lua
new file mode 100644
index 0000000..8ea28a0
--- /dev/null
+++ b/nn/layer_dag.lua
@@ -0,0 +1,224 @@
+local DAGLayer = nerv.class("nerv.DAGLayer", "nerv.Layer")
+
+local function parse_id(str)
+    local id, port, _
+    _, _, id, port = string.find(str, "([a-zA-Z0-9_]+)%[([0-9]+)%]")
+    if id == nil or port == nil then
+        _, _, id, port = string.find(str, "(.+)%[([0-9]+)%]")
+        if not (id == "<input>" or id == "<output>") then
+            nerv.error("wrong format of connection id")
+        end
+    end
+    port = tonumber(port)
+    return id, port
+end
+
+local function discover(id, layers, layer_repo)
+    local ref = layers[id]
+    if id == "<input>" or id == "<output>" then
+        return nil
+    end
+    if ref == nil then
+        local layer = layer_repo:get_layer(id)
+        local dim_in, dim_out = layer:get_dim()
+        ref = {
+            layer = layer,
+            inputs = {},
+            outputs = {},
+            err_inputs = {},
+            err_outputs = {},
+            next_layers = {},
+            input_len = #dim_in,
+            output_len = #dim_out,
+            in_deg = 0,
+            visited = false
+        }
+        layers[id] = ref
+    end
+    return ref
+end
+
+function nerv.DAGLayer:__init(id, global_conf, layer_conf)
+    local layers = {}
+    local inputs = {}
+    local outputs = {}
+    local dim_in = layer_conf.dim_in
+    local dim_out = layer_conf.dim_out
+    for from, to in pairs(layer_conf.connections) do
+        local id_from, port_from = parse_id(from)
+        local id_to, port_to = parse_id(to)
+        local ref_from = discover(id_from, layers, layer_conf.sub_layers)
+        local ref_to = discover(id_to, layers, layer_conf.sub_layers)
+        local input_dim, output_dim, _
+        if ref_from and ref_from.outputs[port_from] ~= nil then
+            nerv.error("%s has already been attached", from)
+        end
+        if ref_to and ref_to.inputs[port_to] ~= nil then
+            nerv.error("%s has already been attached", to)
+        end
+        if id_from == "<input>" then
+            input_dim, _ = ref_to.layer:get_dim()
+            if dim_in[port_from] ~= input_dim[port_to] then
+                nerv.error("mismatching data dimension between %s and %s", from, to)
+            end
+            inputs[port_from] = {ref_to, port_to}
+            ref_to.inputs[port_to] = inputs -- just a place holder
+        elseif id_to == "<output>" then
+            _, output_dim = ref_from.layer:get_dim()
+            if output_dim[port_from] ~= dim_out[port_to] then
+                nerv.error("mismatching data dimension between %s and %s", from, to)
+            end
+            outputs[port_to] = {ref_from, port_from}
+            ref_from.outputs[port_from] = outputs -- just a place holder
+        else
+            _, output_dim = ref_from.layer:get_dim()
+            input_dim, _ = ref_to.layer:get_dim()
+            if output_dim[port_from] ~= input_dim[port_to] then
+                nerv.error("mismatching data dimension between %s and %s", from, to)
+            end
+            local mid = global_conf.mat_type(global_conf.batch_size,
+                                                output_dim[port_from])
+            local err_mid = mid:create() 
+
+            ref_from.outputs[port_from] = mid
+            ref_to.inputs[port_to] = mid
+
+            ref_from.err_inputs[port_from] = err_mid
+            ref_to.err_outputs[port_to] = err_mid
+
+            table.insert(ref_from.next_layers, ref_to) -- add edge
+            ref_to.in_deg = ref_to.in_deg + 1          -- increase the in-degree of the target layer
+        end
+    end
+    self.layers = layers
+    self.inputs = inputs
+    self.outputs = outputs
+    self.dim_in = dim_in
+    self.dim_out = dim_out
+end
+
+function nerv.DAGLayer:init(id) -- topology sort
+    local queue = {}
+    local l = 1
+    local r = 1
+    for id, ref in pairs(self.layers) do
+        if ref.in_deg == 0 then
+            table.insert(queue, ref)
+            nerv.utils.printf("adding source layer: %s\n", id)
+            r = r + 1
+        end
+    end
+    if l == r then
+        nerv.error("loop detected")
+    end
+    while l < r do
+        local cur = queue[l]
+        cur.visited = true
+        l = l + 1
+        for _, nl in pairs(cur.next_layers) do
+            nl.in_deg = nl.in_deg - 1 
+            if nl.in_deg == 0 then
+                table.insert(queue, nl)
+                r = r + 1
+            end
+        end
+    end
+    for i = 1, #queue do
+        nerv.utils.printf("queued layer: %s\n", queue[i].layer.id)
+    end
+    self.queue = queue
+    for id, ref in pairs(self.layers) do
+        -- check wether the graph is connected
+        if ref.visited == false then
+            nerv.utils.printf("warning: layer %s is ignored\n", id)
+        end
+        for i = 1, ref.input_len do
+            if ref.inputs[i] == nil then
+                nerv.error("dangling port %d of layer %s", i, id)
+            end
+        end
+        for i = 1, ref.output_len do
+            if ref.outputs[i] == nil then
+                nerv.error("dangling port %d of layer %s", i, id)
+            end
+        end
+        -- initialize sub layers
+        ref.layer:init()
+    end
+    for i = 1, #self.dim_in do
+        if self.inputs[i] == nil then
+            nerv.error("dangling port %d of layer <input>", i)
+        end
+    end
+    for i = 1, #self.dim_out do
+        if self.outputs[i] == nil then
+            nerv.error("dangling port %d of layer <output>", i)
+        end
+    end
+end
+
+function nerv.DAGLayer:set_inputs(input)
+    for i = 1, #self.dim_in do
+        local layer = self.inputs[i][1]
+        local port = self.inputs[i][2]
+        layer.inputs[port] = input[i]
+    end
+end
+
+function nerv.DAGLayer:set_outputs(output)
+    for i = 1, #self.dim_out do
+        local layer = self.outputs[i][1]
+        local port = self.outputs[i][2]
+        layer.outputs[port] = output[i]
+    end
+end
+
+function nerv.DAGLayer:set_err_inputs(bp_err)
+    for i = 1, #self.dim_out do
+        local layer = self.outputs[i][1]
+        local port = self.outputs[i][2]
+        layer.err_inputs[port] = bp_err[i]
+    end
+end
+
+function nerv.DAGLayer:set_err_outputs(next_bp_err)
+    for i = 1, #self.dim_in do
+        local layer = self.inputs[i][1]
+        local port = self.inputs[i][2]
+        layer.err_outputs[port] = next_bp_err[i]
+    end
+end
+
+function nerv.DAGLayer:update(bp_err, input, output)
+    self:set_err_inputs(bp_err)
+    self:set_inputs(input)
+    self:set_outputs(output)
+    for id, ref in pairs(self.queue) do
+        ref.layer:update(ref.err_inputs, ref.inputs, ref.outputs)
+    end
+end
+
+function nerv.DAGLayer:propagate(input, output)
+    self:set_inputs(input)
+    self:set_outputs(output)
+    for i = 1, #self.queue do
+        local ref = self.queue[i]
+        --[[
+        print(ref.inputs[1])
+        print(ref.outputs[1])
+        print(#ref.inputs, #ref.outputs)
+        --]]
+        ref.layer:propagate(ref.inputs, ref.outputs)
+    end
+end
+
+function nerv.DAGLayer:back_propagate(next_bp_err, bp_err, input, output)
+    self:set_err_outputs(next_bp_err)
+    self:set_err_inputs(bp_err)
+    self:set_inputs(input)
+    self:set_outputs(output)
+    for i = #self.queue, 1, -1 do
+        local ref = self.queue[i]
+        ref.layer:back_propagate(ref.err_outputs, ref.err_inputs, ref.inputs, ref.outputs)
+    end
+end
diff --git a/nn/layer_repo.lua b/nn/layer_repo.lua
new file mode 100644
index 0000000..b1d2248
--- /dev/null
+++ b/nn/layer_repo.lua
@@ -0,0 +1,34 @@
+local LayerRepo = nerv.class("nerv.LayerRepo")
+
+function LayerRepo:__init(layer_spec, param_repo, global_conf)
+    local layers = {}
+    for ltype, llist in pairs(layer_spec) do
+        local layer_type = nerv.get_type(ltype)
+        for id, spec in pairs(llist) do
+            if layers[id] ~= nil then
+                nerv.error("a layer with id %s already exists", id)
+            end
+            nerv.utils.printf("id: %s\n", id)
+            if type(spec[2]) ~= "table" then
+                nerv.error("layer config table is need")
+            end
+            layer_config = spec[2]
+            if type(spec[1]) ~= "table" then
+                nerv.error("parameter description table is needed")
+            end
+            for pname, pid in pairs(spec[1]) do
+                layer_config[pname] = param_repo:get_param(pid, global_conf)
+            end
+            layers[id] = layer_type(id, global_conf, layer_config)
+        end
+    end
+    self.layers = layers
+end
+
+function LayerRepo:get_layer(lid)
+    local layer = self.layers[lid]
+    if layer == nil then
+        nerv.error("layer with id %s not found", lid)
+    end
+    return layer
+end
diff --git a/nn/param_repo.lua b/nn/param_repo.lua
new file mode 100644
index 0000000..3e37c31
--- /dev/null
+++ b/nn/param_repo.lua
@@ -0,0 +1,26 @@
+local ParamRepo = nerv.class("nerv.ParamRepo")
+
+function ParamRepo:__init(param_files)
+    local param_table = {}
+    if type(param_files) ~= "table" then
+        nerv.error("param file table is need")
+    end
+    for i = 1, #param_files do
+        local pf = nerv.ChunkFile(param_files[i], "r")
+        for cid, cspec in pairs(pf.metadata) do
+            if param_table[cid] ~= nil then
+                nerv.error("conflicting chunk id in param files")
+            end
+            param_table[cid] = pf
+        end
+    end
+    self.param_table = param_table
+end
+
+function ParamRepo:get_param(pid, global_conf)
+    local pf = self.param_table[pid]
+    if pf == nil then
+        nerv.error("param with id %s not found", pid)
+    end
+    return pf:read_chunk(pid, global_conf)
+end
diff --git a/speech b/speech
-Subproject 821aec314824b89e9fe9c3ee467793a05ed89ee
+Subproject 0c6ca6a17f06821cd5d612f489ca6cb68c2c4d5