add recipe

7 files changed, 562 insertions, 0 deletions

diff --git a/lua/config.lua b/lua/config.lua
new file mode 100644
index 0000000..9d73b64
--- /dev/null
+++ b/lua/config.lua
@@ -0,0 +1,67 @@
+function get_global_conf()
+    local global_conf = {
+        lrate = 0.15,
+        wcost = 1e-5,
+        momentum = 0,
+        clip = 5,
+        cumat_type = nerv.CuMatrixFloat,
+        mmat_type = nerv.MMatrixFloat,
+        vocab_size = 10000,
+        nn_act_default = 0,
+        hidden_size = 300,
+        layer_num = 1,
+        chunk_size = 15,
+        batch_size = 20,
+        max_iter = 1,
+        param_random = function() return (math.random() / 5 - 0.1) end,
+        dropout = 0.5,
+        timer = nerv.Timer(),
+        pr = nerv.ParamRepo(),
+    }
+    return global_conf
+end
+
+function get_layers(global_conf)
+    local pr = global_conf.pr
+    local layers = {
+        ['nerv.LSTMLayer'] = {},
+        ['nerv.DropoutLayer'] = {},
+        ['nerv.SelectLinearLayer'] = {
+            ['select'] = {dim_in = {1}, dim_out = {global_conf.hidden_size}, vocab = global_conf.vocab_size, pr = pr},
+        },
+        ['nerv.CombinerLayer'] = {},
+        ['nerv.AffineLayer'] = {
+            output = {dim_in = {global_conf.hidden_size}, dim_out = {global_conf.vocab_size}, pr = pr}
+        },
+        ['nerv.SoftmaxCELayer'] = {
+            softmax = {dim_in = {global_conf.vocab_size, global_conf.vocab_size}, dim_out = {1}},
+        },
+    }
+    for i = 1, global_conf.layer_num do
+        layers['nerv.LSTMLayer']['lstm' .. i] = {dim_in = {global_conf.hidden_size, global_conf.hidden_size, global_conf.hidden_size}, dim_out = {global_conf.hidden_size, global_conf.hidden_size}, pr = pr}
+        layers['nerv.DropoutLayer']['dropout' .. i] = {dim_in = {global_conf.hidden_size}, dim_out = {global_conf.hidden_size}}
+        layers['nerv.CombinerLayer']['dup' .. i] = {dim_in = {global_conf.hidden_size}, dim_out = {global_conf.hidden_size, global_conf.hidden_size}, lambda = {1}}
+    end
+    return layers
+end
+
+function get_connections(global_conf)
+    local connections = {
+        {'<input>[1]', 'select[1]', 0},
+        {'select[1]', 'lstm1[1]', 0},
+        {'dropout' .. global_conf.layer_num .. '[1]', 'output[1]', 0},
+        {'output[1]', 'softmax[1]', 0},
+        {'<input>[2]', 'softmax[2]', 0},
+        {'softmax[1]', '<output>[1]', 0},
+    }
+    for i = 1, global_conf.layer_num do
+        table.insert(connections, {'lstm' .. i .. '[1]', 'dup' .. i .. '[1]', 0})
+        table.insert(connections, {'lstm' .. i .. '[2]', 'lstm' .. i .. '[3]', 1})
+        table.insert(connections, {'dup' .. i .. '[1]', 'lstm' .. i .. '[2]', 1})
+        table.insert(connections, {'dup' .. i .. '[2]', 'dropout' .. i .. '[1]', 0})
+        if i > 1 then
+            table.insert(connections, {'dropout' .. (i - 1) .. '[1]', 'lstm' .. i .. '[1]', 0})
+        end
+    end
+    return connections
+end
diff --git a/lua/main.lua b/lua/main.lua
new file mode 100644
index 0000000..684efac
--- /dev/null
+++ b/lua/main.lua
@@ -0,0 +1,43 @@
+nerv.include('reader.lua')
+nerv.include('timer.lua')
+nerv.include('config.lua')
+nerv.include(arg[1])
+
+local global_conf = get_global_conf()
+local timer = global_conf.timer
+
+timer:tic('IO')
+
+local data_path = 'nerv/nerv/examples/lmptb/PTBdata/'
+local train_reader = nerv.Reader(data_path .. 'vocab', data_path .. 'ptb.train.txt.adds')
+local val_reader = nerv.Reader(data_path .. 'vocab', data_path .. 'ptb.valid.txt.adds')
+
+local train_data = train_reader:get_all_batch(global_conf)
+local val_data = val_reader:get_all_batch(global_conf)
+
+local layers = get_layers(global_conf)
+local connections = get_connections(global_conf)
+
+local NN = nerv.NN(global_conf, train_data, val_data, layers, connections)
+
+timer:toc('IO')
+timer:check('IO')
+io.flush()
+
+timer:tic('global')
+local best_cv = 1e10
+for i = 1, global_conf.max_iter do
+    timer:tic('Epoch' .. i)
+    local train_ppl, val_ppl = NN:epoch()
+    if val_ppl < best_cv then
+        best_cv = val_ppl
+    else
+        global_conf.lrate = global_conf.lrate / 2.0
+    end
+    nerv.printf('Epoch %d: %f %f %f\n', i, global_conf.lrate, train_ppl, val_ppl)
+    timer:toc('Epoch' .. i)
+    timer:check('Epoch' .. i)
+    io.flush()
+end
+timer:toc('global')
+timer:check('global')
diff --git a/lua/network.lua b/lua/network.lua
new file mode 100644
index 0000000..6280f24
--- /dev/null
+++ b/lua/network.lua
@@ -0,0 +1,109 @@
+nerv.include('select_linear.lua')
+
+local nn = nerv.class('nerv.NN')
+
+function nn:__init(global_conf, train_data, val_data, layers, connections)
+    self.gconf = global_conf
+    self.network = self:get_network(layers, connections)
+    self.train_data = self:get_data(train_data)
+    self.val_data = self:get_data(val_data)
+end
+
+function nn:get_network(layers, connections)
+    self.gconf.dropout_rate = 0
+    local layer_repo = nerv.LayerRepo(layers, self.gconf.pr, self.gconf)
+    local graph = nerv.GraphLayer('graph', self.gconf, 
+        {dim_in = {1, self.gconf.vocab_size}, dim_out = {1}, 
+        layer_repo = layer_repo, connections = connections})
+    local network = nerv.Network('network', self.gconf, 
+        {network = graph, clip = self.gconf.clip})
+    network:init(self.gconf.batch_size, self.gconf.chunk_size)
+    return network
+end
+
+function nn:get_data(data)
+    local err_output = {}
+    local softmax_output = {}
+    local output = {}
+    for i = 1, self.gconf.chunk_size do
+        err_output[i] = self.gconf.cumat_type(self.gconf.batch_size, 1)
+        softmax_output[i] = self.gconf.cumat_type(self.gconf.batch_size, self.gconf.vocab_size)
+        output[i] = self.gconf.cumat_type(self.gconf.batch_size, 1)
+    end
+    local ret = {}
+    for i = 1, #data do
+        ret[i] = {}
+        ret[i].input = {}
+        ret[i].output = {}
+        ret[i].err_input = {}
+        ret[i].err_output = {}
+        for t = 1, self.gconf.chunk_size do
+            ret[i].input[t] = {}
+            ret[i].output[t] = {}
+            ret[i].err_input[t] = {}
+            ret[i].err_output[t] = {}
+            ret[i].input[t][1] = data[i].input[t]
+            ret[i].input[t][2] = data[i].output[t]
+            ret[i].output[t][1] = output[t]
+            local err_input = self.gconf.mmat_type(self.gconf.batch_size, 1)
+            for j = 1, self.gconf.batch_size do
+                if t <= data[i].seq_len[j] then
+                    err_input[j - 1][0] = 1
+                else
+                    err_input[j - 1][0] = 0
+                end
+            end
+            ret[i].err_input[t][1] = self.gconf.cumat_type.new_from_host(err_input)
+            ret[i].err_output[t][1] = err_output[t]
+            ret[i].err_output[t][2] = softmax_output[t]
+        end
+        ret[i].info = {}
+        ret[i].info.seq_length = data[i].seq_len
+        ret[i].info.new_seq = {}
+        for j = 1, self.gconf.batch_size do
+            if data[i].seq_start[j] then
+                table.insert(ret[i].info.new_seq, j)
+            end
+        end
+    end
+    return ret
+end
+
+function nn:process(data, do_train)
+    local total_err = 0
+    local total_frame = 0
+    for id = 1, #data do
+        if do_train then
+            self.gconf.dropout_rate = self.gconf.dropout
+        else
+            self.gconf.dropout_rate = 0
+        end
+        self.network:mini_batch_init(data[id].info)
+        local input = {}
+        for t = 1, self.gconf.chunk_size do
+            input[t] = {data[id].input[t][1], data[id].input[t][2]:decompress(self.gconf.vocab_size)}
+        end
+        self.network:propagate(input, data[id].output)
+        for t = 1, self.gconf.chunk_size do
+            local tmp = data[id].output[t][1]:new_to_host()
+            for i = 1, self.gconf.batch_size do
+                if t <= data[id].info.seq_length[i] then
+                    total_err = total_err + math.log10(math.exp(tmp[i - 1][0]))
+                    total_frame = total_frame + 1
+                end
+            end
+        end
+        if do_train then
+            self.network:back_propagate(data[id].err_input, data[id].err_output, input, data[id].output)
+            self.network:update(data[id].err_input, input, data[id].output)
+        end
+        collectgarbage('collect')
+    end
+    return math.pow(10, - total_err / total_frame)
+end
+
+function nn:epoch()
+    local train_error = self:process(self.train_data, true)
+    local val_error = self:process(self.val_data, false)
+    return train_error, val_error
+end
diff --git a/lua/reader.lua b/lua/reader.lua
new file mode 100644
index 0000000..2e51a9c
--- /dev/null
+++ b/lua/reader.lua
@@ -0,0 +1,112 @@
+local Reader = nerv.class('nerv.Reader')
+
+function Reader:__init(vocab_file, input_file)
+    self:get_vocab(vocab_file)
+    self:get_seq(input_file)
+end
+
+function Reader:get_vocab(vocab_file)
+    local f = io.open(vocab_file, 'r')
+    local id = 0
+    self.vocab = {}
+    while true do
+        local word = f:read()
+        if word == nil then
+            break
+        end
+        self.vocab[word] = id
+        id = id + 1
+    end
+    self.size = id
+end
+
+function Reader:split(s, t)
+    local ret = {}
+    for x in (s .. t):gmatch('(.-)' .. t) do
+        table.insert(ret, x)
+    end
+    return ret
+end
+
+function Reader:get_seq(input_file)
+    local f = io.open(input_file, 'r')
+    self.seq = {}
+    while true do
+        local seq = f:read()
+        if seq == nil then
+            break
+        end
+        seq = self:split(seq, ' ')
+        local tmp = {}
+        for i = 1, #seq do
+            if seq[i] ~= '' then
+                table.insert(tmp, self.vocab[seq[i]])
+            end
+        end
+        table.insert(self.seq, tmp)
+    end
+end
+
+function Reader:get_in_out(id, pos)
+    return self.seq[id][pos], self.seq[id][pos + 1], pos + 1 == #self.seq[id]
+end
+
+function Reader:get_all_batch(global_conf)
+    local data = {}
+    local pos = {}
+    local offset = 1
+    for i = 1, global_conf.batch_size do
+        pos[i] = nil
+    end
+    while true do
+        local input = {}
+        local output = {}
+        for i = 1, global_conf.chunk_size do
+            input[i] = global_conf.mmat_type(global_conf.batch_size, 1)
+            input[i]:fill(global_conf.nn_act_default)
+            output[i] = global_conf.mmat_type(global_conf.batch_size, 1)
+            output[i]:fill(global_conf.nn_act_default)
+        end
+        local seq_start = {}
+        local seq_end = {}
+        local seq_len = {}
+        for i = 1, global_conf.batch_size do
+            seq_start[i] = false
+            seq_end[i] = false
+            seq_len[i] = 0
+        end
+        local has_new = false
+        for i = 1, global_conf.batch_size do
+            if pos[i] == nil then
+                if offset < #self.seq then
+                    seq_start[i] = true
+                    pos[i] = {offset, 1}
+                    offset = offset + 1
+                end
+            end
+            if pos[i] ~= nil then
+                has_new = true
+                for j = 1, global_conf.chunk_size do
+                    local final
+                    input[j][i-1][0], output[j][i-1][0], final = self:get_in_out(pos[i][1], pos[i][2])
+                    seq_len[i] = j
+                    if final then
+                        seq_end[i] = true
+                        pos[i] = nil
+                        break
+                    end
+                    pos[i][2] = pos[i][2] + 1
+                end
+            end
+        end
+        if not has_new then
+            break
+        end
+        for i = 1, global_conf.chunk_size do
+            input[i] = global_conf.cumat_type.new_from_host(input[i])
+            output[i] = global_conf.cumat_type.new_from_host(output[i])
+        end
+        table.insert(data, {input = input, output = output, seq_start = seq_start, seq_end = seq_end, seq_len = seq_len})
+    end
+    return data
+end
diff --git a/lua/select_linear.lua b/lua/select_linear.lua
new file mode 100644
index 0000000..a7e20cc
--- /dev/null
+++ b/lua/select_linear.lua
@@ -0,0 +1,62 @@
+local SL = nerv.class('nerv.SelectLinearLayer', 'nerv.Layer')
+
+--id: string
+--global_conf: table
+--layer_conf: table
+--Get Parameters
+function SL:__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
+
+    self.vocab = layer_conf.vocab
+    self.ltp = self:find_param("ltp", layer_conf, global_conf, nerv.LinearTransParam, {self.vocab, self.dim_out[1]}) --layer_conf.ltp
+ 
+    self:check_dim_len(1, 1)
+end
+
+--Check parameter 
+function SL:init(batch_size)
+    if (self.dim_in[1] ~= 1) then --one word id 
+        nerv.error("mismatching dimensions of ltp and input")
+    end
+    if (self.dim_out[1] ~= self.ltp.trans:ncol()) then
+        nerv.error("mismatching dimensions of bp and output")
+    end
+    
+    self.batch_size = bath_size
+    self.ltp:train_init()
+end
+
+function SL:update(bp_err, input, output)
+    --use this to produce reproducable result, don't forget to set the dropout to zero!
+    --for i = 1, input[1]:nrow(), 1 do
+    --    local word_vec = self.ltp.trans[input[1][i - 1][0]]
+    --    word_vec:add(word_vec, bp_err[1][i - 1], 1, - self.gconf.lrate / self.gconf.batch_size)
+    --end 
+    
+    --I tried the update_select_rows kernel which uses atomicAdd, but it generates unreproducable result
+    self.ltp.trans:update_select_rows_by_colidx(bp_err[1], input[1], - self.gconf.lrate / self.gconf.batch_size, 0)
+    self.ltp.trans:add(self.ltp.trans, self.ltp.trans, 1.0, - self.gconf.lrate * self.gconf.wcost)
+end
+
+function SL:propagate(input, output)
+    --for i = 0, input[1]:ncol() - 1, 1 do
+    --    if (input[1][0][i] > 0) then
+    --        output[1][i]:copy_fromd(self.ltp.trans[input[1][0][i]])
+    --    else
+    --        output[1][i]:fill(0)
+    --    end
+    --end
+    output[1]:copy_rows_fromd_by_colidx(self.ltp.trans, input[1])
+end
+
+function SL:back_propagate(bp_err, next_bp_err, input, output)
+    --input is compressed, do nothing
+end
+
+function SL:get_params()
+    local paramRepo = nerv.ParamRepo({self.ltp})
+    return paramRepo
+end
diff --git a/lua/timer.lua b/lua/timer.lua
new file mode 100644
index 0000000..2c54ca8
--- /dev/null
+++ b/lua/timer.lua
@@ -0,0 +1,33 @@
+local Timer = nerv.class("nerv.Timer")
+
+function Timer:__init()
+    self.last = {}
+    self.rec = {}
+end
+
+function Timer:tic(item)
+    self.last[item] = os.clock()
+end
+
+function Timer:toc(item)
+    if (self.last[item] == nil) then
+        nerv.error("item not there")
+    end
+    if (self.rec[item] == nil) then
+        self.rec[item] = 0
+    end
+    self.rec[item] = self.rec[item] + os.clock() - self.last[item]
+end
+
+function Timer:check(item)
+    if self.rec[item]==nil then
+        nerv.error('item not there')
+    end
+    nerv.printf('"%s" lasts for %f secs.\n',item,self.rec[item])
+end
+
+function Timer:flush()
+    for key, value in pairs(self.rec) do
+        self.rec[key] = nil
+    end
+end
diff --git a/lua/tnn.lua b/lua/tnn.lua
new file mode 100644
index 0000000..bf9f118
--- /dev/null
+++ b/lua/tnn.lua
@@ -0,0 +1,136 @@
+nerv.include('select_linear.lua')
+
+local reader = nerv.class('nerv.TNNReader')
+
+function reader:__init(global_conf, data)
+    self.gconf = global_conf
+    self.offset = 0
+    self.data = data
+end
+
+function reader:get_batch(feeds)
+    self.offset = self.offset + 1
+    if self.offset > #self.data then
+        return false
+    end
+    for i = 1, self.gconf.chunk_size do
+        feeds.inputs_m[i][1]:copy_from(self.data[self.offset].input[i])
+        feeds.inputs_m[i][2]:copy_from(self.data[self.offset].output[i]:decompress(self.gconf.vocab_size))
+    end
+    feeds.flags_now = self.data[self.offset].flags
+    feeds.flagsPack_now = self.data[self.offset].flagsPack
+    return true
+end
+
+function reader:has_data(t, i)
+    return t <= self.data[self.offset].seq_len[i]
+end
+
+function reader:get_err_input()
+    return self.data[self.offset].err_input
+end
+
+local nn = nerv.class('nerv.NN')
+
+function nn:__init(global_conf, train_data, val_data, layers, connections)
+    self.gconf = global_conf
+    self.tnn = self:get_tnn(layers, connections)
+    self.train_data = self:get_data(train_data)
+    self.val_data = self:get_data(val_data)
+end
+
+function nn:get_tnn(layers, connections)
+    self.gconf.dropout_rate = 0
+    local layer_repo = nerv.LayerRepo(layers, self.gconf.pr, self.gconf)
+    local tnn = nerv.TNN('TNN', self.gconf, {dim_in = {1, self.gconf.vocab_size}, 
+        dim_out = {1}, sub_layers = layer_repo, connections = connections, 
+        clip = self.gconf.clip})
+    tnn:init(self.gconf.batch_size, self.gconf.chunk_size)
+    return tnn
+end
+
+function nn:get_data(data)
+    local ret = {}
+    for i = 1, #data do
+        ret[i] = {}
+        ret[i].input = data[i].input
+        ret[i].output = data[i].output
+        ret[i].flags = {}
+        ret[i].err_input = {}
+        for t = 1, self.gconf.chunk_size do
+            ret[i].flags[t] = {}
+            local err_input = self.gconf.mmat_type(self.gconf.batch_size, 1)
+            for j = 1, self.gconf.batch_size do
+                if t <= data[i].seq_len[j] then
+                    ret[i].flags[t][j] = nerv.TNN.FC.SEQ_NORM
+                    err_input[j - 1][0] = 1
+                else
+                    ret[i].flags[t][j] = 0
+                    err_input[j - 1][0] = 0
+                end
+            end
+            ret[i].err_input[t] = self.gconf.cumat_type.new_from_host(err_input)
+        end
+        for j = 1, self.gconf.batch_size do
+            if data[i].seq_start[j] then
+                ret[i].flags[1][j] = bit.bor(ret[i].flags[1][j], nerv.TNN.FC.SEQ_START)
+            end
+            if data[i].seq_end[j] then
+                local t = data[i].seq_len[j]
+                ret[i].flags[t][j] = bit.bor(ret[i].flags[t][j], nerv.TNN.FC.SEQ_END)
+            end
+        end
+        ret[i].flagsPack = {}
+        for t = 1, self.gconf.chunk_size do
+            ret[i].flagsPack[t] = 0
+            for j = 1, self.gconf.batch_size do
+                ret[i].flagsPack[t] = bit.bor(ret[i].flagsPack[t], ret[i].flags[t][j])
+            end
+        end
+        ret[i].seq_len = data[i].seq_len
+    end
+    return ret
+end
+
+function nn:process(data, do_train)
+    local total_err = 0
+    local total_frame = 0
+    local reader = nerv.TNNReader(self.gconf, data)
+    while true do
+        local r, _ = self.tnn:getfeed_from_reader(reader)
+        if not r then
+            break
+        end
+        if do_train then
+            self.gconf.dropout_rate = self.gconf.dropout
+        else
+            self.gconf.dropout_rate = 0
+        end
+        self.tnn:net_propagate()
+        for t = 1, self.gconf.chunk_size do
+            local tmp = self.tnn.outputs_m[t][1]:new_to_host()
+            for i = 1, self.gconf.batch_size do
+                if reader:has_data(t, i) then
+                    total_err = total_err + math.log10(math.exp(tmp[i - 1][0]))
+                    total_frame = total_frame + 1
+                end
+            end
+        end
+        if do_train then
+            local err_input = reader:get_err_input()
+            for i = 1, self.gconf.chunk_size do
+                self.tnn.err_inputs_m[i][1]:copy_from(err_input[i])
+            end
+            self.tnn:net_backpropagate(false)
+            self.tnn:net_backpropagate(true)
+        end
+        collectgarbage('collect')
+    end
+    return math.pow(10, - total_err / total_frame)
+end
+
+function nn:epoch()
+    local train_error = self:process(self.train_data, true)
+    local val_error = self:process(self.val_data, false)
+    return train_error, val_error
+end