Merge branch 'master' of github.com:liuq901/nerv into liuq901-master

4 files changed, 87 insertions, 140 deletions

diff --git a/nerv/examples/network_debug/config.lua b/nerv/examples/network_debug/config.lua
index e20d5a9..0429e9a 100644
--- a/nerv/examples/network_debug/config.lua
+++ b/nerv/examples/network_debug/config.lua
@@ -35,6 +35,10 @@ function get_layers(global_conf)
         ['nerv.SoftmaxCELayer'] = {
             softmax = {dim_in = {global_conf.vocab_size, global_conf.vocab_size}, dim_out = {1}, compressed = true},
         },
+        ['nerv.DuplicateLayer'] = {
+            dup1 = {dim_in = {1}, dim_out = {1}},
+            dup2 = {dim_in = {1}, dim_out = {1}},
+        },
     }
     for i = 1, global_conf.layer_num do
         layers['nerv.LSTMLayer']['lstm' .. i] = {dim_in = {global_conf.hidden_size}, dim_out = {global_conf.hidden_size}, pr = pr}
@@ -45,12 +49,14 @@ end
 
 function get_connections(global_conf)
     local connections = {
-        {'<input>[1]', 'select[1]', 0},
+        {'<input>[1]', 'dup1[1]', 0},
+        {'dup1[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},
+        {'softmax[1]', 'dup2[1]', 0},
+        {'dup2[1]', '<output>[1]', 0},
     }
     for i = 1, global_conf.layer_num do
         table.insert(connections, {'lstm' .. i .. '[1]', 'dropout' .. i .. '[1]', 0})
diff --git a/nerv/examples/network_debug/main.lua b/nerv/examples/network_debug/main.lua
index 790c404..bbcdb6c 100644
--- a/nerv/examples/network_debug/main.lua
+++ b/nerv/examples/network_debug/main.lua
@@ -6,35 +6,26 @@ nerv.include(arg[1])
 local global_conf = get_global_conf()
 local timer = global_conf.timer
 
-timer:tic('IO')
-
 local data_path = '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()
+local NN = nerv.NN(global_conf, layers, connections)
 
 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()
+    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_ppl, val_ppl = NN:epoch(train_reader, val_reader)
+    nerv.printf('Epoch %d: %f %f %f\n', i, global_conf.lrate, train_ppl, val_ppl)
     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()
@@ -43,3 +34,5 @@ timer:toc('global')
 timer:check('global')
 timer:check('network')
 timer:check('gc')
+timer:check('IO')
+global_conf.cumat_type.print_profile()
diff --git a/nerv/examples/network_debug/network.lua b/nerv/examples/network_debug/network.lua
index 5518e27..386c3b0 100644
--- a/nerv/examples/network_debug/network.lua
+++ b/nerv/examples/network_debug/network.lua
@@ -2,11 +2,17 @@ nerv.include('select_linear.lua')
 
 local nn = nerv.class('nerv.NN')
 
-function nn:__init(global_conf, train_data, val_data, layers, connections)
+function nn:__init(global_conf, 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)
+
+    self.output = {}
+    self.err_output = {}
+    for i = 1, self.gconf.chunk_size do
+        self.output[i] = {self.gconf.cumat_type(self.gconf.batch_size, 1)}
+        self.err_output[i] = {self.gconf.cumat_type(self.gconf.batch_size, 1)}
+        self.err_output[i][2] = self.gconf.cumat_type(self.gconf.batch_size, 1)
+    end
 end
 
 function nn:get_network(layers, connections)
@@ -20,79 +26,67 @@ function nn:get_network(layers, connections)
     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
+function nn:process(data, do_train, reader)
+    local timer = self.gconf.timer
+    local buffer = nerv.SeqBuffer(self.gconf, {
+        batch_size = self.gconf.batch_size, chunk_size = self.gconf.chunk_size,
+        readers = {reader},
+    })
+    local total_err = 0
+    local total_frame = 0
+    self.network:epoch_init()
+    while true do
+        timer:tic('IO')
+        data = buffer:get_data()
+        if data == nil then
+            break
+        end
+        local err_input = {}
+        if do_train then
+            for t = 1, self.gconf.chunk_size do
+                local tmp = self.gconf.mmat_type(self.gconf.batch_size, 1)
+                for i = 1, self.gconf.batch_size do
+                    if t <= data.seq_length[i] then
+                        tmp[i - 1][0] = 1
+                    else
+                        tmp[i - 1][0] = 0
+                    end
                 end
+                err_input[t] = {self.gconf.cumat_type.new_from_host(tmp)}
             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].seq_length = data[i].seq_len
-        ret[i].new_seq = {}
-        for j = 1, self.gconf.batch_size do
-            if data[i].seq_start[j] then
-                table.insert(ret[i].new_seq, j)
-            end
+        local info = {input = {}, output = self.output, err_input = err_input, do_train = do_train,
+            err_output = self.err_output, seq_length = data.seq_length, new_seq = data.new_seq}
+        for t = 1, self.gconf.chunk_size do
+            info.input[t] = {data.data['input'][t]}
+            info.input[t][2] = data.data['label'][t]
         end
-    end
-    return ret
-end
+        timer:toc('IO')
 
-function nn:process(data, do_train)
-    local timer = self.gconf.timer
-    local total_err = 0
-    local total_frame = 0
-    self.network:epoch_init()
-    for id = 1, #data do
-        data[id].do_train = do_train
         timer:tic('network')
-        self.network:mini_batch_init(data[id])
+        self.network:mini_batch_init(info)
         self.network:propagate()
         timer:toc('network')
+
+        timer:tic('IO')
         for t = 1, self.gconf.chunk_size do
-            local tmp = data[id].output[t][1]:new_to_host()
+            local tmp = info.output[t][1]:new_to_host()
             for i = 1, self.gconf.batch_size do
-                if t <= data[id].seq_length[i] then
-                    total_err = total_err + math.log10(math.exp(tmp[i - 1][0]))
-                    total_frame = total_frame + 1
-                end
+                total_err = total_err + math.log10(math.exp(tmp[i - 1][0]))
             end
         end
+        for i = 1, self.gconf.batch_size do
+            total_frame = total_frame + info.seq_length[i]
+        end
+        timer:toc('IO')
+
+        timer:tic('network')
         if do_train then
-            timer:tic('network')
             self.network:back_propagate()
             self.network:update()
-            timer:toc('network')
         end
+        timer:toc('network')
+
         timer:tic('gc')
         collectgarbage('collect')
         timer:toc('gc')
@@ -100,11 +94,11 @@ function nn:process(data, do_train)
     return math.pow(10, - total_err / total_frame)
 end
 
-function nn:epoch()
-    local train_error = self:process(self.train_data, true)
+function nn:epoch(train_reader, val_reader)
+    local train_error = self:process(self.train_data, true, train_reader)
     local tmp = self.gconf.dropout_rate
     self.gconf.dropout_rate = 0
-    local val_error = self:process(self.val_data, false)
+    local val_error = self:process(self.val_data, false, val_reader)
     self.gconf.dropout_rate = tmp
     return train_error, val_error
 end
diff --git a/nerv/examples/network_debug/reader.lua b/nerv/examples/network_debug/reader.lua
index b10baaf..76a78cf 100644
--- a/nerv/examples/network_debug/reader.lua
+++ b/nerv/examples/network_debug/reader.lua
@@ -3,6 +3,7 @@ local Reader = nerv.class('nerv.Reader')
 function Reader:__init(vocab_file, input_file)
     self:get_vocab(vocab_file)
     self:get_seq(input_file)
+    self.offset = 1
 end
 
 function Reader:get_vocab(vocab_file)
@@ -32,6 +33,7 @@ function Reader:get_seq(input_file)
     local f = io.open(input_file, 'r')
     self.seq = {}
     while true do
+    -- for i = 1, 26 do
         local seq = f:read()
         if seq == nil then
             break
@@ -47,67 +49,19 @@ function Reader:get_seq(input_file)
     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
+function Reader:get_data()
+    if self.offset > #self.seq then
+        return nil
     end
-    while true do
-    -- for i = 1, 26 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})
+    local tmp = self.seq[self.offset]
+    local res = {
+        input = nerv.MMatrixFloat(#tmp - 1, 1),
+        label = nerv.MMatrixFloat(#tmp - 1, 1),
+    }
+    for i = 1, #tmp - 1 do
+        res.input[i - 1][0] = tmp[i]
+        res.label[i - 1][0] = tmp[i + 1]
     end
-    return data
+    self.offset = self.offset + 1
+    return res
 end