From 03439902dbd339cfbbc684b6fcc6b1810fa02ede Mon Sep 17 00:00:00 2001
From: Qi Liu <liuq901@163.com>
Date: Mon, 9 May 2016 20:51:10 +0800
Subject: fix bug in affine.lua

---
 nerv/nn/network.lua | 1 -
 1 file changed, 1 deletion(-)

(limited to 'nerv/nn')

diff --git a/nerv/nn/network.lua b/nerv/nn/network.lua
index bf69ccc..d0d5462 100644
--- a/nerv/nn/network.lua
+++ b/nerv/nn/network.lua
@@ -416,7 +416,6 @@ function network:make_initial_store()
             local dim_in, dim_out = self.layers[i]:get_dim()
             for j = 1, #dim_in do
                 if self.input[t][i][j] == nil then
-                    print(t,i,j,self.layers[i].id)
                     nerv.error('input reference dangling')
                 end
                 if self.err_output[t][i][j] == nil then
-- 
cgit v1.2.3


From 4585970021f75d4c9e7154fc1681a80efa0f48ab Mon Sep 17 00:00:00 2001
From: Qi Liu <liuq901@163.com>
Date: Mon, 9 May 2016 21:02:57 +0800
Subject: clip on gradient now

---
 nerv/nn/network.lua | 10 ----------
 nerv/nn/trainer.lua |  2 +-
 2 files changed, 1 insertion(+), 11 deletions(-)

(limited to 'nerv/nn')

diff --git a/nerv/nn/network.lua b/nerv/nn/network.lua
index d0d5462..358b100 100644
--- a/nerv/nn/network.lua
+++ b/nerv/nn/network.lua
@@ -33,8 +33,6 @@ local network = nerv.class('nerv.Network')
 --
 -- * `network`: a `nerv.Layer` instance describing the structure of the network
 --   to be compiled
--- * `clip`: a `number` value indicating the cliping threshold (i.e. preserve
---   the values within [-clip, +clip])
 -- * `nn_act_default`: a `number` value indicating the value used for filling
 --   "holes" in activation values of a batch matrix (0 by default)
 
@@ -49,7 +47,6 @@ function network:__init(id, global_conf, network_conf)
     else
         self.mat_type = self.gconf.cumat_type
     end
-    self.clip = network_conf.clip
     self.nn_act_default = network_conf.nn_act_default
     if self.nn_act_default == nil then
         self.nn_act_default = 0
@@ -660,13 +657,6 @@ function network:back_propagate()
         local t, id = self.queue[i].chunk, self.queue[i].id
         if t <= self.max_length then
             self.layers[id]:back_propagate(self.err_input[t][id], self.err_output[t][id], self.input[t][id], self.output[t][id], t)
-            -- gradient clip
-            if self.clip ~= nil then
-                local dim_in, _ = self.layers[id]:get_dim()
-                for j = 1, #dim_in do
-                    self.err_output[t][id][j]:clip(-self.clip, self.clip)
-                end
-            end
         end
         -- flush border gradient
         if self.flush[t][id].timestamp == self.timestamp then
diff --git a/nerv/nn/trainer.lua b/nerv/nn/trainer.lua
index 44390ea..8357c10 100644
--- a/nerv/nn/trainer.lua
+++ b/nerv/nn/trainer.lua
@@ -25,7 +25,7 @@ function trainer:__init(gconf)
     self.input_order = self:get_input_order()
     self.network = nerv.Network('network', gconf,
                                 {network = graph,
-                                 clip = gconf.clip})
+                                 nn_act_default = gconf.nn_act_default})
     local network = self.network
     network:init(gconf.batch_size, gconf.chunk_size)
 
-- 
cgit v1.2.3


From e85a4af65640aef77378275d478c1ba8b06b785e Mon Sep 17 00:00:00 2001
From: Qi Liu <liuq901@163.com>
Date: Wed, 11 May 2016 16:53:56 +0800
Subject: seq buffer support shuffle

---
 nerv/nn/trainer.lua | 3 +++
 1 file changed, 3 insertions(+)

(limited to 'nerv/nn')

diff --git a/nerv/nn/trainer.lua b/nerv/nn/trainer.lua
index 8357c10..a17b36c 100644
--- a/nerv/nn/trainer.lua
+++ b/nerv/nn/trainer.lua
@@ -77,9 +77,12 @@ function trainer:make_buffer(readers)
         })
     else
         return nerv.SeqBuffer(gconf, {
+            buffer_size = gconf.buffer_size,
             batch_size = gconf.batch_size,
             chunk_size = gconf.chunk_size,
+            randomize = gconf.randomize,
             readers = readers,
+            nn_act_default = gconf.nn_act_default,
         })
     end
 end
-- 
cgit v1.2.3