2013-11-15 - 版本历史

2013年11月18日 (一) 06:48 Cslt

2013-11-18T06:48:50Z

2013年11月18日 (一) 06:46 Cslt

2013-11-18T06:46:42Z

2013年11月18日 (一) 06:45 Cslt

2013-11-18T06:45:42Z

Cslt：/* QA LM */

2013-11-18T06:35:19Z

‎QA LM

Cslt：/* QA LM */

2013-11-18T06:34:58Z

‎QA LM

2013年11月18日 (一) 06:33 Cslt

2013-11-18T06:33:50Z

Cslt：以内容“== Data sharing == * LM count files still undelivered! == AM development == === Sparse DNN === * Optimal Brain Damage(OBD). * Online OBD. * Try 3 configurations...”创建新页面

2013-11-15T02:32:47Z

以内容“== Data sharing == * LM count files still undelivered! == AM development == === Sparse DNN === * Optimal Brain Damage(OBD). * Online OBD. * Try 3 configurations...”创建新页面

新页面

== Data sharing ==

* LM count files still undelivered!

== AM development ==

=== Sparse DNN ===

* Optimal Brain Damage(OBD).

* Online OBD.

* Try 3 configurations: batch size=256, 13000 (10 prunings), whole data. The current results show that the the performance order is: whole data > 256 > 13000.

===Noisy training ===

* Simulated Annealing training.
* Rejected with small noises. With clean training rejected after annealing.

* Noise concentrated training

=== Tencent exps ===
N/A

==LM development==

===NN LM ===

===QA LM===

# Tencent word segmentation system ready.
# Collecting data for Q-LM training.

←上一版本		2013年11月18日 (一) 06:48的版本
第17行：		第17行：

	===Noisy training ===		===Noisy training ===
		+
		+	* An ICASSP paper submitted.

	* Simulated Annealing training.		* Simulated Annealing training.

@@ 第35行： / 第35行： @@
 ==LM development==
-==NN LM===
+===NN LM===
 #  Results show better performance with NN rescoring.
@@ 第53行： / 第53行： @@
 </pre>
-==QA LM ==
+===QA LM ===
 The QA model training done.  Test on the Sogou Q text.

@@ 第62行： / 第62行： @@
 |Q (10G)||15w  ||1.5G ||800M|| 301.64  || 317.19
 |-
-|QA(100G):11w ||4.5G ||1G  || 287.134 || 315.695
+|QA(100G)||11w ||4.5G ||1G  || 287.134 || 315.695
 |-
-|QA(100G):8w8 ||4.5G ||1G  || 559.029 || 626.146
+|QA(100G)||8w8 ||4.5G ||1G  || 559.029 || 626.146
 |-
 |}

@@ 第59行： / 第59行： @@
 {| class="wikitable"
 ! Data !! lexicon !! size !! size2 !! PPL !! PPL2
 |Q (10G)||15w  ||1.5G ||800M|| 301.64  || 317.19
 |-

@@ 第58行： / 第58行： @@
 {| class="wikitable"
-!! Data ! lexicon ! size ! size2 ! PPL ! PPL2
+! Data !! lexicon !! size !! size2 !! PPL !! PPL2
-||Q (10G)|15w  |1.5G |800M| 301.64  | 317.19
+|Q (10G)||15w  ||1.5G ||800M|| 301.64  || 317.19
 |-
-||QA(100G):11w |4.5G |1G  | 287.134 | 315.695
+|QA(100G):11w ||4.5G ||1G  || 287.134 || 315.695
 |-
-||QA(100G):8w8 |4.5G |1G  | 559.029 | 626.146
+|QA(100G):8w8 ||4.5G ||1G  || 559.029 || 626.146
 |-
 |}

@@ 第9行： / 第9行： @@
 * Optimal Brain Damage(OBD).
-* Online OBD.
+# Basic OBD done, with the ICASSP paper submitted.
+# Online OBD running
-* Try 3 configurations: batch size=256, 13000 (10 prunings), whole data. The current results show that the the performance order is: whole data > 256 > 13000.
+:* Try 3 configurations: batch size=256, 13000 (10 prunings), whole data.
+:* The current results show that the the performance follows the order: Acc(whole data) > Acc(256) > Acc(13000).
@@ 第18行： / 第19行： @@
 * Simulated Annealing training.
-* Rejected with small noises. With clean training rejected after annealing.
+:* Rejected with small noises.
 * Noise concentrated training
+* Noise-adding modification
 === Tencent exps ===
@@ 第30行： / 第35行： @@
 ==LM development==
-===NN LM ===
+==NN LM===
-===QA LM===
+The QA model training done.  Test on the Sogou Q text.
-#  Tencent word segmentation system ready.
+{| class="wikitable"
-#  Collecting data for Q-LM training.
+!! Data ! lexicon ! size ! size2 ! PPL ! PPL2