http://cslt.org/mediawiki/index.php?action=history&feed=atom&title=2013-06-28 2026-04-15T03:43:37Z 本wiki的该页面的版本历史 MediaWiki 1.23.3 http://cslt.org/mediawiki/index.php?title=2013-06-28&diff=7974&oldid=prev 2013-06-28T09:56:12Z

<p></p> <table class='diff diff-contentalign-left'> <col class='diff-marker' /> <col class='diff-content' /> <col class='diff-marker' /> <col class='diff-content' /> <tr style='vertical-align: top;'> <td colspan='2' style="background-color: white; color:black; text-align: center;">←上一版本</td> <td colspan='2' style="background-color: white; color:black; text-align: center;">2013年6月28日 (五) 09:56的版本</td> </tr><tr><td colspan="2" class="diff-lineno">第12行：</td> <td colspan="2" class="diff-lineno">第12行：</td></tr> <tr><td class='diff-marker'>&#160;</td><td style="background-color: #f9f9f9; color: #333333; font-size: 88%; border-style: solid; border-width: 1px 1px 1px 4px; border-radius: 0.33em; border-color: #e6e6e6; vertical-align: top; white-space: pre-wrap;"><div>=== Tencent exps ===</div></td><td class='diff-marker'>&#160;</td><td style="background-color: #f9f9f9; color: #333333; font-size: 88%; border-style: solid; border-width: 1px 1px 1px 4px; border-radius: 0.33em; border-color: #e6e6e6; vertical-align: top; white-space: pre-wrap;"><div>=== Tencent exps ===</div></td></tr> <tr><td class='diff-marker'>&#160;</td><td style="background-color: #f9f9f9; color: #333333; font-size: 88%; border-style: solid; border-width: 1px 1px 1px 4px; border-radius: 0.33em; border-color: #e6e6e6; vertical-align: top; white-space: pre-wrap;"></td><td class='diff-marker'>&#160;</td><td style="background-color: #f9f9f9; color: #333333; font-size: 88%; border-style: solid; border-width: 1px 1px 1px 4px; border-radius: 0.33em; border-color: #e6e6e6; vertical-align: top; white-space: pre-wrap;"></td></tr> <tr><td colspan="2">&#160;</td><td class='diff-marker'>+</td><td style="color:black; font-size: 88%; border-style: solid; border-width: 1px 1px 1px 4px; border-radius: 0.33em; border-color: #a3d3ff; vertical-align: top; white-space: pre-wrap;"><div><ins style="font-weight: bold; text-decoration: none;">1：本周大规模数据DNN实验结束，基本在第8次迭代收敛， </ins></div></td></tr> <tr><td colspan="2">&#160;</td><td class='diff-marker'>+</td><td style="color:black; font-size: 88%; border-style: solid; border-width: 1px 1px 1px 4px; border-radius: 0.33em; border-color: #a3d3ff; vertical-align: top; white-space: pre-wrap;"><div><ins style="font-weight: bold; text-decoration: none;"></ins></div></td></tr> <tr><td colspan="2">&#160;</td><td class='diff-marker'>+</td><td style="color:black; font-size: 88%; border-style: solid; border-width: 1px 1px 1px 4px; border-radius: 0.33em; border-color: #a3d3ff; vertical-align: top; white-space: pre-wrap;"><div><ins style="font-weight: bold; text-decoration: none;"> Old Baseline New Baeline DNN-1000小时 DNN-6000小时（7次迭代）DNN-6000小时（8次迭代）</ins></div></td></tr> <tr><td colspan="2">&#160;</td><td class='diff-marker'>+</td><td style="color:black; font-size: 88%; border-style: solid; border-width: 1px 1px 1px 4px; border-radius: 0.33em; border-color: #a3d3ff; vertical-align: top; white-space: pre-wrap;"><div><ins style="font-weight: bold; text-decoration: none;"> 1900 8.4 6.8 4.3 3.9 3.7</ins></div></td></tr> <tr><td colspan="2">&#160;</td><td class='diff-marker'>+</td><td style="color:black; font-size: 88%; border-style: solid; border-width: 1px 1px 1px 4px; border-radius: 0.33em; border-color: #a3d3ff; vertical-align: top; white-space: pre-wrap;"><div><ins style="font-weight: bold; text-decoration: none;"> 2044 22.4 15.7 12.7 10.7 10.6</ins></div></td></tr> <tr><td colspan="2">&#160;</td><td class='diff-marker'>+</td><td style="color:black; font-size: 88%; border-style: solid; border-width: 1px 1px 1px 4px; border-radius: 0.33em; border-color: #a3d3ff; vertical-align: top; white-space: pre-wrap;"><div><ins style="font-weight: bold; text-decoration: none;"> online1 35.6 32.7 25.8 24.6 24.6</ins></div></td></tr> <tr><td colspan="2">&#160;</td><td class='diff-marker'>+</td><td style="color:black; font-size: 88%; border-style: solid; border-width: 1px 1px 1px 4px; border-radius: 0.33em; border-color: #a3d3ff; vertical-align: top; white-space: pre-wrap;"><div><ins style="font-weight: bold; text-decoration: none;"> online2 29.6 27.3 22.1 21.1 20.8</ins></div></td></tr> <tr><td colspan="2">&#160;</td><td class='diff-marker'>+</td><td style="color:black; font-size: 88%; border-style: solid; border-width: 1px 1px 1px 4px; border-radius: 0.33em; border-color: #a3d3ff; vertical-align: top; white-space: pre-wrap;"><div><ins style="font-weight: bold; text-decoration: none;"> map 24.5 15.8 13.4 8.7 8.4</ins></div></td></tr> <tr><td colspan="2">&#160;</td><td class='diff-marker'>+</td><td style="color:black; font-size: 88%; border-style: solid; border-width: 1px 1px 1px 4px; border-radius: 0.33em; border-color: #a3d3ff; vertical-align: top; white-space: pre-wrap;"><div><ins style="font-weight: bold; text-decoration: none;"> general 36 25.1 19.3 15.9 16</ins></div></td></tr> <tr><td colspan="2">&#160;</td><td class='diff-marker'>+</td><td style="color:black; font-size: 88%; border-style: solid; border-width: 1px 1px 1px 4px; border-radius: 0.33em; border-color: #a3d3ff; vertical-align: top; white-space: pre-wrap;"><div><ins style="font-weight: bold; text-decoration: none;"></ins></div></td></tr> <tr><td colspan="2">&#160;</td><td class='diff-marker'>+</td><td style="color:black; font-size: 88%; border-style: solid; border-width: 1px 1px 1px 4px; border-radius: 0.33em; border-color: #a3d3ff; vertical-align: top; white-space: pre-wrap;"><div><ins style="font-weight: bold; text-decoration: none;">2：dnn+dt代码完成，下周可以开始实验。首先将在小数据上进行。</ins></div></td></tr> <tr><td class='diff-marker'>&#160;</td><td style="background-color: #f9f9f9; color: #333333; font-size: 88%; border-style: solid; border-width: 1px 1px 1px 4px; border-radius: 0.33em; border-color: #e6e6e6; vertical-align: top; white-space: pre-wrap;"><div>=== GPU &amp; CPU merge ===</div></td><td class='diff-marker'>&#160;</td><td style="background-color: #f9f9f9; color: #333333; font-size: 88%; border-style: solid; border-width: 1px 1px 1px 4px; border-radius: 0.33em; border-color: #e6e6e6; vertical-align: top; white-space: pre-wrap;"><div>=== GPU &amp; CPU merge ===</div></td></tr> <tr><td class='diff-marker'>&#160;</td><td style="background-color: #f9f9f9; color: #333333; font-size: 88%; border-style: solid; border-width: 1px 1px 1px 4px; border-radius: 0.33em; border-color: #e6e6e6; vertical-align: top; white-space: pre-wrap;"></td><td class='diff-marker'>&#160;</td><td style="background-color: #f9f9f9; color: #333333; font-size: 88%; border-style: solid; border-width: 1px 1px 1px 4px; border-radius: 0.33em; border-color: #e6e6e6; vertical-align: top; white-space: pre-wrap;"></td></tr> </table>

Wangd http://cslt.org/mediawiki/index.php?title=2013-06-28&diff=7962&oldid=prev 2013-06-28T06:47:57Z

<p>以内容“== Data sharing == * LM count files still undelivered! == DNN progress == === Experiments === * Sparse DNN. 1. With Atlas without any change, on the ARM platform, ...”创建新页面</p> <p><b>新页面</b></p><div>== Data sharing ==<br /> <br /> * LM count files still undelivered!<br /> <br /> == DNN progress ==<br /> <br /> === Experiments ===<br /> * Sparse DNN. <br /> <br /> 1. With Atlas without any change, on the ARM platform, obtained RT 2.0. Will change the Atlas code to support sparse matrices.<br /> <br /> === Tencent exps ===<br /> <br /> === GPU &amp; CPU merge ===<br /> <br /> #Hold<br /> <br /> == RNN LM progress ==<br /> <br /> * Use 100M text, 10k lexicon in training. Validation test set is obtained from the transcription of the Tencent online1 speech data.<br /> <br /> :100 hidden layer, 1 hidden layer, 3-gram<br /> :training time: 7 hour, 8GB<br /> :prediction time: quick, 8GB<br /> <br /> :3-gram PPL: 227.323 WER: 36%<br /> :RNN PPL: 170.056129 WER: 41%<br /> :3-gram+RNN: PPL: 0.25RNN+0.75 3-gram: 180.0 WER 35%<br /> <br /> *possibly a bug when computing PPL with the RNN toolkit.<br /> <br /> <br /> == Embedded progress ==<br /> <br /> *Status:<br /> <br /> : 1000 test words + 2000 noise words<br /> <br /> before | after <br /> utt 952 3317<br /> %wer 6.26% 11.04%<br /> RT 0.07 0.20<br /> <br /> This means the GMM-based system highly relies on the vocabulary. It may work well with small lexica, but difficult with large ones.<br /> <br /> :Run other optimization parameters:<br /> <br /> option | RT | %wer<br /> <br /> original 0.07 6.28<br /> <br /> -ds 0.06 6.33%<br /> -topn 0.06 6.80% <br /> -maxwpf - -<br /> -maxhmmpf - - <br /> -kdmaxdepth - -<br /> -kdmaxbbi - -<br /> -pl_window - -<br /> <br /> <br /> *To be done<br /> :# sparse DNN based Kaldi engine<br /> :# sparse DNN based PS engine</div>

Wangd