Both sides previous revision
Previous revision
Next revision
|
Previous revision
Next revision
Both sides next revision
|
user:dusek:vystadial:flect [2012/12/19 10:04] dusek |
user:dusek:vystadial:flect [2013/01/18 12:55] dusek |
2012-11-27 09:53:07,196 Loaded. Vectorizing... | 2012-11-27 09:53:07,196 Loaded. Vectorizing... |
2012-11-27 09:55:51,228 Data shape (652544, 301784) | 2012-11-27 09:55:51,228 Data shape (652544, 301784) |
/home/odusek/.local-x86_64/lib/python2.7/site-packages/sklearn/feature_selection/univariate_selection.py:94: RuntimeWarning: divide by zero encountered in divide | /home/odusek/.local-x86_64/lib/python2.7/site-packages/sklearn/feature_selection/univariate_selection.py:94: RuntimeWarning: divide by zero encountered in divide f = msb / msw |
f = msb / msw | |
2012-11-27 09:56:28,985 Filt shape (652544, 30183) | 2012-11-27 09:56:28,985 Filt shape (652544, 30183) |
2012-11-27 09:56:28,986 Training ... | 2012-11-27 09:56:28,986 Training ... |
| |
| |
*** Přeprogramování trénování *** | ** Přeprogramování trénování ** |
| |
2012-12-18 11:14:06,606 TREEX-INFO: Loading data set from data/train.arff.gz... | 2012-12-18 11:14:06,606 TREEX-INFO: Loading data set from data/train.arff.gz... |
2012-12-18 11:16:20,214 TREEX-INFO: Preparing data set... | 2012-12-18 11:16:20,214 TREEX-INFO: Preparing data set... |
2012-12-18 11:18:04,516 TREEX-INFO: Filtering... | 2012-12-18 11:18:04,516 TREEX-INFO: Filtering... |
/home/odusek/.local-x86_64/lib/python2.7/site-packages/sklearn/feature_selection/univariate_selection.py:94: RuntimeWarning: divide by zero encountered in divide | /home/odusek/.local-x86_64/lib/python2.7/site-packages/sklearn/feature_selection/univariate_selection.py:94: RuntimeWarning: divide by zero encountered in divide f = msb / msw |
f = msb / msw | |
2012-12-18 11:18:24,378 TREEX-INFO: Training... | 2012-12-18 11:18:24,378 TREEX-INFO: Training... |
2012-12-18 17:26:14,641 TREEX-INFO: Training done. | 2012-12-18 17:26:14,641 TREEX-INFO: Training done. |
2012-12-18 17:28:30,627 TREEX-INFO: Model successfully saved. | 2012-12-18 17:28:30,627 TREEX-INFO: Model successfully saved. |
| |
| |
| ** Trénování na datech vyprodukovaných v pythonu ** |
| |
| andromeda2:~/od-playground/test/exp-flect$ ../../src/experiment/train_model.py runs/train-plain_pydata/config.py data/train.arff.gz runs/train-plain_pydata/model.pickle.gz data/dtest.arff.gz runs/train-plain_pydata/classif.arff.gz |
| 2013-01-04 13:28:17,539 TREEX-INFO: Loading data set from data/train.arff.gz... |
| 2013-01-04 13:29:56,363 TREEX-INFO: Preparing data set... |
| 2013-01-04 13:31:04,559 TREEX-INFO: Filtering... |
| /home/odusek/.local-x86_64/lib/python2.7/site-packages/sklearn/feature_selection/univariate_selection.py:94: RuntimeWarning: divide by zero encountered in divide f = msb / msw |
| 2013-01-04 13:31:49,558 TREEX-INFO: Training... |
| 2013-01-05 00:55:21,004 TREEX-INFO: Training done. |
| 2013-01-05 00:55:21,374 TREEX-INFO: Evaluation on file: data/dtest.arff.gz |
| 2013-01-05 00:56:10,144 TREEX-INFO: Score: 0.961291064956 |
| 2013-01-05 00:56:10,149 TREEX-INFO: Saving model to file runs/train-plain_pydata/model.pickle.gz |
| 2013-01-05 00:57:22,448 TREEX-INFO: Model successfully saved. |
| |
| ** Výsledky pro různé parametry logreg ** |
| |
| * Nepomáhá příliš malé C = 0.1, ani příliš malé tol = 0.1 |
| * Spíš ani C = 1 není nic moc, C = 10 nebo 100 je mnohem lepší |
| * Tol taky radši = 0.001 nebo 0.0001 |
| * Na L2 / L1 druhu regularizace zřejmě moc nezávisí |
| * Rozpětí 96.92 - 94.01, naprostá většina nad 96.5 |
| * L2 regularizace tvoří nechutně velké modely, L1 jsou mnooohem menší |
| |
| train-l2_1000_001.py.o6633181:2013-01-11 03:15:37,871 TREEX-INFO: Score: 0.968472611875 |
| train-l1_100_0001.py.o6633154:2013-01-10 18:35:26,878 TREEX-INFO: Score: 0.968484109828 |
| train-l1_100_00001.py.o6633155:2013-01-11 04:25:29,295 TREEX-INFO: Score: 0.968886538196 |
| train-l1_10_0001.py.o6633150:2013-01-11 08:14:22,650 TREEX-INFO: Score: 0.96910499931 |
| train-l1_10_00001.py.o6633151:2013-01-11 03:05:33,517 TREEX-INFO: Score: 0.969254472704 |
| |
| ** Použití SVM ** |
| * Lineární SVM -- trvá dýl trénování, jinak není rozdíl -- nedosahují ani nejlepších výsledků. |
| * hlavně s L2 je dlouhé. |
| |
| train-l1_l2_1_False_0001.py.o6636505:2013-01-12 21:33:26,919 TREEX-INFO: Score: 0.964501467119 |
| train-l2_l2_1_False_0001.py.o6636541:2013-01-14 16:01:52,936 TREEX-INFO: Score: 0.964363385306 |
| train-l2_l2_10_False_0001.py.o6636544:2013-01-14 06:30:18,333 TREEX-INFO: Score: 0.964363385306 |
| train-l1_l2_1_False_00001.py.o6636506:2013-01-13 09:29:26,141 TREEX-INFO: Score: 0.964363385306 |
| |
| * Normální SVC s 16G paměti spadne |
| * S 32G to doběhne, ale s mizivým výsledkem |
| |
| ** Zkrácení sufixů a filtrace ** |
| |
| * Bez použití teček se prodlouží trénování, zatím L2 vyhrává |
| * Pokud se sufixy zkrátí na 4 znaky, funguje to dobře -- ale jen bez filtrování; s ním je to už moc slabé |
| * Bez filtrování to funguje dobře |