зеркало из https://github.com/mozilla/kaldi.git
Add faster-rnnlm rescoring to librispeech
Add installation script Update rescoring results
This commit is contained in:
Anton Bakhtin
Родитель
250c69c4d8
Коммит
338ef754d5
|
|
@ -164,37 +164,98 @@
|
|||
%WER 17.16 [ 8982 / 52343, 855 ins, 1421 del, 6706 sub ] exp/nnet2_online/nnet_a_online/decode_test_other_tgmed/wer_12
|
||||
%WER 18.90 [ 9891 / 52343, 798 ins, 1786 del, 7307 sub ] exp/nnet2_online/nnet_a_online/decode_test_other_tgsmall/wer_13
|
||||
|
||||
# RNNLM rescoring of tri6b
|
||||
# RNNLM rescoring of tri6b (rnnlm-hs-0.1b hidden=150 direct=4.0Gb)
|
||||
%WER 7.39 [ 4023 / 54402, 540 ins, 444 del, 3039 sub ] exp/tri6b/decode_tglarge_dev_clean/wer_13_1.0
|
||||
%WER 7.06 [ 3839 / 54402, 547 ins, 401 del, 2891 sub ] exp/tri6b/decode_tglarge_dev_clean_rnnlm-hs-0.1b_h150-me5-1000_L0.25/wer_13_1.0
|
||||
%WER 7.14 [ 3883 / 54402, 579 ins, 364 del, 2940 sub ] exp/tri6b/decode_tglarge_dev_clean_rnnlm-hs-0.1b_h150-me5-1000_L0.5/wer_13_1.0
|
||||
%WER 7.54 [ 4103 / 54402, 651 ins, 343 del, 3109 sub ] exp/tri6b/decode_tglarge_dev_clean_rnnlm-hs-0.1b_h150-me5-1000_L0.75/wer_12_1.0
|
||||
|
||||
%WER 7.50 [ 4080 / 54402, 617 ins, 416 del, 3047 sub ] exp/tri6b/decode_tglarge_dev_clean/wer_14
|
||||
%WER 7.09 [ 3859 / 54402, 611 ins, 354 del, 2894 sub ] exp/tri6b/decode_tglarge_dev_clean_rnnlm_h150_me5-1000_L0.25/wer_14
|
||||
%WER 7.29 [ 3968 / 54402, 661 ins, 332 del, 2975 sub ] exp/tri6b/decode_tglarge_dev_clean_rnnlm_h150_me5-1000_L0.5/wer_13
|
||||
%WER 7.73 [ 4205 / 54402, 709 ins, 349 del, 3147 sub ] exp/tri6b/decode_tglarge_dev_clean_rnnlm_h150_me5-1000_L0.75/wer_12
|
||||
%WER 21.31 [ 10858 / 50948, 1525 ins, 1151 del, 8182 sub ] exp/tri6b/decode_tglarge_dev_other/wer_17_0.0
|
||||
%WER 20.77 [ 10583 / 50948, 1387 ins, 1222 del, 7974 sub ] exp/tri6b/decode_tglarge_dev_other_rnnlm-hs-0.1b_h150-me5-1000_L0.25/wer_17_0.5
|
||||
%WER 20.88 [ 10638 / 50948, 1456 ins, 1148 del, 8034 sub ] exp/tri6b/decode_tglarge_dev_other_rnnlm-hs-0.1b_h150-me5-1000_L0.5/wer_16_0.5
|
||||
%WER 21.37 [ 10890 / 50948, 1518 ins, 1126 del, 8246 sub ] exp/tri6b/decode_tglarge_dev_other_rnnlm-hs-0.1b_h150-me5-1000_L0.75/wer_17_0.5
|
||||
|
||||
%WER 21.94 [ 11180 / 50948, 1264 ins, 1506 del, 8410 sub ] exp/tri6b/decode_tglarge_dev_other/wer_16
|
||||
%WER 21.36 [ 10881 / 50948, 1309 ins, 1362 del, 8210 sub ] exp/tri6b/decode_tglarge_dev_other_rnnlm_h150_me5-1000_L0.25/wer_16
|
||||
%WER 21.29 [ 10848 / 50948, 1330 ins, 1324 del, 8194 sub ] exp/tri6b/decode_tglarge_dev_other_rnnlm_h150_me5-1000_L0.5/wer_16
|
||||
%WER 21.75 [ 11082 / 50948, 1351 ins, 1346 del, 8385 sub ] exp/tri6b/decode_tglarge_dev_other_rnnlm_h150_me5-1000_L0.75/wer_17
|
||||
%WER 9.21 [ 5012 / 54402, 703 ins, 510 del, 3799 sub ] exp/tri6b/decode_tgmed_dev_clean/wer_14_0.0
|
||||
%WER 7.97 [ 4334 / 54402, 579 ins, 465 del, 3290 sub ] exp/tri6b/decode_tgmed_dev_clean_rnnlm-hs-0.1b_h150-me5-1000_L0.25/wer_14_0.5
|
||||
%WER 7.77 [ 4226 / 54402, 615 ins, 394 del, 3217 sub ] exp/tri6b/decode_tgmed_dev_clean_rnnlm-hs-0.1b_h150-me5-1000_L0.5/wer_14_0.5
|
||||
%WER 7.97 [ 4338 / 54402, 614 ins, 420 del, 3304 sub ] exp/tri6b/decode_tgmed_dev_clean_rnnlm-hs-0.1b_h150-me5-1000_L0.75/wer_13_1.0
|
||||
|
||||
%WER 9.39 [ 5106 / 54402, 597 ins, 648 del, 3861 sub ] exp/tri6b/decode_tgmed_dev_clean/wer_14
|
||||
%WER 8.09 [ 4400 / 54402, 564 ins, 517 del, 3319 sub ] exp/tri6b/decode_tgmed_dev_clean_rnnlm_h150_me5-1000_L0.25/wer_15
|
||||
%WER 8.00 [ 4350 / 54402, 609 ins, 472 del, 3269 sub ] exp/tri6b/decode_tgmed_dev_clean_rnnlm_h150_me5-1000_L0.5/wer_15
|
||||
%WER 8.21 [ 4467 / 54402, 692 ins, 415 del, 3360 sub ] exp/tri6b/decode_tgmed_dev_clean_rnnlm_h150_me5-1000_L0.75/wer_12
|
||||
%WER 24.27 [ 12365 / 50948, 1365 ins, 1591 del, 9409 sub ] exp/tri6b/decode_tgmed_dev_other/wer_17_0.0
|
||||
%WER 22.76 [ 11596 / 50948, 1557 ins, 1239 del, 8800 sub ] exp/tri6b/decode_tgmed_dev_other_rnnlm-hs-0.1b_h150-me5-1000_L0.25/wer_15_0.0
|
||||
%WER 22.50 [ 11465 / 50948, 1586 ins, 1155 del, 8724 sub ] exp/tri6b/decode_tgmed_dev_other_rnnlm-hs-0.1b_h150-me5-1000_L0.5/wer_15_0.0
|
||||
%WER 22.56 [ 11495 / 50948, 1597 ins, 1143 del, 8755 sub ] exp/tri6b/decode_tgmed_dev_other_rnnlm-hs-0.1b_h150-me5-1000_L0.75/wer_17_0.0
|
||||
|
||||
%WER 25.16 [ 12816 / 50948, 1175 ins, 2076 del, 9565 sub ] exp/tri6b/decode_tgmed_dev_other/wer_16
|
||||
%WER 23.28 [ 11861 / 50948, 1289 ins, 1546 del, 9026 sub ] exp/tri6b/decode_tgmed_dev_other_rnnlm_h150_me5-1000_L0.25/wer_14
|
||||
%WER 23.03 [ 11732 / 50948, 1341 ins, 1467 del, 8924 sub ] exp/tri6b/decode_tgmed_dev_other_rnnlm_h150_me5-1000_L0.5/wer_14
|
||||
%WER 23.12 [ 11779 / 50948, 1351 ins, 1476 del, 8952 sub ] exp/tri6b/decode_tgmed_dev_other_rnnlm_h150_me5-1000_L0.75/wer_15
|
||||
%WER 10.50 [ 5711 / 54402, 693 ins, 674 del, 4344 sub ] exp/tri6b/decode_tgsmall_dev_clean/wer_15_0.0
|
||||
%WER 8.60 [ 4680 / 54402, 700 ins, 445 del, 3535 sub ] exp/tri6b/decode_tgsmall_dev_clean_rnnlm-hs-0.1b_h150-me5-1000_L0.25/wer_14_0.0
|
||||
%WER 8.17 [ 4447 / 54402, 723 ins, 352 del, 3372 sub ] exp/tri6b/decode_tgsmall_dev_clean_rnnlm-hs-0.1b_h150-me5-1000_L0.5/wer_13_0.0
|
||||
%WER 8.34 [ 4536 / 54402, 732 ins, 354 del, 3450 sub ] exp/tri6b/decode_tgsmall_dev_clean_rnnlm-hs-0.1b_h150-me5-1000_L0.75/wer_14_0.0
|
||||
|
||||
%WER 10.66 [ 5800 / 54402, 558 ins, 854 del, 4388 sub ] exp/tri6b/decode_tgsmall_dev_clean/wer_15
|
||||
%WER 8.78 [ 4779 / 54402, 586 ins, 588 del, 3605 sub ] exp/tri6b/decode_tgsmall_dev_clean_rnnlm_h150_me5-1000_L0.25/wer_14
|
||||
%WER 8.50 [ 4624 / 54402, 661 ins, 505 del, 3458 sub ] exp/tri6b/decode_tgsmall_dev_clean_rnnlm_h150_me5-1000_L0.5/wer_13
|
||||
%WER 8.56 [ 4659 / 54402, 674 ins, 485 del, 3500 sub ] exp/tri6b/decode_tgsmall_dev_clean_rnnlm_h150_me5-1000_L0.75/wer_13
|
||||
%WER 26.22 [ 13358 / 50948, 1330 ins, 1955 del, 10073 sub ] exp/tri6b/decode_tgsmall_dev_other/wer_17_0.0
|
||||
%WER 24.06 [ 12258 / 50948, 1526 ins, 1394 del, 9338 sub ] exp/tri6b/decode_tgsmall_dev_other_rnnlm-hs-0.1b_h150-me5-1000_L0.25/wer_15_0.0
|
||||
%WER 23.53 [ 11988 / 50948, 1585 ins, 1267 del, 9136 sub ] exp/tri6b/decode_tgsmall_dev_other_rnnlm-hs-0.1b_h150-me5-1000_L0.5/wer_15_0.0
|
||||
%WER 23.61 [ 12030 / 50948, 1571 ins, 1268 del, 9191 sub ] exp/tri6b/decode_tgsmall_dev_other_rnnlm-hs-0.1b_h150-me5-1000_L0.75/wer_18_0.0
|
||||
|
||||
%WER 27.18 [ 13850 / 50948, 1192 ins, 2340 del, 10318 sub ] exp/tri6b/decode_tgsmall_dev_other/wer_15
|
||||
%WER 24.72 [ 12596 / 50948, 1291 ins, 1803 del, 9502 sub ] exp/tri6b/decode_tgsmall_dev_other_rnnlm_h150_me5-1000_L0.25/wer_14
|
||||
%WER 24.18 [ 12317 / 50948, 1284 ins, 1732 del, 9301 sub ] exp/tri6b/decode_tgsmall_dev_other_rnnlm_h150_me5-1000_L0.5/wer_15
|
||||
%WER 24.19 [ 12323 / 50948, 1327 ins, 1686 del, 9310 sub ] exp/tri6b/decode_tgsmall_dev_other_rnnlm_h150_me5-1000_L0.75/wer_15
|
||||
# RNNLM rescoring of tri6b (faster-rnnlm hidden=150 direct=4.0Gb, Hierarchical Softmax)
|
||||
%WER 7.39 [ 4023 / 54402, 540 ins, 444 del, 3039 sub ] exp/tri6b/decode_tglarge_dev_clean/wer_13_1.0
|
||||
%WER 7.03 [ 3823 / 54402, 608 ins, 343 del, 2872 sub ] exp/tri6b/decode_tglarge_dev_clean_faster-rnnlm_h150-me5-1000_L0.25/wer_13_0.5
|
||||
%WER 7.03 [ 3827 / 54402, 606 ins, 320 del, 2901 sub ] exp/tri6b/decode_tglarge_dev_clean_faster-rnnlm_h150-me5-1000_L0.5/wer_14_0.5
|
||||
%WER 7.25 [ 3946 / 54402, 564 ins, 368 del, 3014 sub ] exp/tri6b/decode_tglarge_dev_clean_faster-rnnlm_h150-me5-1000_L0.75/wer_14_1.0
|
||||
|
||||
%WER 21.31 [ 10858 / 50948, 1525 ins, 1151 del, 8182 sub ] exp/tri6b/decode_tglarge_dev_other/wer_17_0.0
|
||||
%WER 20.62 [ 10504 / 50948, 1377 ins, 1180 del, 7947 sub ] exp/tri6b/decode_tglarge_dev_other_faster-rnnlm_h150-me5-1000_L0.25/wer_15_0.5
|
||||
%WER 20.64 [ 10515 / 50948, 1253 ins, 1313 del, 7949 sub ] exp/tri6b/decode_tglarge_dev_other_faster-rnnlm_h150-me5-1000_L0.5/wer_16_1.0
|
||||
%WER 20.91 [ 10652 / 50948, 1344 ins, 1233 del, 8075 sub ] exp/tri6b/decode_tglarge_dev_other_faster-rnnlm_h150-me5-1000_L0.75/wer_15_1.0
|
||||
|
||||
%WER 9.21 [ 5012 / 54402, 703 ins, 510 del, 3799 sub ] exp/tri6b/decode_tgmed_dev_clean/wer_14_0.0
|
||||
%WER 7.99 [ 4345 / 54402, 554 ins, 487 del, 3304 sub ] exp/tri6b/decode_tgmed_dev_clean_faster-rnnlm_h150-me5-1000_L0.25/wer_15_0.5
|
||||
%WER 7.68 [ 4177 / 54402, 596 ins, 414 del, 3167 sub ] exp/tri6b/decode_tgmed_dev_clean_faster-rnnlm_h150-me5-1000_L0.5/wer_14_0.5
|
||||
%WER 7.70 [ 4190 / 54402, 582 ins, 422 del, 3186 sub ] exp/tri6b/decode_tgmed_dev_clean_faster-rnnlm_h150-me5-1000_L0.75/wer_13_1.0
|
||||
|
||||
%WER 24.27 [ 12365 / 50948, 1365 ins, 1591 del, 9409 sub ] exp/tri6b/decode_tgmed_dev_other/wer_17_0.0
|
||||
%WER 22.51 [ 11468 / 50948, 1496 ins, 1235 del, 8737 sub ] exp/tri6b/decode_tgmed_dev_other_faster-rnnlm_h150-me5-1000_L0.25/wer_15_0.0
|
||||
%WER 22.11 [ 11267 / 50948, 1494 ins, 1163 del, 8610 sub ] exp/tri6b/decode_tgmed_dev_other_faster-rnnlm_h150-me5-1000_L0.5/wer_16_0.0
|
||||
%WER 22.10 [ 11262 / 50948, 1532 ins, 1131 del, 8599 sub ] exp/tri6b/decode_tgmed_dev_other_faster-rnnlm_h150-me5-1000_L0.75/wer_16_0.0
|
||||
|
||||
%WER 10.50 [ 5711 / 54402, 693 ins, 674 del, 4344 sub ] exp/tri6b/decode_tgsmall_dev_clean/wer_15_0.0
|
||||
%WER 8.53 [ 4641 / 54402, 582 ins, 555 del, 3504 sub ] exp/tri6b/decode_tgsmall_dev_clean_faster-rnnlm_h150-me5-1000_L0.25/wer_14_0.5
|
||||
%WER 8.09 [ 4400 / 54402, 605 ins, 469 del, 3326 sub ] exp/tri6b/decode_tgsmall_dev_clean_faster-rnnlm_h150-me5-1000_L0.5/wer_14_0.5
|
||||
%WER 8.02 [ 4363 / 54402, 594 ins, 460 del, 3309 sub ] exp/tri6b/decode_tgsmall_dev_clean_faster-rnnlm_h150-me5-1000_L0.75/wer_13_1.0
|
||||
|
||||
%WER 26.22 [ 13358 / 50948, 1330 ins, 1955 del, 10073 sub ] exp/tri6b/decode_tgsmall_dev_other/wer_17_0.0
|
||||
%WER 23.95 [ 12202 / 50948, 1523 ins, 1381 del, 9298 sub ] exp/tri6b/decode_tgsmall_dev_other_faster-rnnlm_h150-me5-1000_L0.25/wer_14_0.0
|
||||
%WER 23.22 [ 11828 / 50948, 1553 ins, 1247 del, 9028 sub ] exp/tri6b/decode_tgsmall_dev_other_faster-rnnlm_h150-me5-1000_L0.5/wer_14_0.0
|
||||
%WER 23.22 [ 11832 / 50948, 1435 ins, 1376 del, 9021 sub ] exp/tri6b/decode_tgsmall_dev_other_faster-rnnlm_h150-me5-1000_L0.75/wer_15_0.5
|
||||
|
||||
# RNNLM rescoring of tri6b (faster-rnnlm hidden=150 direct=1.6Gb Noise contrastive Estimation)
|
||||
%WER 7.39 [ 4023 / 54402, 540 ins, 444 del, 3039 sub ] exp/tri6b/decode_tglarge_dev_clean/wer_13_1.0
|
||||
%WER 7.05 [ 3835 / 54402, 487 ins, 447 del, 2901 sub ] exp/tri6b/decode_tglarge_dev_clean_faster-rnnlm_h150-me3-400-nce20_L0.25/wer_15_1.0
|
||||
%WER 6.84 [ 3723 / 54402, 524 ins, 394 del, 2805 sub ] exp/tri6b/decode_tglarge_dev_clean_faster-rnnlm_h150-me3-400-nce20_L0.5/wer_13_1.0
|
||||
%WER 6.92 [ 3766 / 54402, 564 ins, 376 del, 2826 sub ] exp/tri6b/decode_tglarge_dev_clean_faster-rnnlm_h150-me3-400-nce20_L0.75/wer_12_1.0
|
||||
|
||||
%WER 21.31 [ 10858 / 50948, 1525 ins, 1151 del, 8182 sub ] exp/tri6b/decode_tglarge_dev_other/wer_17_0.0
|
||||
%WER 20.90 [ 10648 / 50948, 1404 ins, 1227 del, 8017 sub ] exp/tri6b/decode_tglarge_dev_other_faster-rnnlm_h150-me3-400-nce20_L0.25/wer_15_0.5
|
||||
%WER 20.70 [ 10544 / 50948, 1271 ins, 1364 del, 7909 sub ] exp/tri6b/decode_tglarge_dev_other_faster-rnnlm_h150-me3-400-nce20_L0.5/wer_15_1.0
|
||||
%WER 20.82 [ 10605 / 50948, 1295 ins, 1347 del, 7963 sub ] exp/tri6b/decode_tglarge_dev_other_faster-rnnlm_h150-me3-400-nce20_L0.75/wer_15_1.0
|
||||
|
||||
%WER 9.21 [ 5012 / 54402, 703 ins, 510 del, 3799 sub ] exp/tri6b/decode_tgmed_dev_clean/wer_14_0.0
|
||||
%WER 8.01 [ 4360 / 54402, 669 ins, 402 del, 3289 sub ] exp/tri6b/decode_tgmed_dev_clean_faster-rnnlm_h150-me3-400-nce20_L0.25/wer_14_0.0
|
||||
%WER 7.46 [ 4056 / 54402, 584 ins, 422 del, 3050 sub ] exp/tri6b/decode_tgmed_dev_clean_faster-rnnlm_h150-me3-400-nce20_L0.5/wer_14_0.5
|
||||
%WER 7.28 [ 3962 / 54402, 536 ins, 451 del, 2975 sub ] exp/tri6b/decode_tgmed_dev_clean_faster-rnnlm_h150-me3-400-nce20_L0.75/wer_14_1.0
|
||||
|
||||
%WER 24.27 [ 12365 / 50948, 1365 ins, 1591 del, 9409 sub ] exp/tri6b/decode_tgmed_dev_other/wer_17_0.0
|
||||
%WER 22.82 [ 11628 / 50948, 1530 ins, 1244 del, 8854 sub ] exp/tri6b/decode_tgmed_dev_other_faster-rnnlm_h150-me3-400-nce20_L0.25/wer_15_0.0
|
||||
%WER 22.21 [ 11315 / 50948, 1554 ins, 1152 del, 8609 sub ] exp/tri6b/decode_tgmed_dev_other_faster-rnnlm_h150-me3-400-nce20_L0.5/wer_15_0.0
|
||||
%WER 22.01 [ 11213 / 50948, 1609 ins, 1086 del, 8518 sub ] exp/tri6b/decode_tgmed_dev_other_faster-rnnlm_h150-me3-400-nce20_L0.75/wer_15_0.0
|
||||
|
||||
%WER 10.50 [ 5711 / 54402, 693 ins, 674 del, 4344 sub ] exp/tri6b/decode_tgsmall_dev_clean/wer_15_0.0
|
||||
%WER 8.56 [ 4659 / 54402, 677 ins, 467 del, 3515 sub ] exp/tri6b/decode_tgsmall_dev_clean_faster-rnnlm_h150-me3-400-nce20_L0.25/wer_14_0.0
|
||||
%WER 7.81 [ 4250 / 54402, 657 ins, 387 del, 3206 sub ] exp/tri6b/decode_tgsmall_dev_clean_faster-rnnlm_h150-me3-400-nce20_L0.5/wer_14_0.0
|
||||
%WER 7.58 [ 4125 / 54402, 618 ins, 406 del, 3101 sub ] exp/tri6b/decode_tgsmall_dev_clean_faster-rnnlm_h150-me3-400-nce20_L0.75/wer_13_0.5
|
||||
|
||||
%WER 26.22 [ 13358 / 50948, 1330 ins, 1955 del, 10073 sub ] exp/tri6b/decode_tgsmall_dev_other/wer_17_0.0
|
||||
%WER 24.07 [ 12264 / 50948, 1482 ins, 1435 del, 9347 sub ] exp/tri6b/decode_tgsmall_dev_other_faster-rnnlm_h150-me3-400-nce20_L0.25/wer_15_0.0
|
||||
%WER 23.15 [ 11797 / 50948, 1526 ins, 1276 del, 8995 sub ] exp/tri6b/decode_tgsmall_dev_other_faster-rnnlm_h150-me3-400-nce20_L0.5/wer_15_0.0
|
||||
%WER 22.92 [ 11677 / 50948, 1544 ins, 1241 del, 8892 sub ] exp/tri6b/decode_tgsmall_dev_other_faster-rnnlm_h150-me3-400-nce20_L0.75/wer_16_0.0
|
||||
|
||||
## Multi-splice version of online recipe.
|
||||
# for x in exp/nnet2_online/nnet_ms_a/decode_*; do grep WER $x/wer_* | utils/best_wer.sh ; done
|
||||
|
|
|
|||
|
|
@ -3,12 +3,12 @@
|
|||
# Copyright 2014 Yandex (Author: Ilya Edrenkin)
|
||||
# Apache 2.0
|
||||
|
||||
# Begin configuration section.
|
||||
hidden=150
|
||||
maxent_order=5
|
||||
maxent_size=1000
|
||||
num_threads=16
|
||||
# Begin configuration section.
|
||||
rnnlm_options="-hidden 150 -direct 1000 -direct-order 5"
|
||||
rnnlm_tag="h150_me5-1000"
|
||||
num_threads=8 # set this value to the number of physical cores on your CPU
|
||||
stage=0
|
||||
rnnlm_ver=rnnlm-hs-0.1b
|
||||
# End configuration section.
|
||||
|
||||
echo "$0 $@" # Print the command line for logging
|
||||
|
|
@ -25,9 +25,8 @@ if [ $# -ne 2 ]; then
|
|||
echo " <data-dir> is the directory in which the text corpus is downloaded"
|
||||
echo " <lm-dir> is the directory in which the language model is stored"
|
||||
echo "Main options:"
|
||||
echo " --hidden <int> # default 150. Hidden layer size"
|
||||
echo " --maxent-order <int> # default 5. Maxent features order size"
|
||||
echo " --maxent-size <int> # default 1000. Maxent features hash size"
|
||||
echo " --rnnlm-options <int> # default '$rnnlm_options'. Command line arguments to pass to rnnlm"
|
||||
echo " --rnnlm-tag <str> # default '$rnnlm_tag' The tag is appended to exp/ folder name"
|
||||
echo " --num-threads <int> # default 16. Number of concurrent threadss to train RNNLM"
|
||||
echo " --stage <int> # 1 to download and prepare data, 2 to train RNNLM, 3 to rescore tri6b with a trained RNNLM"
|
||||
exit 1
|
||||
|
|
@ -36,51 +35,79 @@ fi
|
|||
s5_dir=`pwd`
|
||||
data_dir=`readlink -f $1`
|
||||
lm_dir=`readlink -f $2`
|
||||
rnnlm_ver=rnnlm-hs-0.1b # Probably could make this an option, but Tomas's RNN will take long to train on 200K vocab
|
||||
rnnlmdir=data/lang_rnnlm_h${hidden}_me${maxent_order}-${maxent_size}
|
||||
export PATH=$KALDI_ROOT/tools/$rnnlm_ver:$PATH
|
||||
modeldir=data/lang_${rnnlm_ver}_${rnnlm_tag}
|
||||
|
||||
if [ $stage -le 1 ]; then
|
||||
echo "$0: Prepare training data for RNNLM"
|
||||
cd $data_dir
|
||||
wget http://www.openslr.org/resources/11/librispeech-lm-norm.txt.gz
|
||||
gunzip librispeech-lm-norm.txt.gz
|
||||
$s5_dir/utils/filt.py $lm_dir/librispeech-vocab.txt librispeech-lm-norm.txt | shuf > librispeech-lm-norm.train.txt
|
||||
$s5_dir/utils/filt.py $lm_dir/librispeech-vocab.txt <(awk '{$1=""; print $0}' $s5_dir/data/train_960/text) > librispeech-lm-norm.dev.txt
|
||||
rm librispeech-lm-norm.txt
|
||||
if [ -f "librispeech-lm-norm.dev.txt" ]; then
|
||||
echo "$0: SKIP File librispeech-lm-norm.dev.txt already exists"
|
||||
else
|
||||
wget http://www.openslr.org/resources/11/librispeech-lm-norm.txt.gz
|
||||
gunzip librispeech-lm-norm.txt.gz
|
||||
$s5_dir/utils/filt.py $lm_dir/librispeech-vocab.txt librispeech-lm-norm.txt | shuf > librispeech-lm-norm.train.txt
|
||||
$s5_dir/utils/filt.py $lm_dir/librispeech-vocab.txt <(awk '{$1=""; print $0}' $s5_dir/data/train_960/text) > librispeech-lm-norm.dev.txt.tmp
|
||||
mv librispeech-lm-norm.dev.txt.tmp librispeech-lm-norm.dev.txt
|
||||
rm librispeech-lm-norm.txt
|
||||
fi
|
||||
cd $s5_dir
|
||||
|
||||
|
||||
fi
|
||||
|
||||
if [ $stage -le 2 ]; then
|
||||
echo "$0: Training RNNLM. It will probably take several hours."
|
||||
cd $KALDI_ROOT/tools
|
||||
if [ -f $rnnlm_ver/rnnlm ]; then
|
||||
echo "Not installing the rnnlm toolkit since it is already there."
|
||||
else
|
||||
echo "$0: Not installing the rnnlm toolkit since it is already there."
|
||||
elif [ $rnnlm_ver == "rnnlm-hs-0.1b" ]; then
|
||||
extras/install_rnnlm_hs.sh
|
||||
elif [ $rnnlm_ver == "faster-rnnlm" ]; then
|
||||
extras/install_faster_rnnlm.sh
|
||||
else
|
||||
echo "$0: ERROR Cannot find $rnnlm_ver"
|
||||
exit 1
|
||||
fi
|
||||
cd $s5_dir
|
||||
mkdir -p $rnnlmdir
|
||||
rnnlm -rnnlm $rnnlmdir/rnnlm -train $data_dir/librispeech-lm-norm.train.txt -valid $data_dir/librispeech-lm-norm.dev.txt \
|
||||
-threads $num_threads -hidden $hidden -direct-order $maxent_order -direct $maxent_size -retry 1 -stop 1.0
|
||||
touch $rnnlmdir/unk.probs
|
||||
awk '{print $1}' $rnnlmdir/rnnlm > $rnnlmdir/wordlist.rnn
|
||||
mkdir -p $modeldir
|
||||
echo "$0: Model file: $modeldir/rnnlm"
|
||||
if [ -f "$modeldir/rnnlm" ]; then
|
||||
echo "$0: SKIP file '$modeldir/rnnlm' already exists"
|
||||
else
|
||||
rm -f $modeldir/rnnlm.tmp
|
||||
rnnlm_path="$(readlink -f "$KALDI_ROOT/tools/$rnnlm_ver/rnnlm")"
|
||||
rnnlm_cmd="$rnnlm_path"
|
||||
if type taskset >/dev/null 2>&1 ; then
|
||||
# HogWild works much faster if all threads are binded to the same phisical cpu
|
||||
rnnlm_cmd="taskset -c $(seq -s, 0 $(( $num_threads - 1 )) ) $rnnlm_cmd"
|
||||
fi
|
||||
$rnnlm_cmd -rnnlm $modeldir/rnnlm.tmp \
|
||||
-train $data_dir/librispeech-lm-norm.train.txt \
|
||||
-valid $data_dir/librispeech-lm-norm.dev.txt \
|
||||
-threads $num_threads $rnnlm_options -retry 1 -stop 1.0 2>&1 | tee $modeldir/rnnlm.log
|
||||
touch $modeldir/unk.probs
|
||||
awk '{print $1}' $modeldir/rnnlm.tmp > $modeldir/wordlist.rnn
|
||||
mv $modeldir/rnnlm.tmp $modeldir/rnnlm
|
||||
mv $modeldir/rnnlm.tmp.nnet $modeldir/rnnlm.nnet
|
||||
fi
|
||||
fi
|
||||
|
||||
if [ $stage -le 3 ]; then
|
||||
echo "$0: Performing RNNLM rescoring on tri6b decoding results"
|
||||
for lm in tgsmall tgmed; do
|
||||
for lm in tgsmall tgmed tglarge; do
|
||||
for devset in dev_clean dev_other; do
|
||||
sourcedir=exp/tri6b/decode_${lm}_${devset}
|
||||
resultsdir=${sourcedir}_rnnlm_h${hidden}_me${maxent_order}-${maxent_size}
|
||||
steps/rnnlmrescore.sh --rnnlm_ver $rnnlm_ver --N 100 0.5 data/lang_test_$lm $rnnlmdir data/$devset $sourcedir ${resultsdir}_L0.5
|
||||
cp -r ${resultsdir}_L0.5 ${resultsdir}_L0.25
|
||||
cp -r ${resultsdir}_L0.5 ${resultsdir}_L0.75
|
||||
steps/rnnlmrescore.sh --rnnlm_ver $rnnlm_ver --N 100 --stage 7 0.25 data/lang_test_$lm $rnnlmdir data/$devset $sourcedir ${resultsdir}_L0.25
|
||||
steps/rnnlmrescore.sh --rnnlm_ver $rnnlm_ver --N 100 --stage 7 0.75 data/lang_test_$lm $rnnlmdir data/$devset $sourcedir ${resultsdir}_L0.75
|
||||
if [ ! -d "$sourcedir" ]; then
|
||||
echo "$0: WARNING cannot find source dir '$sourcedir' to rescore"
|
||||
continue
|
||||
fi
|
||||
resultsdir=${sourcedir}_${rnnlm_ver}_${rnnlm_tag}
|
||||
rm -rf ${resultsdir}_L0.5
|
||||
steps/rnnlmrescore.sh --skip_scoring false --rnnlm_ver $rnnlm_ver --N 100 0.5 data/lang_test_$lm $modeldir data/$devset $sourcedir ${resultsdir}_L0.5
|
||||
for coef in 0.25 0.75; do
|
||||
rm -rf ${resultsdir}_L${coef}
|
||||
cp -r ${resultsdir}_L0.5 ${resultsdir}_L${coef}
|
||||
steps/rnnlmrescore.sh --skip_scoring false --rnnlm_ver $rnnlm_ver --N 100 --stage 7 $coef data/lang_test_$lm $modeldir data/$devset $sourcedir ${resultsdir}_L${coef}
|
||||
done
|
||||
done
|
||||
done
|
||||
fi
|
||||
|
||||
|
||||
|
|
|
|||
|
|
@ -330,9 +330,31 @@ steps/train_quick.sh --cmd "$train_cmd" \
|
|||
# steps/cleanup/debug_lexicon.sh --remove-stress true --nj 200 --cmd "$train_cmd" data/train_clean_100 \
|
||||
# data/lang exp/tri6b data/local/dict/lexicon.txt exp/debug_lexicon_100h
|
||||
|
||||
# #Perform RNNLM rescoring of tri6b
|
||||
# #Perform rescoring of tri6b be means of rnnlm-hs
|
||||
# #Attention: with default settings requires 4 GB of memory per rescoring job, so commenting this out by default
|
||||
# local/run_rnnlm.sh $data data/local/lm
|
||||
# wait && local/run_rnnlm.sh \
|
||||
# --rnnlm-ver "rnnlm-hs-0.1b" \
|
||||
# --rnnlm-options "-hidden 150 -direct 1000 -direct-order 5" \
|
||||
# --rnnlm_tag "h150-me5-1000" $data data/local/lm
|
||||
|
||||
# #Perform rescoring of tri6b be means of faster-rnnlm
|
||||
# #Results are identical to results obtained by rnnlm-hs (or better)
|
||||
# #Hovewer, faster-rnnlm training is several times faster
|
||||
# wait && local/run_rnnlm.sh \
|
||||
# --rnnlm-ver "faster-rnnlm" \
|
||||
# --rnnlm-options "-hidden 150 -direct 1000 -direct-order 5" \
|
||||
# --rnnlm-tag "h150-me5-1000" $data data/local/lm
|
||||
|
||||
# #Perform rescoring of tri6b be means of faster-rnnlm using Noise contrastive estimation
|
||||
# #Note, that could be extremely slow without CUDA
|
||||
# #We use smaller direct layer size so that it could be stored in GPU memory (~2Gb)
|
||||
# #Suprisingly, bottleneck here is validation rather then learning
|
||||
# #Therefore you can use smaller validation dataset to speed up training
|
||||
# wait && local/run_rnnlm.sh \
|
||||
# --rnnlm-ver "faster-rnnlm" \
|
||||
# --rnnlm-options "-hidden 150 -direct 400 -direct-order 3 --nce 20" \
|
||||
# --rnnlm-tag "h150-me3-400-nce20" $data data/local/lm
|
||||
|
||||
|
||||
# train NN models on the entire dataset
|
||||
local/nnet2/run_7a_960.sh || exit 1
|
||||
|
|
|
|||
|
|
@ -0,0 +1,25 @@
|
|||
#!/bin/bash
|
||||
|
||||
# The script downloads and installs faster-rnnlm
|
||||
# https://github.com/yandex/faster-rnnlm
|
||||
|
||||
set -e
|
||||
|
||||
# Make sure we are in the tools/ directory.
|
||||
if [ `basename $PWD` == extras ]; then
|
||||
cd ..
|
||||
fi
|
||||
|
||||
! [ `basename $PWD` == tools ] && \
|
||||
echo "You must call this script from the tools/ directory" && exit 1;
|
||||
|
||||
echo "Installing Faster RNNLM"
|
||||
|
||||
if [ ! -d "faster-rnnlm" ]; then
|
||||
git clone https://github.com/yandex/faster-rnnlm.git
|
||||
fi
|
||||
|
||||
cd faster-rnnlm
|
||||
git pull
|
||||
./build.sh
|
||||
ln -sf faster-rnnlm/rnnlm
|
||||
Загрузка…
Ссылка в новой задаче