Chainer CSJ results

[kan-bayashi]

I finished testing CSJ recipe.

exp/train_nodup_vggblstmp_e4_subsample1_2_2_1_1_unit320_proj320_d1_unit300_location_aconvc10_aconvf100_mtlalpha0.5_adadelta_bs30_mli800_mlo150/decode_eval1_beam20_eacc.best_p0_len0.0-0.8/result.txt:
|        Sum/Avg         |        1272                 43897        |        84.9                  6.2                   8.9                  1.4                 16.5                 70.6        |
exp/train_nodup_vggblstmp_e4_subsample1_2_2_1_1_unit320_proj320_d1_unit300_location_aconvc10_aconvf100_mtlalpha0.5_adadelta_bs30_mli800_mlo150/decode_eval2_beam20_eacc.best_p0_len0.0-0.8/result.txt:
|        Sum/Avg         |        1292                 43623        |        89.2                  5.0                   5.8                  1.0                 11.7                 65.9        |
exp/train_nodup_vggblstmp_e4_subsample1_2_2_1_1_unit320_proj320_d1_unit300_location_aconvc10_aconvf100_mtlalpha0.5_adadelta_bs30_mli800_mlo150/decode_eval3_beam20_eacc.best_p0_len0.0-0.8/result.txt:
|        Sum/Avg         |        1385                 28225        |        89.3                  5.6                   5.1                  1.6                 12.3                 53.8        |

[sw005320]

Thanks.
Can you change --maxlenratio 0.5 --minlenratio 0.1?

[kan-bayashi]

Here is the results.
Slightly worse than the results in papers.

$ grep -e Avg -e SPRK -m 2 exp/train_nodup_vggblstmp_e4_subsample1_2_2_1_1_unit320_proj320_d1_unit300_location_aconvc10_aconvf100_mtlalpha0.5_adadelta_bs30_mli800_mlo150/decode_eval*_beam20_eacc.best_p0_len0.1-0.5/result.txt
exp/train_nodup_vggblstmp_e4_subsample1_2_2_1_1_unit320_proj320_d1_unit300_location_aconvc10_aconvf100_mtlalpha0.5_adadelta_bs30_mli800_mlo150/decode_eval1_beam20_eacc.best_p0_len0.1-0.5/result.txt:
|        Sum/Avg         |        1272                 43897        |        90.1                  7.0                   2.9                  1.6                 11.4                 70.6        |
exp/train_nodup_vggblstmp_e4_subsample1_2_2_1_1_unit320_proj320_d1_unit300_location_aconvc10_aconvf100_mtlalpha0.5_adadelta_bs30_mli800_mlo150/decode_eval2_beam20_eacc.best_p0_len0.1-0.5/result.txt:
|        Sum/Avg         |        1292                 43623        |        93.2                  5.3                   1.5                  1.0                  7.8                 66.0        |
exp/train_nodup_vggblstmp_e4_subsample1_2_2_1_1_unit320_proj320_d1_unit300_location_aconvc10_aconvf100_mtlalpha0.5_adadelta_bs30_mli800_mlo150/decode_eval3_beam20_eacc.best_p0_len0.1-0.5/result.txt:
|        Sum/Avg         |        1385                 28225        |        92.2                  5.9                   1.9                  1.7                  9.5                 53.9        |

I will try to use penalty=0.1

[kan-bayashi]

Results with maxlenratio=0.5 & minlenratio=0.1 & penalty=0.1.
Slightly improved.

$ grep -e Avg -e SPRK -m 2 exp/train_nodup_vggblstmp_e4_subsample1_2_2_1_1_unit320_proj320_d1_unit300_location_aconvc10_aconvf100_mtlalpha0.5_adadelta_bs30_mli800_mlo150/decode_eval*_beam20_eacc.best_p0.1_len0.1-0.5/result.txt
exp/train_nodup_vggblstmp_e4_subsample1_2_2_1_1_unit320_proj320_d1_unit300_location_aconvc10_aconvf100_mtlalpha0.5_adadelta_bs30_mli800_mlo150/decode_eval1_beam20_eacc.best_p0.1_len0.1-0.5/result.txt:
|        Sum/Avg         |        1272                  43897        |        90.3                  7.0                   2.6                  1.6                 11.3                  70.5        |
exp/train_nodup_vggblstmp_e4_subsample1_2_2_1_1_unit320_proj320_d1_unit300_location_aconvc10_aconvf100_mtlalpha0.5_adadelta_bs30_mli800_mlo150/decode_eval2_beam20_eacc.best_p0.1_len0.1-0.5/result.txt:
|        Sum/Avg         |        1292                  43623        |        93.3                  5.3                   1.4                  1.1                  7.8                  66.1        |
exp/train_nodup_vggblstmp_e4_subsample1_2_2_1_1_unit320_proj320_d1_unit300_location_aconvc10_aconvf100_mtlalpha0.5_adadelta_bs30_mli800_mlo150/decode_eval3_beam20_eacc.best_p0.1_len0.1-0.5/result.txt:
|        Sum/Avg         |        1385                  28225        |        92.6                  6.0                   1.5                  1.8                  9.2                  53.9        |

[sw005320]

Thanks.
Please write them in RESULTS, and commit it.
It's good to know that our initial results are not so crazy.