|
|
@ -34,18 +34,14 @@ training:
|
|
|
|
lr_decay: 0.83
|
|
|
|
lr_decay: 0.83
|
|
|
|
weight_decay: 1e-06
|
|
|
|
weight_decay: 1e-06
|
|
|
|
global_grad_clip: 5.0
|
|
|
|
global_grad_clip: 5.0
|
|
|
|
max_iteration: 500000
|
|
|
|
|
|
|
|
plot_interval: 1000
|
|
|
|
|
|
|
|
save_interval: 1000
|
|
|
|
|
|
|
|
valid_interval: 1000
|
|
|
|
|
|
|
|
decoding:
|
|
|
|
decoding:
|
|
|
|
batch_size: 10
|
|
|
|
batch_size: 128
|
|
|
|
error_rate_type: cer
|
|
|
|
error_rate_type: cer
|
|
|
|
decoding_method: ctc_beam_search
|
|
|
|
decoding_method: ctc_beam_search
|
|
|
|
lang_model_path: models/lm/zh_giga.no_cna_cmn.prune01244.klm
|
|
|
|
lang_model_path: models/lm/zh_giga.no_cna_cmn.prune01244.klm
|
|
|
|
alpha: 2.6
|
|
|
|
alpha: 2.6
|
|
|
|
beta: 5.0
|
|
|
|
beta: 5.0
|
|
|
|
beam_size: 300
|
|
|
|
beam_size: 300
|
|
|
|
cutoff_prob: 1.0
|
|
|
|
cutoff_prob: 0.99
|
|
|
|
cutoff_top_n: 40
|
|
|
|
cutoff_top_n: 40
|
|
|
|
num_proc_bsearch: 10
|
|
|
|
num_proc_bsearch: 10
|
|
|
|