ASLP-lab commited on
Commit
e56d177
·
verified ·
1 Parent(s): 391b67b

Upload 2 files

Browse files
u2pp_conformer_yue/bpe.model ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:ee1b3b4dbe74bbad77f76efd271577db44377ca5a502a8aa44ac4d7dfd58c2d2
3
+ size 253122
u2pp_conformer_yue/train.yaml ADDED
@@ -0,0 +1,95 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ accum_grad: 4
2
+ ctc: ctc
3
+ ctc_conf:
4
+ ctc_blank_id: 0
5
+ dataset_conf:
6
+ batch_conf:
7
+ batch_type: dynamic
8
+ max_frames_in_batch: 36000
9
+ fbank_conf:
10
+ dither: 0.1
11
+ frame_length: 25
12
+ frame_shift: 10
13
+ num_mel_bins: 80
14
+ filter_conf:
15
+ max_length: 10240
16
+ min_length: 10
17
+ token_max_length: 200
18
+ token_min_length: 1
19
+ resample_conf:
20
+ resample_rate: 16000
21
+ shuffle: true
22
+ shuffle_conf:
23
+ shuffle_size: 1500
24
+ sort: true
25
+ sort_conf:
26
+ sort_size: 500
27
+ spec_aug: true
28
+ spec_aug_conf:
29
+ max_f: 10
30
+ max_t: 50
31
+ num_f_mask: 2
32
+ num_t_mask: 2
33
+ speed_perturb: true
34
+ decoder: bitransformer
35
+ decoder_conf:
36
+ attention_heads: 8
37
+ dropout_rate: 0.1
38
+ linear_units: 2048
39
+ num_blocks: 3
40
+ positional_dropout_rate: 0.1
41
+ r_num_blocks: 3
42
+ self_attention_dropout_rate: 0.1
43
+ src_attention_dropout_rate: 0.1
44
+ dtype: fp32
45
+ encoder: conformer
46
+ encoder_conf:
47
+ activation_type: swish
48
+ attention_dropout_rate: 0.1
49
+ attention_heads: 8
50
+ cnn_module_kernel: 15
51
+ dropout_rate: 0.1
52
+ input_layer: conv2d
53
+ linear_units: 2048
54
+ normalize_before: true
55
+ num_blocks: 12
56
+ output_size: 512
57
+ pos_enc_layer_type: rel_pos
58
+ positional_dropout_rate: 0.1
59
+ selfattention_layer_type: rel_selfattn
60
+ use_cnn_module: true
61
+ grad_clip: 5
62
+ input_dim: 80
63
+ log_interval: 100
64
+ max_epoch: 20
65
+ model: asr_model
66
+ model_conf:
67
+ ctc_weight: 0.3
68
+ length_normalized_loss: false
69
+ lsm_weight: 0.1
70
+ reverse_weight: 0.3
71
+ model_dir: /home/work_nfs17/lhli/workspace/checkpoint/WenetSpeech_Yue_open_source
72
+ optim: adam
73
+ optim_conf:
74
+ lr: 0.001
75
+ output_dim: 8629
76
+ save_states: model_only
77
+ scheduler: warmuplr
78
+ scheduler_conf:
79
+ warmup_steps: 25000
80
+ tokenizer: bpe
81
+ tokenizer_conf:
82
+ bpe_path: bpe.model
83
+ is_multilingual: true
84
+ non_lang_syms_path: null
85
+ num_languages: 2
86
+ special_tokens:
87
+ <blank>: 0
88
+ <eos>: 8628
89
+ <sos>: 8628
90
+ <unk>: 1
91
+ split_with_space: false
92
+ symbol_table_path: data/dict_bpe/lang_char.txt
93
+ train_engine: torch_ddp
94
+ use_amp: false
95
+ vocab_size: 8629