Upload train.yaml
Browse files- whisper_medium_yue/train.yaml +121 -0
whisper_medium_yue/train.yaml
ADDED
|
@@ -0,0 +1,121 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
accum_grad: 4
|
| 2 |
+
cmvn: null
|
| 3 |
+
cmvn_conf:
|
| 4 |
+
cmvn_file: null
|
| 5 |
+
is_json_cmvn: null
|
| 6 |
+
ctc: ctc
|
| 7 |
+
ctc_conf:
|
| 8 |
+
ctc_blank_id: 50363
|
| 9 |
+
dataset: asr
|
| 10 |
+
dataset_conf:
|
| 11 |
+
batch_conf:
|
| 12 |
+
batch_type: dynamic
|
| 13 |
+
max_frames_in_batch: 8000
|
| 14 |
+
cycle: 6
|
| 15 |
+
feats_type: log_mel_spectrogram
|
| 16 |
+
filter_conf:
|
| 17 |
+
max_length: 2700
|
| 18 |
+
min_length: 0
|
| 19 |
+
token_max_length: 448
|
| 20 |
+
token_min_length: 1
|
| 21 |
+
log_mel_spectrogram_conf:
|
| 22 |
+
hop_length: 160
|
| 23 |
+
n_fft: 400
|
| 24 |
+
num_mel_bins: 80
|
| 25 |
+
padding: 0
|
| 26 |
+
resample_conf:
|
| 27 |
+
resample_rate: 16000
|
| 28 |
+
shuffle: true
|
| 29 |
+
shuffle_conf:
|
| 30 |
+
shuffle_size: 1500
|
| 31 |
+
sort: true
|
| 32 |
+
sort_conf:
|
| 33 |
+
sort_size: 500
|
| 34 |
+
spec_aug: true
|
| 35 |
+
spec_aug_conf:
|
| 36 |
+
max_f: 10
|
| 37 |
+
max_t: 50
|
| 38 |
+
num_f_mask: 2
|
| 39 |
+
num_t_mask: 2
|
| 40 |
+
spec_sub: true
|
| 41 |
+
spec_sub_conf:
|
| 42 |
+
max_t: 30
|
| 43 |
+
num_t_sub: 3
|
| 44 |
+
spec_trim: false
|
| 45 |
+
speed_perturb: true
|
| 46 |
+
decoder: transformer
|
| 47 |
+
decoder_conf:
|
| 48 |
+
activation_type: gelu
|
| 49 |
+
attention_heads: 16
|
| 50 |
+
dropout_rate: 0.1
|
| 51 |
+
gradient_checkpointing: true
|
| 52 |
+
input_layer: embed_learnable_pe
|
| 53 |
+
key_bias: false
|
| 54 |
+
linear_units: 4096
|
| 55 |
+
normalize_before: true
|
| 56 |
+
num_blocks: 24
|
| 57 |
+
positional_dropout_rate: 0.1
|
| 58 |
+
self_attention_dropout_rate: 0.0
|
| 59 |
+
src_attention: true
|
| 60 |
+
src_attention_dropout_rate: 0.0
|
| 61 |
+
src_key_bias: false
|
| 62 |
+
tie_word_embedding: true
|
| 63 |
+
use_output_layer: true
|
| 64 |
+
dtype: fp32
|
| 65 |
+
encoder: transformer
|
| 66 |
+
encoder_conf:
|
| 67 |
+
activation_type: gelu
|
| 68 |
+
attention_dropout_rate: 0.0
|
| 69 |
+
attention_heads: 16
|
| 70 |
+
dropout_rate: 0.0
|
| 71 |
+
gradient_checkpointing: true
|
| 72 |
+
input_layer: conv1d2
|
| 73 |
+
key_bias: false
|
| 74 |
+
linear_units: 4096
|
| 75 |
+
normalize_before: true
|
| 76 |
+
num_blocks: 24
|
| 77 |
+
output_size: 1024
|
| 78 |
+
pos_enc_layer_type: abs_pos_whisper
|
| 79 |
+
positional_dropout_rate: 0.1
|
| 80 |
+
static_chunk_size: -1
|
| 81 |
+
use_dynamic_chunk: false
|
| 82 |
+
use_dynamic_left_chunk: false
|
| 83 |
+
grad_clip: 5
|
| 84 |
+
input_dim: 80
|
| 85 |
+
log_interval: 100
|
| 86 |
+
max_epoch: 6
|
| 87 |
+
model: whisper
|
| 88 |
+
model_conf:
|
| 89 |
+
ctc_weight: 0.0
|
| 90 |
+
length_normalized_loss: false
|
| 91 |
+
lsm_weight: 0.1
|
| 92 |
+
model_dir: exp/whisper_medium_fintune_wenetspeech_yue_opensource
|
| 93 |
+
optim: adam
|
| 94 |
+
optim_conf:
|
| 95 |
+
lr: 1.0e-05
|
| 96 |
+
output_dim: 51865
|
| 97 |
+
save_interval: 2000
|
| 98 |
+
save_states: model_only
|
| 99 |
+
scheduler: warmuplr
|
| 100 |
+
scheduler_conf:
|
| 101 |
+
warmup_steps: 12000
|
| 102 |
+
tokenizer: whisper
|
| 103 |
+
tokenizer_conf:
|
| 104 |
+
bpe_path: null
|
| 105 |
+
is_multilingual: true
|
| 106 |
+
non_lang_syms_path: null
|
| 107 |
+
num_languages: 99
|
| 108 |
+
special_tokens:
|
| 109 |
+
eot: 50257
|
| 110 |
+
no_speech: 50362
|
| 111 |
+
no_timestamps: 50363
|
| 112 |
+
sot: 50258
|
| 113 |
+
sot_prev: 50361
|
| 114 |
+
timestamp_begin: 50364
|
| 115 |
+
transcribe: 50359
|
| 116 |
+
translate: 50358
|
| 117 |
+
split_with_space: false
|
| 118 |
+
symbol_table_path: null
|
| 119 |
+
train_engine: torch_ddp
|
| 120 |
+
use_amp: false
|
| 121 |
+
vocab_size: 51865
|