From fa6f60fa762f271d096b8749f3cc9bfc61a6ed48 Mon Sep 17 00:00:00 2001
From: 游雁 <zhifu.gzf@alibaba-inc.com>
Date: 星期五, 23 二月 2024 14:01:44 +0800
Subject: [PATCH] update
---
examples/aishell/llm_asr_nar/conf/template.yaml | 34 ++++++++++++++++++----------------
1 files changed, 18 insertions(+), 16 deletions(-)
diff --git a/examples/aishell/llm_asr_nar/conf/template.yaml b/examples/aishell/llm_asr_nar/conf/template.yaml
index 0b26969..d529635 100644
--- a/examples/aishell/llm_asr_nar/conf/template.yaml
+++ b/examples/aishell/llm_asr_nar/conf/template.yaml
@@ -24,11 +24,11 @@
init_param_path: "/nfs/maziyang.mzy/models/vicuna-7b-v1.5"
freeze: true
-adaptor: linear
+adaptor: Linear
adaptor_conf:
downsample_rate: 1
llm_dim: 4096
- encoder_dim: 2048
+ encoder_dim: 512
# frontend related
frontend: WavFrontend
@@ -38,54 +38,56 @@
n_mels: 80
frame_length: 25
frame_shift: 10
- dither: 0.0
- lfr_m: 1
- lfr_n: 1
+ lfr_m: 7
+ lfr_n: 6
+ cmvn_file: "/root/.cache/modelscope/hub/iic/speech_paraformer-large_asr_nat-zh-cn-16k-common-vocab8404-pytorch/am.mvn"
-specaug: SpecAug
+specaug: SpecAugLFR
specaug_conf:
- apply_time_warp: true
+ apply_time_warp: false
time_warp_window: 5
time_warp_mode: bicubic
apply_freq_mask: true
freq_mask_width_range:
- 0
- 30
- num_freq_mask: 2
+ lfr_rate: 6
+ num_freq_mask: 1
apply_time_mask: true
time_mask_width_range:
- 0
- - 40
- num_time_mask: 2
+ - 12
+ num_time_mask: 1
train_conf:
accum_grad: 1
grad_clip: 5
max_epoch: 150
keep_nbest_models: 10
- log_interval: 50
+ log_interval: 10
-optim: adam
+optim: adamw
optim_conf:
- lr: 0.001
+ lr: 0.0001
weight_decay: 0.000001
scheduler: warmuplr
scheduler_conf:
- warmup_steps: 35000
+ warmup_steps: 1500
dataset: AudioLLMDataset
dataset_conf:
index_ds: IndexDSJsonl
batch_sampler: RankFullLocalShuffleBatchSampler
batch_type: example # example or length
- batch_size: 4 # if batch_type is example, batch_size is the numbers of samples; if length, batch_size is source_token_len+target_token_len;
+ batch_size: 8 # if batch_type is example, batch_size is the numbers of samples; if length, batch_size is source_token_len+target_token_len;
max_token_length: 2048 # filter samples if source_token_len+target_token_len > max_token_length,
buffer_size: 500
shuffle: True
num_workers: 4
+ preprocessor_text: TextPreprocessRemovePunctuation
tokenizer: HuggingfaceTokenizer
tokenizer_conf:
unk_symbol: <unk>
- init_param_path: null
+ init_param_path: "/nfs/maziyang.mzy/models/vicuna-7b-v1.5"
--
Gitblit v1.9.1