From 97a689d65da434345a641a909f13b78e5690c86b Mon Sep 17 00:00:00 2001
From: zhifu gao <zhifu.gzf@alibaba-inc.com>
Date: 星期四, 18 五月 2023 19:35:08 +0800
Subject: [PATCH] Merge pull request #526 from alibaba-damo-academy/dev_infer
---
egs/librispeech_100h/conformer/conf/train_asr_conformer.yaml | 32 ++++++++++++++++++++++----------
1 files changed, 22 insertions(+), 10 deletions(-)
diff --git a/egs/librispeech/conformer/conf/train_asr_conformer_uttnorm.yaml b/egs/librispeech_100h/conformer/conf/train_asr_conformer.yaml
similarity index 77%
rename from egs/librispeech/conformer/conf/train_asr_conformer_uttnorm.yaml
rename to egs/librispeech_100h/conformer/conf/train_asr_conformer.yaml
index 16b7cc0..bd92bb0 100644
--- a/egs/librispeech/conformer/conf/train_asr_conformer_uttnorm.yaml
+++ b/egs/librispeech_100h/conformer/conf/train_asr_conformer.yaml
@@ -1,8 +1,8 @@
encoder: conformer
encoder_conf:
- output_size: 512
- attention_heads: 8
- linear_units: 2048
+ output_size: 256
+ attention_heads: 4
+ linear_units: 1024
num_blocks: 12
dropout_rate: 0.1
positional_dropout_rate: 0.1
@@ -19,7 +19,7 @@
decoder: transformer
decoder_conf:
- attention_heads: 8
+ attention_heads: 4
linear_units: 2048
num_blocks: 6
dropout_rate: 0.1
@@ -27,13 +27,25 @@
self_attention_dropout_rate: 0.1
src_attention_dropout_rate: 0.1
+# frontend related
+frontend: wav_frontend
+frontend_conf:
+ fs: 16000
+ window: hamming
+ n_mels: 80
+ frame_length: 25
+ frame_shift: 10
+ lfr_m: 1
+ lfr_n: 1
+
+# hybrid CTC/attention
model_conf:
ctc_weight: 0.3
lsm_weight: 0.1
length_normalized_loss: false
-accum_grad: 2
-max_epoch: 50
+accum_grad: 1
+max_epoch: 210
patience: none
init: none
best_model_criterion:
@@ -44,11 +56,11 @@
optim: adam
optim_conf:
- lr: 0.0025
+ lr: 0.002
weight_decay: 0.000001
scheduler: warmuplr
scheduler_conf:
- warmup_steps: 40000
+ warmup_steps: 15000
specaug: specaug
specaug_conf:
@@ -64,7 +76,7 @@
time_mask_width_ratio_range:
- 0.
- 0.05
- num_time_mask: 10
+ num_time_mask: 5
dataset_conf:
shuffle: True
@@ -77,4 +89,4 @@
num_workers: 8
log_interval: 50
-normalize: utterance_mvn
\ No newline at end of file
+normalize: None
\ No newline at end of file
--
Gitblit v1.9.1