msb-public
/
PaddleSpeech
mirror of https://github.com/PaddlePaddle/PaddleSpeech
1
0
Code Issues Projects Releases Wiki Activity

add the whole of aishell asr1

Browse Source
pull/1225/head
huangyuxin 3 years ago
parent c40b6f4062
commit 960658f669
6 changed files with 177 additions and 192 deletions
Show Stats Download Patch File Download Diff File

183
examples/aishell/asr1/conf/chunk_conformer.yaml
Unescape View File

@ -1,103 +1,94 @@
# network architecture ############################################
model: # Network Architecture #
cmvn_file: ############################################
cmvn_file_type: "json" cmvn_file:
# encoder related cmvn_file_type: "json"
encoder: conformer # encoder related
encoder_conf: encoder: conformer
output_size: 256 # dimension of attention encoder_conf:
attention_heads: 4 output_size: 256 # dimension of attention
linear_units: 2048 # the number of units of position-wise feed forward attention_heads: 4
num_blocks: 12 # the number of encoder blocks linear_units: 2048 # the number of units of position-wise feed forward
dropout_rate: 0.1 num_blocks: 12 # the number of encoder blocks
positional_dropout_rate: 0.1 dropout_rate: 0.1
attention_dropout_rate: 0.0 positional_dropout_rate: 0.1
input_layer: conv2d # encoder input type, you can chose conv2d, conv2d6 and conv2d8 attention_dropout_rate: 0.0
normalize_before: True input_layer: conv2d # encoder input type, you can chose conv2d, conv2d6 and conv2d8
cnn_module_kernel: 15 normalize_before: True
use_cnn_module: True cnn_module_kernel: 15
activation_type: 'swish' use_cnn_module: True
pos_enc_layer_type: 'rel_pos' activation_type: 'swish'
selfattention_layer_type: 'rel_selfattn' pos_enc_layer_type: 'rel_pos'
causal: true selfattention_layer_type: 'rel_selfattn'
use_dynamic_chunk: true causal: true
cnn_module_norm: 'layer_norm' # using nn.LayerNorm makes model converge faster use_dynamic_chunk: true
use_dynamic_left_chunk: false cnn_module_norm: 'layer_norm' # using nn.LayerNorm makes model converge faster
use_dynamic_left_chunk: false
# decoder related
decoder: transformer
decoder_conf:
attention_heads: 4
linear_units: 2048
num_blocks: 6
dropout_rate: 0.1
positional_dropout_rate: 0.1
self_attention_dropout_rate: 0.0
src_attention_dropout_rate: 0.0
# hybrid CTC/attention
model_conf:
ctc_weight: 0.3
lsm_weight: 0.1 # label smoothing option
length_normalized_loss: false
# decoder related ###########################################
decoder: transformer # Data #
decoder_conf: ###########################################
attention_heads: 4
linear_units: 2048
num_blocks: 6
dropout_rate: 0.1
positional_dropout_rate: 0.1
self_attention_dropout_rate: 0.0
src_attention_dropout_rate: 0.0
# hybrid CTC/attention train_manifest: data/manifest.train
model_conf: dev_manifest: data/manifest.dev
ctc_weight: 0.3 test_manifest: data/manifest.test
lsm_weight: 0.1 # label smoothing option
length_normalized_loss: false
data: ###########################################
train_manifest: data/manifest.train # Dataloader #
dev_manifest: data/manifest.dev ###########################################
test_manifest: data/manifest.test
vocab_filepath: data/lang_char/vocab.txt
unit_type: 'char'
augmentation_config: conf/preprocess.yaml
feat_dim: 80
stride_ms: 10.0
window_ms: 25.0
sortagrad: 0 # Feed samples from shortest to longest ; -1: enabled for all epochs, 0: disabled, other: enabled for 'other' epochs
batch_size: 64
maxlen_in: 512 # if input length > maxlen-in, batchsize is automatically reduced
maxlen_out: 150 # if output length > maxlen-out, batchsize is automatically reduced
minibatches: 0 # for debug
batch_count: auto
batch_bins: 0
batch_frames_in: 0
batch_frames_out: 0
batch_frames_inout: 0
num_workers: 0
subsampling_factor: 1
num_encs: 1
collator: ###########################################
vocab_filepath: data/lang_char/vocab.txt # training #
unit_type: 'char' ###########################################
augmentation_config: conf/preprocess.yaml n_epoch: 240
feat_dim: 80 accum_grad: 2
stride_ms: 10.0 global_grad_clip: 5.0
window_ms: 25.0 optim: adam
sortagrad: 0 # Feed samples from shortest to longest ; -1: enabled for all epochs, 0: disabled, other: enabled for 'other' epochs optim_conf:
batch_size: 64 lr: 0.002
maxlen_in: 512 # if input length > maxlen-in, batchsize is automatically reduced weight_decay: 1e-6
maxlen_out: 150 # if output length > maxlen-out, batchsize is automatically reduced scheduler: warmuplr
minibatches: 0 # for debug scheduler_conf:
batch_count: auto warmup_steps: 25000
batch_bins: 0 lr_decay: 1.0
batch_frames_in: 0 log_interval: 100
batch_frames_out: 0 checkpoint:
batch_frames_inout: 0 kbest_n: 50
num_workers: 0 latest_n: 5
subsampling_factor: 1
num_encs: 1
training:
n_epoch: 240
accum_grad: 2
global_grad_clip: 5.0
optim: adam
optim_conf:
lr: 0.002
weight_decay: 1e-6
scheduler: warmuplr
scheduler_conf:
warmup_steps: 25000
lr_decay: 1.0
log_interval: 100
checkpoint:
kbest_n: 50
latest_n: 5
decoding:
beam_size: 10
batch_size: 128
error_rate_type: cer
decoding_method: attention # 'attention', 'ctc_greedy_search', 'ctc_prefix_beam_search', 'attention_rescoring'
ctc_weight: 0.5 # ctc weight for attention rescoring decode mode.
decoding_chunk_size: -1 # decoding chunk size. Defaults to -1.
# <0: for decoding, use full chunk.
# >0: for decoding, use fixed chunk size as set.
# 0: used for training, it's prohibited here.
num_decoding_left_chunks: -1 # number of left chunks for decoding. Defaults to -1.
simulate_streaming: False # simulate streaming inference. Defaults to False.

5
examples/aishell/asr1/conf/conformer.yaml
Unescape View File

@ -1,7 +1,6 @@
############################################ ############################################
# Network Architecture # # Network Architecture #
############################################ ############################################
#model:
cmvn_file: cmvn_file:
cmvn_file_type: "json" cmvn_file_type: "json"
# encoder related # encoder related
@ -42,7 +41,6 @@ model_conf:
########################################### ###########################################
# Data # # Data #
########################################### ###########################################
#data:
train_manifest: data/manifest.train train_manifest: data/manifest.train
dev_manifest: data/manifest.dev dev_manifest: data/manifest.dev
test_manifest: data/manifest.test test_manifest: data/manifest.test
@ -50,11 +48,9 @@ test_manifest: data/manifest.test
########################################### ###########################################
# Dataloader # # Dataloader #
########################################### ###########################################
#collator:
vocab_filepath: data/lang_char/vocab.txt vocab_filepath: data/lang_char/vocab.txt
unit_type: 'char' unit_type: 'char'
augmentation_config: conf/preprocess.yaml augmentation_config: conf/preprocess.yaml
spm_model_prefix: ''
feat_dim: 80 feat_dim: 80
stride_ms: 10.0 stride_ms: 10.0
window_ms: 25.0 window_ms: 25.0
@ -75,7 +71,6 @@ num_encs: 1
########################################### ###########################################
# training # # training #
########################################### ###########################################
#training:
n_epoch: 240 n_epoch: 240
accum_grad: 2 accum_grad: 2
global_grad_clip: 5.0 global_grad_clip: 5.0

167
examples/aishell/asr1/conf/transformer.yaml
Unescape View File

@ -1,95 +1,84 @@
# network architecture ############################################
model: # Network Architecture #
cmvn_file: ############################################
cmvn_file_type: "json" cmvn_file:
# encoder related cmvn_file_type: "json"
encoder: transformer # encoder related
encoder_conf: encoder: transformer
output_size: 256 # dimension of attention encoder_conf:
attention_heads: 4 output_size: 256 # dimension of attention
linear_units: 2048 # the number of units of position-wise feed forward attention_heads: 4
num_blocks: 12 # the number of encoder blocks linear_units: 2048 # the number of units of position-wise feed forward
dropout_rate: 0.1 num_blocks: 12 # the number of encoder blocks
positional_dropout_rate: 0.1 dropout_rate: 0.1
attention_dropout_rate: 0.0 positional_dropout_rate: 0.1
input_layer: conv2d # encoder input type, you can chose conv2d, conv2d6 and conv2d8 attention_dropout_rate: 0.0
normalize_before: true input_layer: conv2d # encoder input type, you can chose conv2d, conv2d6 and conv2d8
normalize_before: true
# decoder related # decoder related
decoder: transformer decoder: transformer
decoder_conf: decoder_conf:
attention_heads: 4 attention_heads: 4
linear_units: 2048 linear_units: 2048
num_blocks: 6 num_blocks: 6
dropout_rate: 0.1 dropout_rate: 0.1
positional_dropout_rate: 0.1 positional_dropout_rate: 0.1
self_attention_dropout_rate: 0.0 self_attention_dropout_rate: 0.0
src_attention_dropout_rate: 0.0 src_attention_dropout_rate: 0.0
# hybrid CTC/attention
model_conf:
ctc_weight: 0.3
lsm_weight: 0.1 # label smoothing option
length_normalized_loss: false
# hybrid CTC/attention
model_conf:
ctc_weight: 0.3
lsm_weight: 0.1 # label smoothing option
length_normalized_loss: false
###########################################
# Data #
###########################################
# https://yaml.org/type/float.html # https://yaml.org/type/float.html
data: train_manifest: data/manifest.train
train_manifest: data/manifest.train dev_manifest: data/manifest.dev
dev_manifest: data/manifest.dev test_manifest: data/manifest.test
test_manifest: data/manifest.test
collator:
unit_type: 'char'
vocab_filepath: data/lang_char/vocab.txt
feat_dim: 80
stride_ms: 10.0
window_ms: 25.0
sortagrad: 0 # Feed samples from shortest to longest ; -1: enabled for all epochs, 0: disabled, other: enabled for 'other' epochs
batch_size: 64
maxlen_in: 512 # if input length > maxlen-in, batchsize is automatically reduced
maxlen_out: 150 # if output length > maxlen-out, batchsize is automatically reduced
minibatches: 0 # for debug
batch_count: auto
batch_bins: 0
batch_frames_in: 0
batch_frames_out: 0
batch_frames_inout: 0
augmentation_config: conf/preprocess.yaml
num_workers: 0
subsampling_factor: 1
num_encs: 1
training:
n_epoch: 240
accum_grad: 2
global_grad_clip: 5.0
optim: adam
optim_conf:
lr: 0.002
weight_decay: 1e-6
scheduler: warmuplr
scheduler_conf:
warmup_steps: 25000
lr_decay: 1.0
log_interval: 100
checkpoint:
kbest_n: 50
latest_n: 5
###########################################
# Dataloader #
###########################################
unit_type: 'char'
vocab_filepath: data/lang_char/vocab.txt
feat_dim: 80
stride_ms: 10.0
window_ms: 25.0
sortagrad: 0 # Feed samples from shortest to longest ; -1: enabled for all epochs, 0: disabled, other: enabled for 'other' epochs
batch_size: 64
maxlen_in: 512 # if input length > maxlen-in, batchsize is automatically reduced
maxlen_out: 150 # if output length > maxlen-out, batchsize is automatically reduced
minibatches: 0 # for debug
batch_count: auto
batch_bins: 0
batch_frames_in: 0
batch_frames_out: 0
batch_frames_inout: 0
augmentation_config: conf/preprocess.yaml
num_workers: 0
subsampling_factor: 1
num_encs: 1
decoding: ###########################################
beam_size: 10 # training #
batch_size: 128 ###########################################
error_rate_type: cer n_epoch: 240
decoding_method: attention # 'attention', 'ctc_greedy_search', 'ctc_prefix_beam_search', 'attention_rescoring' accum_grad: 2
ctc_weight: 0.5 # ctc weight for attention rescoring decode mode. global_grad_clip: 5.0
decoding_chunk_size: -1 # decoding chunk size. Defaults to -1. optim: adam
# <0: for decoding, use full chunk. optim_conf:
# >0: for decoding, use fixed chunk size as set. lr: 0.002
# 0: used for training, it's prohibited here. weight_decay: 1e-6
num_decoding_left_chunks: -1 # number of left chunks for decoding. Defaults to -1. scheduler: warmuplr
simulate_streaming: False # simulate streaming inference. Defaults to False. scheduler_conf:
warmup_steps: 25000
lr_decay: 1.0
log_interval: 100
checkpoint:
kbest_n: 50
latest_n: 5

11
examples/aishell/asr1/conf/tuning/chunk_decode.yaml
Unescape View File

@ -0,0 +1,11 @@
beam_size: 10
decode_batch_size: 128
error_rate_type: cer
decoding_method: attention # 'attention', 'ctc_greedy_search', 'ctc_prefix_beam_search', 'attention_rescoring'
ctc_weight: 0.5 # ctc weight for attention rescoring decode mode.
decoding_chunk_size: -1 # decoding chunk size. Defaults to -1.
# <0: for decoding, use full chunk.
# >0: for decoding, use fixed chunk size as set.
# 0: used for training, it's prohibited here.
num_decoding_left_chunks: -1 # number of left chunks for decoding. Defaults to -1.
simulate_streaming: False # simulate streaming inference. Defaults to False.

1
examples/aishell/asr1/conf/decode.yaml → examples/aishell/asr1/conf/tuning/decode.yaml
Unescape View File

@ -1,4 +1,3 @@
#decoding:
beam_size: 10 beam_size: 10
decode_batch_size: 128 decode_batch_size: 128
error_rate_type: cer error_rate_type: cer

2
examples/aishell/asr1/run.sh
Unescape View File

@ -6,7 +6,7 @@ gpus=0,1,2,3
stage=0 stage=0
stop_stage=50 stop_stage=50
conf_path=conf/conformer.yaml conf_path=conf/conformer.yaml
decode_conf_path=conf/decode.yaml decode_conf_path=conf/tuning/decode.yaml
avg_num=20 avg_num=20
audio_file=data/demo_01_03.wav audio_file=data/demo_01_03.wav

Write Preview
Loading…
Cancel
Save