Skip to content

Persephone asr refactor #131

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged
merged 18 commits into from
Mar 22, 2023
Merged

Persephone asr refactor #131

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
18 commits
Select commit Hold shift + click to select a range
13dd879
update configs for w2v
jesus-villalba Dec 27, 2022
4ba13a7
isort data
jesus-villalba Dec 27, 2022
dd7b3ed
fixed filter_func_args
jesus-villalba Dec 27, 2022
b274895
trainers use dictionaries
jesus-villalba Dec 28, 2022
4f7bc3d
fixed amp
jesus-villalba Dec 28, 2022
5a138e8
update feature, embeddings datasets and samplers, added fwse and cfws…
jesus-villalba Dec 29, 2022
15db427
added scripts to extract xvectors from feats back to bin
jesus-villalba Dec 29, 2022
f9e4a24
isorted everything
jesus-villalba Dec 29, 2022
e26f5b9
new hyperparams for voxceleb
jesus-villalba Jan 26, 2023
5f0ac99
updated qmf script
jesus-villalba Jan 27, 2023
b364798
xxx
jesus-villalba Jan 27, 2023
06adb57
solved conflicts
jesus-villalba Jan 27, 2023
76ac6f3
audio dataset with dictionary for asr
jesus-villalba Feb 9, 2023
aabbef1
start refactorizing rnn-t
jesus-villalba Feb 10, 2023
1d78ea3
refactored beam search
jesus-villalba Mar 10, 2023
97050f9
working align length sycn dec
jesus-villalba Mar 10, 2023
db14742
finised refactoring rnn transducer
jesus-villalba Mar 22, 2023
16b1eae
Merge branch 'persephone-asr' into persephone-asr-refactor
jesus-villalba Mar 22, 2023
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
The table of contents is too big for display.
Diff view
Diff view
  •  
  •  
  •  
4 changes: 4 additions & 0 deletions egs/librispeech/v1/conf/infer.yaml
View file
Open in desktop
Original file line number Diff line number Diff line change
@@ -0,0 +1,4 @@
beam_width: 5
decoding_method: time_sync_beam_search
#decoding_method: greedy
#decoding_method: align_length_sync_beam_search
69 changes: 69 additions & 0 deletions egs/librispeech/v1/conf/train_wav2vec2base_rnnt_k2_pruned_stage1_v1.0.yaml
View file
Open in desktop
Original file line number Diff line number Diff line change
@@ -0,0 +1,69 @@
data:
train:
dataset:
wav_scale: 1
aug_cfgs:
- conf/reverb_noise_aug.yaml
return_segment_info:
- text
sampler:
sampler_type: bucketing_seg_sampler
max_batch_length: 70.
min_batch_size: 1
drop_last: false
data_loader:
num_workers: 4
val:
dataset:
aug_cfgs:
- conf/reverb_noise_aug.yaml
wav_scale: 1
return_segment_info:
- text
sampler:
sampler_type: bucketing_seg_sampler
max_batch_length: 70.
min_batch_size: 1
drop_last: true
data_loader:
num_workers: 4
model:
hf_feats:
pretrained_model_path: facebook/wav2vec2-base-960h
transducer:
decoder:
rnnt_loss: k2_pruned
predictor:
embed_dim: 1024
num_layers: 2
hid_feats: 512
embed_dropout_rate: 0.4
rnn_dropout_rate: 0.4
rnn_type: lstm
joiner:
hid_feats: 512
feat_fusion_method: weighted-avg
feat_fusion_start: 2
trainer:
optim:
opt_type: sgd
lr: 0.003
momentum: 0.9
weight_decay: 4e-4
lrsched:
lrsch_type: exp_lr
decay_rate: 0.5
decay_steps: 4200
hold_steps: 1500
min_lr: 4e-5
warmup_steps: 1500
update_lr_on_opt_step: true
grad_clip: 100
use_amp: true
log_interval: 1000
epochs: 120
# eff_batch_size: 1024
eff_batch_size: 128
train_mode: hf-feats-frozen-nograd


69 changes: 69 additions & 0 deletions egs/librispeech/v1/conf/train_wav2vec2base_rnnt_k2_pruned_stage1_v1.2.yaml
View file
Open in desktop
Original file line number Diff line number Diff line change
@@ -0,0 +1,69 @@
data:
train:
dataset:
wav_scale: 1
aug_cfgs:
- conf/reverb_noise_aug.yaml
return_segment_info:
- text
sampler:
sampler_type: bucketing_seg_sampler
max_batch_length: 70.
min_batch_size: 1
drop_last: false
data_loader:
num_workers: 4
val:
dataset:
aug_cfgs:
- conf/reverb_noise_aug.yaml
wav_scale: 1
return_segment_info:
- text
sampler:
sampler_type: bucketing_seg_sampler
max_batch_length: 70.
min_batch_size: 1
drop_last: true
data_loader:
num_workers: 4
model:
hf_feats:
pretrained_model_path: facebook/wav2vec2-base-960h
transducer:
decoder:
rnnt_loss: k2_pruned
predictor:
embed_dim: 1024
num_layers: 2
hid_feats: 512
embed_dropout_rate: 0.4
rnn_dropout_rate: 0.4
rnn_type: lstm
joiner:
hid_feats: 512
feat_fusion_method: weighted-avg
feat_fusion_start: 2
trainer:
optim:
opt_type: sgd
lr: 0.005
momentum: 0.9
weight_decay: 4e-4
lrsched:
lrsch_type: exp_lr
decay_rate: 0.5
decay_steps: 4200
hold_steps: 1500
min_lr: 4e-5
warmup_steps: 1500
update_lr_on_opt_step: true
grad_clip: 100
use_amp: true
log_interval: 1000
epochs: 120
# eff_batch_size: 1024
eff_batch_size: 128
train_mode: hf-feats-frozen-nograd


70 changes: 70 additions & 0 deletions egs/librispeech/v1/conf/train_wav2vec2base_rnnt_k2_pruned_stage1_v1.3.yaml
View file
Open in desktop
Original file line number Diff line number Diff line change
@@ -0,0 +1,70 @@
data:
train:
dataset:
wav_scale: 1
aug_cfgs:
- conf/reverb_noise_aug.yaml
return_segment_info:
- text
sampler:
sampler_type: bucketing_seg_sampler
max_batch_length: 70.
min_batch_size: 1
drop_last: false
data_loader:
num_workers: 4
val:
dataset:
aug_cfgs:
- conf/reverb_noise_aug.yaml
wav_scale: 1
return_segment_info:
- text
sampler:
sampler_type: bucketing_seg_sampler
max_batch_length: 70.
min_batch_size: 1
drop_last: true
data_loader:
num_workers: 4
model:
hf_feats:
pretrained_model_path: facebook/wav2vec2-base-960h
transducer:
decoder:
rnnt_loss: k2_pruned
simple_loss_scale: 0.2
predictor:
embed_dim: 1024
num_layers: 2
hid_feats: 512
embed_dropout_rate: 0.4
rnn_dropout_rate: 0.4
rnn_type: lstm
joiner:
hid_feats: 512
feat_fusion_method: weighted-avg
feat_fusion_start: 2
trainer:
optim:
opt_type: sgd
lr: 0.005
momentum: 0.9
weight_decay: 4e-4
lrsched:
lrsch_type: exp_lr
decay_rate: 0.5
decay_steps: 4200
hold_steps: 1500
min_lr: 4e-5
warmup_steps: 1500
update_lr_on_opt_step: true
grad_clip: 100
use_amp: true
log_interval: 1000
epochs: 120
# eff_batch_size: 1024
eff_batch_size: 128
train_mode: hf-feats-frozen-nograd


69 changes: 69 additions & 0 deletions egs/librispeech/v1/conf/train_wav2vec2base_rnnt_k2_stage1_v1.0.yaml
View file
Open in desktop
Original file line number Diff line number Diff line change
@@ -0,0 +1,69 @@
data:
train:
dataset:
wav_scale: 1
aug_cfgs:
- conf/reverb_noise_aug.yaml
return_segment_info:
- text
sampler:
sampler_type: bucketing_seg_sampler
max_batch_length: 70.
min_batch_size: 1
drop_last: false
data_loader:
num_workers: 4
val:
dataset:
aug_cfgs:
- conf/reverb_noise_aug.yaml
wav_scale: 1
return_segment_info:
- text
sampler:
sampler_type: bucketing_seg_sampler
max_batch_length: 70.
min_batch_size: 1
drop_last: true
data_loader:
num_workers: 4
model:
hf_feats:
pretrained_model_path: facebook/wav2vec2-base-960h
transducer:
decoder:
rnnt_loss: k2
predictor:
embed_dim: 1024
num_layers: 2
hid_feats: 512
embed_dropout_rate: 0.4
rnn_dropout_rate: 0.4
rnn_type: lstm
joiner:
hid_feats: 512
feat_fusion_method: weighted-avg
feat_fusion_start: 2
trainer:
optim:
opt_type: sgd
lr: 0.003
momentum: 0.9
weight_decay: 4e-4
lrsched:
lrsch_type: exp_lr
decay_rate: 0.5
decay_steps: 4200
hold_steps: 1500
min_lr: 4e-5
warmup_steps: 1500
update_lr_on_opt_step: true
grad_clip: 100
use_amp: true
log_interval: 1000
epochs: 120
# eff_batch_size: 1024
eff_batch_size: 128
train_mode: hf-feats-frozen-nograd


53 changes: 53 additions & 0 deletions egs/librispeech/v1/conf/train_wav2vec2base_transducer_stage1_v5.0.yaml
View file
Open in desktop
Original file line number Diff line number Diff line change
@@ -0,0 +1,53 @@
data:
train:
dataset:
wav_scale: 1
aug_cfgs:
- conf/reverb_noise_aug.yaml
return_segment_info:
- text
sampler:
sampler_type: 'bucketing_seg_sampler'
max_batch_length: 75.
min_batch_size: 1
drop_last: false
data_loader:
num_workers: 4
val:
dataset:
aug_cfgs:
- conf/reverb_noise_aug.yaml
wav_scale: 1
return_segment_info:
- text
sampler:
sampler_type: 'bucketing_seg_sampler'
max_batch_length: 75.
min_batch_size: 1
drop_last: true
data_loader:
num_workers: 4
model: wav2vec2base_transducer_do0.4.yaml
trainer:
optim:
opt_type: sgd
lr: 0.003
momentum: 0.9
weight_decay: 4e-4
lrsched:
lrsch_type: exp_lr
decay_rate: 0.5
decay_steps: 42000
hold_steps: 15000
min_lr: 4e-5
warmup_steps: 15000
update_lr_on_opt_step: true
grad_clip: 100
use_amp: true
log_interval: 1000
epochs: 1200
# eff_batch_size: 1024
eff_batch_size: 128
train_mode: hf-feats-frozen-nograd


Loading