wyz97
commited on
Commit
•
db03c34
1
Parent(s):
e3a60b5
Update model
Browse files- README.md +336 -0
- exp/enh_train_raw/99epoch.pth +3 -0
- exp/enh_train_raw/RESULTS.md +20 -0
- exp/enh_train_raw/config.yaml +234 -0
- exp/enh_train_raw/images/L1_on_Spec.png +0 -0
- exp/enh_train_raw/images/MSE_on_Spec.png +0 -0
- exp/enh_train_raw/images/TD_L1_loss.png +0 -0
- exp/enh_train_raw/images/TD_MSE_loss.png +0 -0
- exp/enh_train_raw/images/backward_time.png +0 -0
- exp/enh_train_raw/images/clip.png +0 -0
- exp/enh_train_raw/images/forward_time.png +0 -0
- exp/enh_train_raw/images/gpu_max_cached_mem_GB.png +0 -0
- exp/enh_train_raw/images/grad_norm.png +0 -0
- exp/enh_train_raw/images/iter_time.png +0 -0
- exp/enh_train_raw/images/loss.png +0 -0
- exp/enh_train_raw/images/loss_scale.png +0 -0
- exp/enh_train_raw/images/optim0_lr0.png +0 -0
- exp/enh_train_raw/images/optim_step_time.png +0 -0
- exp/enh_train_raw/images/snr_loss.png +0 -0
- exp/enh_train_raw/images/train_time.png +0 -0
- meta.yaml +8 -0
README.md
CHANGED
@@ -1,3 +1,339 @@
|
|
1 |
---
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2 |
license: cc-by-4.0
|
3 |
---
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
---
|
2 |
+
tags:
|
3 |
+
- espnet
|
4 |
+
- audio
|
5 |
+
- audio-to-audio
|
6 |
+
language: en
|
7 |
+
datasets:
|
8 |
+
- librimix
|
9 |
license: cc-by-4.0
|
10 |
---
|
11 |
+
|
12 |
+
## ESPnet2 ENH model
|
13 |
+
|
14 |
+
### `espnet/Wangyou_Zhang_librimix_train_enh_tse_td_speakerbeam_raw`
|
15 |
+
|
16 |
+
This model was trained by Wangyou Zhang using librimix recipe in [espnet](https://github.com/espnet/espnet/).
|
17 |
+
|
18 |
+
### Demo: How to use in ESPnet2
|
19 |
+
|
20 |
+
Follow the [ESPnet installation instructions](https://espnet.github.io/espnet/installation.html)
|
21 |
+
if you haven't done that already.
|
22 |
+
|
23 |
+
```bash
|
24 |
+
cd espnet
|
25 |
+
|
26 |
+
pip install -e .
|
27 |
+
cd egs2/librimix/tse1
|
28 |
+
./run.sh --skip_data_prep false --skip_train true --download_model espnet/Wangyou_Zhang_librimix_train_enh_tse_td_speakerbeam_raw
|
29 |
+
```
|
30 |
+
|
31 |
+
<!-- Generated by ./scripts/utils/show_enh_score.sh -->
|
32 |
+
# RESULTS
|
33 |
+
## Environments
|
34 |
+
- date: `Mon Jun 5 22:42:07 CST 2023`
|
35 |
+
- python version: `3.8.16 (default, Mar 2 2023, 03:21:46) [GCC 11.2.0]`
|
36 |
+
- espnet version: `espnet 202301`
|
37 |
+
- pytorch version: `pytorch 2.0.1`
|
38 |
+
- Git hash: ``
|
39 |
+
- Commit date: ``
|
40 |
+
|
41 |
+
|
42 |
+
## enh_train_raw
|
43 |
+
|
44 |
+
config: ./conf/train.yaml
|
45 |
+
|
46 |
+
|dataset|PESQ_WB|STOI|SAR|SDR|SIR|SI_SNR|OVRL|SIG|BAK|P808_MOS|
|
47 |
+
|---|---|---|---|---|---|---|---|---|---|---|
|
48 |
+
|enhanced_dev|1.73|86.50|12.50|11.40|24.83|10.58|2.95|3.24|3.92|3.23|
|
49 |
+
|enhanced_test|1.73|87.36|12.34|11.47|24.51|10.74|2.99|3.29|3.91|3.25|
|
50 |
+
|
51 |
+
## ENH config
|
52 |
+
|
53 |
+
<details><summary>expand</summary>
|
54 |
+
|
55 |
+
```
|
56 |
+
config: ./conf/train.yaml
|
57 |
+
print_config: false
|
58 |
+
log_level: INFO
|
59 |
+
dry_run: false
|
60 |
+
iterator_type: chunk
|
61 |
+
output_dir: exp/enh_train_raw
|
62 |
+
ngpu: 1
|
63 |
+
seed: 0
|
64 |
+
num_workers: 2
|
65 |
+
num_att_plot: 3
|
66 |
+
dist_backend: nccl
|
67 |
+
dist_init_method: env://
|
68 |
+
dist_world_size: 4
|
69 |
+
dist_rank: 0
|
70 |
+
local_rank: 0
|
71 |
+
dist_master_addr: localhost
|
72 |
+
dist_master_port: 43837
|
73 |
+
dist_launcher: null
|
74 |
+
multiprocessing_distributed: true
|
75 |
+
unused_parameters: true
|
76 |
+
sharded_ddp: false
|
77 |
+
cudnn_enabled: true
|
78 |
+
cudnn_benchmark: false
|
79 |
+
cudnn_deterministic: true
|
80 |
+
collect_stats: false
|
81 |
+
write_collected_feats: false
|
82 |
+
skip_stats_npz: false
|
83 |
+
max_epoch: 100
|
84 |
+
patience: 20
|
85 |
+
val_scheduler_criterion:
|
86 |
+
- valid
|
87 |
+
- loss
|
88 |
+
early_stopping_criterion:
|
89 |
+
- valid
|
90 |
+
- loss
|
91 |
+
- min
|
92 |
+
best_model_criterion:
|
93 |
+
- - valid
|
94 |
+
- snr
|
95 |
+
- max
|
96 |
+
- - valid
|
97 |
+
- loss
|
98 |
+
- min
|
99 |
+
keep_nbest_models: 1
|
100 |
+
nbest_averaging_interval: 0
|
101 |
+
grad_clip: 5.0
|
102 |
+
grad_clip_type: 2.0
|
103 |
+
grad_noise: false
|
104 |
+
accum_grad: 1
|
105 |
+
no_forward_run: false
|
106 |
+
resume: true
|
107 |
+
train_dtype: float32
|
108 |
+
use_amp: false
|
109 |
+
log_interval: null
|
110 |
+
use_matplotlib: true
|
111 |
+
use_tensorboard: true
|
112 |
+
create_graph_in_tensorboard: false
|
113 |
+
use_wandb: false
|
114 |
+
wandb_project: null
|
115 |
+
wandb_id: null
|
116 |
+
wandb_entity: null
|
117 |
+
wandb_name: null
|
118 |
+
wandb_model_log_interval: -1
|
119 |
+
detect_anomaly: false
|
120 |
+
pretrain_path: null
|
121 |
+
init_param: []
|
122 |
+
ignore_init_mismatch: false
|
123 |
+
freeze_param: []
|
124 |
+
num_iters_per_epoch: null
|
125 |
+
batch_size: 16
|
126 |
+
valid_batch_size: null
|
127 |
+
batch_bins: 1000000
|
128 |
+
valid_batch_bins: null
|
129 |
+
train_shape_file:
|
130 |
+
- exp/enh_stats_train_dev_16k/train/speech_mix_shape
|
131 |
+
- exp/enh_stats_train_dev_16k/train/speech_ref1_shape
|
132 |
+
- exp/enh_stats_train_dev_16k/train/enroll_ref1_shape
|
133 |
+
- exp/enh_stats_train_dev_16k/train/speech_ref2_shape
|
134 |
+
- exp/enh_stats_train_dev_16k/train/enroll_ref2_shape
|
135 |
+
valid_shape_file:
|
136 |
+
- exp/enh_stats_train_dev_16k/valid/speech_mix_shape
|
137 |
+
- exp/enh_stats_train_dev_16k/valid/speech_ref1_shape
|
138 |
+
- exp/enh_stats_train_dev_16k/valid/enroll_ref1_shape
|
139 |
+
- exp/enh_stats_train_dev_16k/valid/speech_ref2_shape
|
140 |
+
- exp/enh_stats_train_dev_16k/valid/enroll_ref2_shape
|
141 |
+
batch_type: folded
|
142 |
+
valid_batch_type: null
|
143 |
+
fold_length:
|
144 |
+
- 80000
|
145 |
+
- 80000
|
146 |
+
- 80000
|
147 |
+
- 80000
|
148 |
+
- 80000
|
149 |
+
sort_in_batch: descending
|
150 |
+
sort_batch: descending
|
151 |
+
multiple_iterator: false
|
152 |
+
chunk_length: 48000
|
153 |
+
chunk_shift_ratio: 0.5
|
154 |
+
num_cache_chunks: 1024
|
155 |
+
chunk_excluded_key_prefixes:
|
156 |
+
- enroll_ref
|
157 |
+
train_data_path_and_name_and_type:
|
158 |
+
- - dump/raw/train/wav.scp
|
159 |
+
- speech_mix
|
160 |
+
- sound
|
161 |
+
- - dump/raw/train/spk1.scp
|
162 |
+
- speech_ref1
|
163 |
+
- sound
|
164 |
+
- - dump/raw/train/enroll_spk1.scp
|
165 |
+
- enroll_ref1
|
166 |
+
- text
|
167 |
+
- - dump/raw/train/spk2.scp
|
168 |
+
- speech_ref2
|
169 |
+
- sound
|
170 |
+
- - dump/raw/train/enroll_spk2.scp
|
171 |
+
- enroll_ref2
|
172 |
+
- text
|
173 |
+
valid_data_path_and_name_and_type:
|
174 |
+
- - dump/raw/dev/wav.scp
|
175 |
+
- speech_mix
|
176 |
+
- sound
|
177 |
+
- - dump/raw/dev/spk1.scp
|
178 |
+
- speech_ref1
|
179 |
+
- sound
|
180 |
+
- - dump/raw/dev/enroll_spk1.scp
|
181 |
+
- enroll_ref1
|
182 |
+
- text
|
183 |
+
- - dump/raw/dev/spk2.scp
|
184 |
+
- speech_ref2
|
185 |
+
- sound
|
186 |
+
- - dump/raw/dev/enroll_spk2.scp
|
187 |
+
- enroll_ref2
|
188 |
+
- text
|
189 |
+
allow_variable_data_keys: false
|
190 |
+
max_cache_size: 0.0
|
191 |
+
max_cache_fd: 32
|
192 |
+
valid_max_cache_size: null
|
193 |
+
exclude_weight_decay: false
|
194 |
+
exclude_weight_decay_conf: {}
|
195 |
+
optim: adam
|
196 |
+
optim_conf:
|
197 |
+
lr: 0.001
|
198 |
+
eps: 1.0e-08
|
199 |
+
weight_decay: 0
|
200 |
+
scheduler: reducelronplateau
|
201 |
+
scheduler_conf:
|
202 |
+
mode: min
|
203 |
+
factor: 0.7
|
204 |
+
patience: 3
|
205 |
+
init: null
|
206 |
+
model_conf:
|
207 |
+
num_spk: 2
|
208 |
+
share_encoder: true
|
209 |
+
criterions:
|
210 |
+
- name: snr
|
211 |
+
conf:
|
212 |
+
eps: 1.0e-07
|
213 |
+
wrapper: fixed_order
|
214 |
+
wrapper_conf:
|
215 |
+
weight: 1.0
|
216 |
+
- name: l1_fd
|
217 |
+
conf:
|
218 |
+
only_for_test: true
|
219 |
+
wrapper: fixed_order
|
220 |
+
wrapper_conf:
|
221 |
+
weight: 0.0
|
222 |
+
- name: l1_td
|
223 |
+
conf:
|
224 |
+
only_for_test: true
|
225 |
+
wrapper: fixed_order
|
226 |
+
wrapper_conf:
|
227 |
+
weight: 0.0
|
228 |
+
- name: mse_fd
|
229 |
+
conf:
|
230 |
+
only_for_test: true
|
231 |
+
wrapper: fixed_order
|
232 |
+
wrapper_conf:
|
233 |
+
weight: 0.0
|
234 |
+
- name: mse_td
|
235 |
+
conf:
|
236 |
+
only_for_test: true
|
237 |
+
wrapper: fixed_order
|
238 |
+
wrapper_conf:
|
239 |
+
weight: 0.0
|
240 |
+
train_spk2enroll: data/train-100/spk2enroll.json
|
241 |
+
enroll_segment: 48000
|
242 |
+
load_spk_embedding: false
|
243 |
+
load_all_speakers: false
|
244 |
+
rir_scp: null
|
245 |
+
rir_apply_prob: 1.0
|
246 |
+
noise_scp: null
|
247 |
+
noise_apply_prob: 1.0
|
248 |
+
noise_db_range: '13_15'
|
249 |
+
short_noise_thres: 0.5
|
250 |
+
speech_volume_normalize: null
|
251 |
+
use_reverberant_ref: false
|
252 |
+
num_spk: 1
|
253 |
+
num_noise_type: 1
|
254 |
+
sample_rate: 8000
|
255 |
+
force_single_channel: false
|
256 |
+
channel_reordering: false
|
257 |
+
categories: []
|
258 |
+
encoder: conv
|
259 |
+
encoder_conf:
|
260 |
+
channel: 256
|
261 |
+
kernel_size: 32
|
262 |
+
stride: 16
|
263 |
+
extractor: td_speakerbeam
|
264 |
+
extractor_conf:
|
265 |
+
layer: 8
|
266 |
+
stack: 4
|
267 |
+
bottleneck_dim: 256
|
268 |
+
hidden_dim: 512
|
269 |
+
skip_dim: 256
|
270 |
+
kernel: 3
|
271 |
+
causal: false
|
272 |
+
norm_type: gLN
|
273 |
+
pre_nonlinear: prelu
|
274 |
+
nonlinear: relu
|
275 |
+
i_adapt_layer: 7
|
276 |
+
adapt_layer_type: mul
|
277 |
+
adapt_enroll_dim: 256
|
278 |
+
use_spk_emb: false
|
279 |
+
decoder: conv
|
280 |
+
decoder_conf:
|
281 |
+
channel: 256
|
282 |
+
kernel_size: 32
|
283 |
+
stride: 16
|
284 |
+
preprocessor: tse
|
285 |
+
preprocessor_conf: {}
|
286 |
+
required:
|
287 |
+
- output_dir
|
288 |
+
version: '202301'
|
289 |
+
distributed: true
|
290 |
+
```
|
291 |
+
|
292 |
+
</details>
|
293 |
+
|
294 |
+
|
295 |
+
|
296 |
+
### Citing ESPnet
|
297 |
+
|
298 |
+
```BibTex
|
299 |
+
@inproceedings{watanabe2018espnet,
|
300 |
+
author={Shinji Watanabe and Takaaki Hori and Shigeki Karita and Tomoki Hayashi and Jiro Nishitoba and Yuya Unno and Nelson Yalta and Jahn Heymann and Matthew Wiesner and Nanxin Chen and Adithya Renduchintala and Tsubasa Ochiai},
|
301 |
+
title={{ESPnet}: End-to-End Speech Processing Toolkit},
|
302 |
+
year={2018},
|
303 |
+
booktitle={Proceedings of Interspeech},
|
304 |
+
pages={2207--2211},
|
305 |
+
doi={10.21437/Interspeech.2018-1456},
|
306 |
+
url={http://dx.doi.org/10.21437/Interspeech.2018-1456}
|
307 |
+
}
|
308 |
+
|
309 |
+
|
310 |
+
@inproceedings{ESPnet-SE,
|
311 |
+
author = {Chenda Li and Jing Shi and Wangyou Zhang and Aswin Shanmugam Subramanian and Xuankai Chang and
|
312 |
+
Naoyuki Kamo and Moto Hira and Tomoki Hayashi and Christoph B{"{o}}ddeker and Zhuo Chen and Shinji Watanabe},
|
313 |
+
title = {ESPnet-SE: End-To-End Speech Enhancement and Separation Toolkit Designed for {ASR} Integration},
|
314 |
+
booktitle = {{IEEE} Spoken Language Technology Workshop, {SLT} 2021, Shenzhen, China, January 19-22, 2021},
|
315 |
+
pages = {785--792},
|
316 |
+
publisher = {{IEEE}},
|
317 |
+
year = {2021},
|
318 |
+
url = {https://doi.org/10.1109/SLT48900.2021.9383615},
|
319 |
+
doi = {10.1109/SLT48900.2021.9383615},
|
320 |
+
timestamp = {Mon, 12 Apr 2021 17:08:59 +0200},
|
321 |
+
biburl = {https://dblp.org/rec/conf/slt/Li0ZSCKHHBC021.bib},
|
322 |
+
bibsource = {dblp computer science bibliography, https://dblp.org}
|
323 |
+
}
|
324 |
+
|
325 |
+
|
326 |
+
```
|
327 |
+
|
328 |
+
or arXiv:
|
329 |
+
|
330 |
+
```bibtex
|
331 |
+
@misc{watanabe2018espnet,
|
332 |
+
title={ESPnet: End-to-End Speech Processing Toolkit},
|
333 |
+
author={Shinji Watanabe and Takaaki Hori and Shigeki Karita and Tomoki Hayashi and Jiro Nishitoba and Yuya Unno and Nelson Yalta and Jahn Heymann and Matthew Wiesner and Nanxin Chen and Adithya Renduchintala and Tsubasa Ochiai},
|
334 |
+
year={2018},
|
335 |
+
eprint={1804.00015},
|
336 |
+
archivePrefix={arXiv},
|
337 |
+
primaryClass={cs.CL}
|
338 |
+
}
|
339 |
+
```
|
exp/enh_train_raw/99epoch.pth
ADDED
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
1 |
+
version https://git-lfs.github.com/spec/v1
|
2 |
+
oid sha256:7e1d6831d0df187ae8b23ed4d6804d07d0ec314336266e99f7da41a2f8ad055f
|
3 |
+
size 65038857
|
exp/enh_train_raw/RESULTS.md
ADDED
@@ -0,0 +1,20 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
<!-- Generated by ./scripts/utils/show_enh_score.sh -->
|
2 |
+
# RESULTS
|
3 |
+
## Environments
|
4 |
+
- date: `Mon Jun 5 22:42:07 CST 2023`
|
5 |
+
- python version: `3.8.16 (default, Mar 2 2023, 03:21:46) [GCC 11.2.0]`
|
6 |
+
- espnet version: `espnet 202301`
|
7 |
+
- pytorch version: `pytorch 2.0.1`
|
8 |
+
- Git hash: ``
|
9 |
+
- Commit date: ``
|
10 |
+
|
11 |
+
|
12 |
+
## enh_train_raw
|
13 |
+
|
14 |
+
config: ./conf/train.yaml
|
15 |
+
|
16 |
+
|dataset|PESQ_WB|STOI|SAR|SDR|SIR|SI_SNR|OVRL|SIG|BAK|P808_MOS|
|
17 |
+
|---|---|---|---|---|---|---|---|---|---|---|
|
18 |
+
|enhanced_dev|1.73|86.50|12.50|11.40|24.83|10.58|2.95|3.24|3.92|3.23|
|
19 |
+
|enhanced_test|1.73|87.36|12.34|11.47|24.51|10.74|2.99|3.29|3.91|3.25|
|
20 |
+
|
exp/enh_train_raw/config.yaml
ADDED
@@ -0,0 +1,234 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
config: ./conf/train.yaml
|
2 |
+
print_config: false
|
3 |
+
log_level: INFO
|
4 |
+
dry_run: false
|
5 |
+
iterator_type: chunk
|
6 |
+
output_dir: exp/enh_train_raw
|
7 |
+
ngpu: 1
|
8 |
+
seed: 0
|
9 |
+
num_workers: 2
|
10 |
+
num_att_plot: 3
|
11 |
+
dist_backend: nccl
|
12 |
+
dist_init_method: env://
|
13 |
+
dist_world_size: 4
|
14 |
+
dist_rank: 0
|
15 |
+
local_rank: 0
|
16 |
+
dist_master_addr: localhost
|
17 |
+
dist_master_port: 43837
|
18 |
+
dist_launcher: null
|
19 |
+
multiprocessing_distributed: true
|
20 |
+
unused_parameters: true
|
21 |
+
sharded_ddp: false
|
22 |
+
cudnn_enabled: true
|
23 |
+
cudnn_benchmark: false
|
24 |
+
cudnn_deterministic: true
|
25 |
+
collect_stats: false
|
26 |
+
write_collected_feats: false
|
27 |
+
skip_stats_npz: false
|
28 |
+
max_epoch: 100
|
29 |
+
patience: 20
|
30 |
+
val_scheduler_criterion:
|
31 |
+
- valid
|
32 |
+
- loss
|
33 |
+
early_stopping_criterion:
|
34 |
+
- valid
|
35 |
+
- loss
|
36 |
+
- min
|
37 |
+
best_model_criterion:
|
38 |
+
- - valid
|
39 |
+
- snr
|
40 |
+
- max
|
41 |
+
- - valid
|
42 |
+
- loss
|
43 |
+
- min
|
44 |
+
keep_nbest_models: 1
|
45 |
+
nbest_averaging_interval: 0
|
46 |
+
grad_clip: 5.0
|
47 |
+
grad_clip_type: 2.0
|
48 |
+
grad_noise: false
|
49 |
+
accum_grad: 1
|
50 |
+
no_forward_run: false
|
51 |
+
resume: true
|
52 |
+
train_dtype: float32
|
53 |
+
use_amp: false
|
54 |
+
log_interval: null
|
55 |
+
use_matplotlib: true
|
56 |
+
use_tensorboard: true
|
57 |
+
create_graph_in_tensorboard: false
|
58 |
+
use_wandb: false
|
59 |
+
wandb_project: null
|
60 |
+
wandb_id: null
|
61 |
+
wandb_entity: null
|
62 |
+
wandb_name: null
|
63 |
+
wandb_model_log_interval: -1
|
64 |
+
detect_anomaly: false
|
65 |
+
pretrain_path: null
|
66 |
+
init_param: []
|
67 |
+
ignore_init_mismatch: false
|
68 |
+
freeze_param: []
|
69 |
+
num_iters_per_epoch: null
|
70 |
+
batch_size: 16
|
71 |
+
valid_batch_size: null
|
72 |
+
batch_bins: 1000000
|
73 |
+
valid_batch_bins: null
|
74 |
+
train_shape_file:
|
75 |
+
- exp/enh_stats_train_dev_16k/train/speech_mix_shape
|
76 |
+
- exp/enh_stats_train_dev_16k/train/speech_ref1_shape
|
77 |
+
- exp/enh_stats_train_dev_16k/train/enroll_ref1_shape
|
78 |
+
- exp/enh_stats_train_dev_16k/train/speech_ref2_shape
|
79 |
+
- exp/enh_stats_train_dev_16k/train/enroll_ref2_shape
|
80 |
+
valid_shape_file:
|
81 |
+
- exp/enh_stats_train_dev_16k/valid/speech_mix_shape
|
82 |
+
- exp/enh_stats_train_dev_16k/valid/speech_ref1_shape
|
83 |
+
- exp/enh_stats_train_dev_16k/valid/enroll_ref1_shape
|
84 |
+
- exp/enh_stats_train_dev_16k/valid/speech_ref2_shape
|
85 |
+
- exp/enh_stats_train_dev_16k/valid/enroll_ref2_shape
|
86 |
+
batch_type: folded
|
87 |
+
valid_batch_type: null
|
88 |
+
fold_length:
|
89 |
+
- 80000
|
90 |
+
- 80000
|
91 |
+
- 80000
|
92 |
+
- 80000
|
93 |
+
- 80000
|
94 |
+
sort_in_batch: descending
|
95 |
+
sort_batch: descending
|
96 |
+
multiple_iterator: false
|
97 |
+
chunk_length: 48000
|
98 |
+
chunk_shift_ratio: 0.5
|
99 |
+
num_cache_chunks: 1024
|
100 |
+
chunk_excluded_key_prefixes:
|
101 |
+
- enroll_ref
|
102 |
+
train_data_path_and_name_and_type:
|
103 |
+
- - dump/raw/train/wav.scp
|
104 |
+
- speech_mix
|
105 |
+
- sound
|
106 |
+
- - dump/raw/train/spk1.scp
|
107 |
+
- speech_ref1
|
108 |
+
- sound
|
109 |
+
- - dump/raw/train/enroll_spk1.scp
|
110 |
+
- enroll_ref1
|
111 |
+
- text
|
112 |
+
- - dump/raw/train/spk2.scp
|
113 |
+
- speech_ref2
|
114 |
+
- sound
|
115 |
+
- - dump/raw/train/enroll_spk2.scp
|
116 |
+
- enroll_ref2
|
117 |
+
- text
|
118 |
+
valid_data_path_and_name_and_type:
|
119 |
+
- - dump/raw/dev/wav.scp
|
120 |
+
- speech_mix
|
121 |
+
- sound
|
122 |
+
- - dump/raw/dev/spk1.scp
|
123 |
+
- speech_ref1
|
124 |
+
- sound
|
125 |
+
- - dump/raw/dev/enroll_spk1.scp
|
126 |
+
- enroll_ref1
|
127 |
+
- text
|
128 |
+
- - dump/raw/dev/spk2.scp
|
129 |
+
- speech_ref2
|
130 |
+
- sound
|
131 |
+
- - dump/raw/dev/enroll_spk2.scp
|
132 |
+
- enroll_ref2
|
133 |
+
- text
|
134 |
+
allow_variable_data_keys: false
|
135 |
+
max_cache_size: 0.0
|
136 |
+
max_cache_fd: 32
|
137 |
+
valid_max_cache_size: null
|
138 |
+
exclude_weight_decay: false
|
139 |
+
exclude_weight_decay_conf: {}
|
140 |
+
optim: adam
|
141 |
+
optim_conf:
|
142 |
+
lr: 0.001
|
143 |
+
eps: 1.0e-08
|
144 |
+
weight_decay: 0
|
145 |
+
scheduler: reducelronplateau
|
146 |
+
scheduler_conf:
|
147 |
+
mode: min
|
148 |
+
factor: 0.7
|
149 |
+
patience: 3
|
150 |
+
init: null
|
151 |
+
model_conf:
|
152 |
+
num_spk: 2
|
153 |
+
share_encoder: true
|
154 |
+
criterions:
|
155 |
+
- name: snr
|
156 |
+
conf:
|
157 |
+
eps: 1.0e-07
|
158 |
+
wrapper: fixed_order
|
159 |
+
wrapper_conf:
|
160 |
+
weight: 1.0
|
161 |
+
- name: l1_fd
|
162 |
+
conf:
|
163 |
+
only_for_test: true
|
164 |
+
wrapper: fixed_order
|
165 |
+
wrapper_conf:
|
166 |
+
weight: 0.0
|
167 |
+
- name: l1_td
|
168 |
+
conf:
|
169 |
+
only_for_test: true
|
170 |
+
wrapper: fixed_order
|
171 |
+
wrapper_conf:
|
172 |
+
weight: 0.0
|
173 |
+
- name: mse_fd
|
174 |
+
conf:
|
175 |
+
only_for_test: true
|
176 |
+
wrapper: fixed_order
|
177 |
+
wrapper_conf:
|
178 |
+
weight: 0.0
|
179 |
+
- name: mse_td
|
180 |
+
conf:
|
181 |
+
only_for_test: true
|
182 |
+
wrapper: fixed_order
|
183 |
+
wrapper_conf:
|
184 |
+
weight: 0.0
|
185 |
+
train_spk2enroll: data/train-100/spk2enroll.json
|
186 |
+
enroll_segment: 48000
|
187 |
+
load_spk_embedding: false
|
188 |
+
load_all_speakers: false
|
189 |
+
rir_scp: null
|
190 |
+
rir_apply_prob: 1.0
|
191 |
+
noise_scp: null
|
192 |
+
noise_apply_prob: 1.0
|
193 |
+
noise_db_range: '13_15'
|
194 |
+
short_noise_thres: 0.5
|
195 |
+
speech_volume_normalize: null
|
196 |
+
use_reverberant_ref: false
|
197 |
+
num_spk: 1
|
198 |
+
num_noise_type: 1
|
199 |
+
sample_rate: 8000
|
200 |
+
force_single_channel: false
|
201 |
+
channel_reordering: false
|
202 |
+
categories: []
|
203 |
+
encoder: conv
|
204 |
+
encoder_conf:
|
205 |
+
channel: 256
|
206 |
+
kernel_size: 32
|
207 |
+
stride: 16
|
208 |
+
extractor: td_speakerbeam
|
209 |
+
extractor_conf:
|
210 |
+
layer: 8
|
211 |
+
stack: 4
|
212 |
+
bottleneck_dim: 256
|
213 |
+
hidden_dim: 512
|
214 |
+
skip_dim: 256
|
215 |
+
kernel: 3
|
216 |
+
causal: false
|
217 |
+
norm_type: gLN
|
218 |
+
pre_nonlinear: prelu
|
219 |
+
nonlinear: relu
|
220 |
+
i_adapt_layer: 7
|
221 |
+
adapt_layer_type: mul
|
222 |
+
adapt_enroll_dim: 256
|
223 |
+
use_spk_emb: false
|
224 |
+
decoder: conv
|
225 |
+
decoder_conf:
|
226 |
+
channel: 256
|
227 |
+
kernel_size: 32
|
228 |
+
stride: 16
|
229 |
+
preprocessor: tse
|
230 |
+
preprocessor_conf: {}
|
231 |
+
required:
|
232 |
+
- output_dir
|
233 |
+
version: '202301'
|
234 |
+
distributed: true
|
exp/enh_train_raw/images/L1_on_Spec.png
ADDED
![]() |
exp/enh_train_raw/images/MSE_on_Spec.png
ADDED
![]() |
exp/enh_train_raw/images/TD_L1_loss.png
ADDED
![]() |
exp/enh_train_raw/images/TD_MSE_loss.png
ADDED
![]() |
exp/enh_train_raw/images/backward_time.png
ADDED
![]() |
exp/enh_train_raw/images/clip.png
ADDED
![]() |
exp/enh_train_raw/images/forward_time.png
ADDED
![]() |
exp/enh_train_raw/images/gpu_max_cached_mem_GB.png
ADDED
![]() |
exp/enh_train_raw/images/grad_norm.png
ADDED
![]() |
exp/enh_train_raw/images/iter_time.png
ADDED
![]() |
exp/enh_train_raw/images/loss.png
ADDED
![]() |
exp/enh_train_raw/images/loss_scale.png
ADDED
![]() |
exp/enh_train_raw/images/optim0_lr0.png
ADDED
![]() |
exp/enh_train_raw/images/optim_step_time.png
ADDED
![]() |
exp/enh_train_raw/images/snr_loss.png
ADDED
![]() |
exp/enh_train_raw/images/train_time.png
ADDED
![]() |
meta.yaml
ADDED
@@ -0,0 +1,8 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
espnet: '202301'
|
2 |
+
files:
|
3 |
+
model_file: exp/enh_train_raw/99epoch.pth
|
4 |
+
python: "3.8.16 (default, Mar 2 2023, 03:21:46) \n[GCC 11.2.0]"
|
5 |
+
timestamp: 1685992283.505222
|
6 |
+
torch: 2.0.1
|
7 |
+
yaml_files:
|
8 |
+
train_config: exp/enh_train_raw/config.yaml
|