Fangjun Kuang
e22bc78f98
Export streaming zipformer2 to RKNN ( #1977 )
2025-07-11 13:24:01 +08:00
Fangjun Kuang
fba5e67d5e
Fix CI tests. ( #1974 )
...
- Introduce unified AMP helpers (create_grad_scaler, torch_autocast) to handle
deprecations in PyTorch ≥2.3.0
- Replace direct uses of torch.cuda.amp.GradScaler and torch.cuda.amp.autocast
with the new utilities across all training and inference scripts
- Update all torch.load calls to include weights_only=False for compatibility with
newer PyTorch versions
2025-07-01 13:47:55 +08:00
Fangjun Kuang
71377d21cd
Export streaming zipformer models with whisper feature to onnx ( #1973 )
2025-06-30 19:01:15 +08:00
Zengwei Yao
ffb7d05635
refactor branch exchange in cr-ctc ( #1954 )
2025-05-27 12:09:59 +08:00
Tianxiang Zhao
30e7ea4b5a
Fix a bug in finetune.py --use-mux ( #1949 )
2025-05-22 12:05:01 +08:00
Yifan Yang
e79833aad2
ensure SwooshL/SwooshR output dtype matches input dtype ( #1940 )
2025-05-12 19:28:48 +08:00
Fangjun Kuang
171cf8c9fe
Avoid redundant computation in PiecewiseLinear. ( #1915 )
2025-04-09 11:52:37 +08:00
Han Zhu
ab91112909
Improve infinity-check ( #1862 )
...
1. Attach the inf-check hooks if the grad scale is getting too small.
2. Add try-catch to avoid OOM in the inf-check hooks.
3. Set warmup_start=0.1 to reduce chances of divergence
2025-01-09 15:05:38 +08:00
Han Zhu
48088cb807
Refactor optimizer ( #1837 )
...
* Print indexes of largest grad
2024-12-30 15:30:02 +08:00
Fangjun Kuang
d4d4f281ec
Revert "Replace deprecated pytorch methods ( #1814 )" ( #1841 )
...
This reverts commit 3e4da5f78160d3dba3bdf97968bd7ceb8c11631f.
2024-12-18 16:49:57 +08:00
Li Peng
3e4da5f781
Replace deprecated pytorch methods ( #1814 )
...
* Replace deprecated pytorch methods
- torch.cuda.amp.GradScaler(...) => torch.amp.GradScaler("cuda", ...)
- torch.cuda.amp.autocast(...) => torch.amp.autocast("cuda", ...)
* Replace `with autocast(...)` with `with autocast("cuda", ...)`
Co-authored-by: Li Peng <lipeng@unisound.ai>
2024-12-16 10:24:16 +08:00
zr_jin
87cadfcd2e
fixed formatting issue ( #1791 )
...
* isort fixed formatting issue
2024-10-30 21:14:12 +08:00
Wei Kang
d513d456b8
Add prefix beam search and corresponding decoding methods ( #1786 )
...
* Add prefix beam search / shallow fussion / hotwords in librispeech ctc decode
* Add librispeech cr-ctc prefix beam search results
2024-10-30 10:14:34 +08:00
Fangjun Kuang
05f756390c
Avoid using lr from checkpoint. ( #1781 )
2024-10-28 00:59:04 +08:00
zr_jin
88bacfb9e6
minor fixes for the repo ( #1775 )
...
* minor fixes for the repo
Co-authored-by: Fangjun Kuang <csukuangfj@gmail.com>
2024-10-21 13:51:56 +08:00
zr_jin
e8b6b920c0
A LibriTTS recipe on both ASR & Neural Codec Tasks ( #1746 )
...
* added ASR & CODEC recipes for LibriTTS corpus
2024-10-21 11:30:14 +08:00
Zengwei Yao
693d84a301
Add Consistency-Regularized CTC ( #1766 )
...
* support consistency-regularized CTC
* update arguments of cr-ctc
* set default value of cr_loss_masked_scale to 1.0
* minor fix
* refactor codes
* update RESULTS.md
2024-10-21 10:35:26 +08:00
Zengwei Yao
fbba712887
Fix issue with eval mode in ActivationDropoutLinear ( #1770 )
...
* Fix issue with eval mode in ActivationDropoutLinear
---------
Co-authored-by: Daniel Povey <dpovey@gmail.com>
2024-10-12 19:09:05 +08:00
Fangjun Kuang
6f1abd832d
Fix exporting streaming zipformer models. ( #1755 )
2024-09-11 21:04:52 +08:00
Xiaoyu Yang
a6c02a4d8c
zipformer BF16 training recipe ( #1700 )
...
Support Zipformer AMP +BF16 training
2024-08-23 09:42:22 +08:00
Yuekai Zhang
3b434fe83c
fix triton onnx export ( #1730 )
2024-08-23 09:33:46 +08:00
Karel Vesely
1730fce688
split save_results() -> save_asr_output() + save_wer_results() ( #1712 )
...
- the idea is to support `--skip-scoring` argument passed to a decoding
script
- created for Transducer decoding (non-streaming, streaming)
- it can be done also for CTC decoding... (not yet)
- also added `--label` for extra label in `streaming_decode.py`
- and also added `set_caching_enabled(True)`, which has no effect on
librispeech, but it leads to faster runtime on DBs with long
recordings (assuming `librispeech/zipformer` scripts are the
example scripts for other setups)
2024-08-13 23:02:14 +08:00
Zengwei Yao
d47c078286
add decoding method of ctc-greedy-search in zipformer recipe ( #1690 )
2024-07-14 17:30:13 +08:00
Zengwei Yao
334beed2af
fix usages of returned losses after adding attention-decoder in zipformer ( #1689 )
2024-07-12 16:50:58 +08:00
Teo Wen Shen
19048e155b
Cast grad_scale in whiten to float ( #1663 )
...
* cast grad_scale in whiten to float
* fix cast in zipformer_lora
2024-07-11 15:12:30 +08:00
Yifan Yang
d65187ec52
Small fix ( #1686 )
2024-07-11 14:45:35 +08:00
zr_jin
2d64228efa
Update attention_decoder.py ( #1681 )
2024-07-06 09:01:34 +08:00
Zengwei Yao
f76afff741
Support CTC/AED option for Zipformer recipe ( #1389 )
...
* add attention-decoder loss option for zipformer recipe
* add attention-decoder-rescoring
* update export.py and pretrained_ctc.py
* update RESULTS.md
2024-07-05 20:19:18 +08:00
Yifan Yang
cbcac23d26
Fix typos, remove unused packages, normalize comments ( #1678 )
2024-07-04 14:19:45 +08:00
Manix
eaab2c819f
Zipformer Onnx FP16 ( #1671 )
...
Signed-off-by: manickavela29 <manickavela1998@gmail.com>
2024-06-27 16:08:24 +08:00
Fangjun Kuang
b88062292b
Typo fixes ( #1643 )
2024-06-03 16:49:21 +08:00
Zengwei Yao
0df406c5da
Initialize BiasNorm bias with small random values ( #1630 )
2024-05-20 22:32:02 +08:00
Yifan Yang
ed6bc200e3
Update train.py ( #1590 )
2024-04-11 19:35:25 +08:00
Zengwei Yao
353469182c
fix issue in zipformer.py ( #1566 )
2024-03-21 15:59:43 +08:00
Fangjun Kuang
489263e5bb
Add streaming HLG decoding for zipformer CTC. ( #1557 )
...
Note it supports only CPU.
2024-03-18 20:11:47 +08:00
Karel Vesely
4917ac8bab
allow export of onnx-streaming-models with other than 80dim input features ( #1556 )
2024-03-18 18:43:29 +08:00
zr_jin
eb132da00d
additional instruction for the grad_scale is too small error ( #1550 )
2024-03-14 11:33:49 +08:00
zr_jin
242002e0bd
Strengthened style constraints ( #1527 )
2024-03-04 23:28:04 +08:00
Xiaoyu Yang
777074046d
Fine-tune recipe for Zipformer ( #1484 )
...
1. support finetune zipformer
2. update the usage; set a very large batch count
2024-02-06 18:25:43 +08:00
Teo Wen Shen
b9e6327adf
Fixing torch.ctc err ( #1485 )
...
* fixing torch.ctc err
* Move targets & lengths to CPU
2024-02-03 06:25:27 +08:00
Zengwei Yao
c401a2646b
minor fix of zipformer/optim.py ( #1474 )
2024-01-26 15:50:11 +08:00
zr_jin
5445ea6df6
Use shuffled LibriSpeech cuts instead ( #1450 )
...
* use shuffled LibriSpeech cuts instead
* leave the old code in comments for reference
2024-01-08 15:09:21 +08:00
Karel Vesely
716b82cc3a
streaming_decode.py, relax the audio range from [-1,+1] to [-10,+10] ( #1448 )
...
- some AudioTransform classes produce audio signals out of range [-1,+1]
- Resample produced 1.0079
- The range [-10,+10] was chosen to still be able to reliably
distinguish from the [-32k,+32k] signal...
- this is related to : https://github.com/lhotse-speech/lhotse/issues/1254
2024-01-05 10:21:27 +08:00
Fangjun Kuang
8136ad775b
Use high_freq -400 in computing fbank features. ( #1447 )
...
See also https://github.com/k2-fsa/sherpa-onnx/issues/514
2024-01-04 13:59:32 +08:00
Fangjun Kuang
79a42148db
Add CI test to cover zipformer/train.py ( #1424 )
2023-12-23 00:38:36 +08:00
Fangjun Kuang
f85f0252a9
Add greedy search for streaming zipformer CTC. ( #1415 )
2023-12-13 17:34:12 +08:00
zr_jin
d0da509055
Support ONNX export for Streaming CTC Encoder ( #1413 )
...
* Create export-onnx-streaming-ctc.py
* doc_str updated
Co-authored-by: Fangjun Kuang <csukuangfj@gmail.com>
---------
Co-authored-by: Fangjun Kuang <csukuangfj@gmail.com>
2023-12-13 10:33:28 +08:00
Fangjun Kuang
20a82c9abf
first commit ( #1411 )
2023-12-12 18:13:26 +08:00
Wei Kang
11d816d174
Add cumstomized score for hotwords ( #1385 )
...
* add custom score for each hotword
* Add more comments
* Fix deocde
* fix style
* minor fixes
2023-11-18 18:47:55 +08:00
zr_jin
231bbcd2b6
Update optim.py ( #1366 )
2023-11-03 12:06:29 +08:00
