From fd8f8780fac68d959df0bf521f1b76aac35f88fd Mon Sep 17 00:00:00 2001 From: Fangjun Kuang Date: Wed, 21 May 2025 12:04:57 +0800 Subject: [PATCH 1/4] Fix logging torch.dtype. (#1947) --- icefall/utils.py | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/icefall/utils.py b/icefall/utils.py index aab479e56..ffb926566 100644 --- a/icefall/utils.py +++ b/icefall/utils.py @@ -186,7 +186,7 @@ class AttributeDict(dict): tmp = {} for k, v in self.items(): # PosixPath is ont JSON serializable - if isinstance(v, pathlib.Path) or isinstance(v, torch.device): + if isinstance(v, (pathlib.Path, torch.device, torch.dtype)): v = str(v) tmp[k] = v return json.dumps(tmp, indent=indent, sort_keys=True) From 30e7ea4b5a0449bfac4e7fa5eef9a23238230013 Mon Sep 17 00:00:00 2001 From: Tianxiang Zhao <162714929+Redemption-ZTX@users.noreply.github.com> Date: Thu, 22 May 2025 12:05:01 +0800 Subject: [PATCH 2/4] Fix a bug in finetune.py --use-mux (#1949) --- egs/librispeech/ASR/zipformer/finetune.py | 6 +++--- 1 file changed, 3 insertions(+), 3 deletions(-) diff --git a/egs/librispeech/ASR/zipformer/finetune.py b/egs/librispeech/ASR/zipformer/finetune.py index 2ff631914..2c869a57a 100755 --- a/egs/librispeech/ASR/zipformer/finetune.py +++ b/egs/librispeech/ASR/zipformer/finetune.py @@ -140,8 +140,8 @@ def add_finetune_arguments(parser: argparse.ArgumentParser): type=str2bool, default=False, help=""" - Whether to adapt. If true, we will mix 5% of the new data - with 95% of the original data to fine-tune. This is useful + Whether to adapt. If true, we will mix 5%% of the new data + with 95%% of the original data to fine-tune. This is useful if you want to maintain the performance on the original domain """, ) @@ -1134,7 +1134,7 @@ def train_one_epoch( f"Epoch {params.cur_epoch}, " f"batch {batch_idx}, loss[{loss_info}], " f"tot_loss[{tot_loss}], batch size: {batch_size}, " - f"lr: {cur_lr:.2e}, " + f"lr: {cur_lr: .2e}, " + (f"grad_scale: {scaler._scale.item()}" if params.use_fp16 else "") ) From 021e1a88469f5ca8299daa4000d05a5dd89b8623 Mon Sep 17 00:00:00 2001 From: Mahsa Yarmohammadi <35142904+mahsa7823@users.noreply.github.com> Date: Thu, 22 May 2025 10:06:35 -0400 Subject: [PATCH 3/4] Add acknowledgment to README (#1950) --- README.md | 4 ++++ 1 file changed, 4 insertions(+) diff --git a/README.md b/README.md index 0e550ffb1..498f7e3b4 100644 --- a/README.md +++ b/README.md @@ -383,3 +383,7 @@ Please see: [![Open In Colab](https://colab.research.google.com/assets/colab-bad [vctk]: egs/vctk/TTS [ljspeech]: egs/ljspeech/TTS [libritts_tts]: egs/libritts/TTS + +## Acknowledgements + +Some contributors to this project were supported by Xiaomi Corporation. Others were supported by National Science Foundation CCRI award 2120435. This is not an exhaustive list of sources of support. From ffb7d0563526f42123d2c65c284646a5c375b74f Mon Sep 17 00:00:00 2001 From: Zengwei Yao Date: Tue, 27 May 2025 12:09:59 +0800 Subject: [PATCH 4/4] refactor branch exchange in cr-ctc (#1954) --- egs/librispeech/ASR/zipformer/model.py | 8 ++++---- 1 file changed, 4 insertions(+), 4 deletions(-) diff --git a/egs/librispeech/ASR/zipformer/model.py b/egs/librispeech/ASR/zipformer/model.py index c7dbe1e0a..f2791e51f 100644 --- a/egs/librispeech/ASR/zipformer/model.py +++ b/egs/librispeech/ASR/zipformer/model.py @@ -210,10 +210,10 @@ class AsrModel(nn.Module): ) # Compute consistency regularization loss - exchanged_targets = ctc_output.detach().chunk(2, dim=0) - exchanged_targets = torch.cat( - [exchanged_targets[1], exchanged_targets[0]], dim=0 - ) # exchange: [x1, x2] -> [x2, x1] + batch_size = ctc_output.shape[0] + assert batch_size % 2 == 0, batch_size + # exchange: [x1, x2] -> [x2, x1] + exchanged_targets = torch.roll(ctc_output.detach(), batch_size // 2, dims=0) cr_loss = nn.functional.kl_div( input=ctc_output, target=exchanged_targets,