Merge branch 'k2-fsa:master' into dev/cv-zipformer

k2-fsa · Mar 23, 2024 · 1d92107 · 1d92107
2 parents c274003 + bb9ebcf
commit 1d92107
Show file tree

Hide file tree

Showing 7 changed files with 545 additions and 4 deletions.
diff --git a/.github/workflows/ljspeech.yml b/.github/workflows/ljspeech.yml
@@ -90,7 +90,7 @@ jobs:
           path: ./*.wav
 
       - name: Release exported onnx models
-        if: matrix.python-version == '3.9' && matrix.torch-version == '2.2.0'
+        if: matrix.python-version == '3.9' && matrix.torch-version == '2.2.0' && github.event_name == 'push'
         uses: svenstaro/upload-release-action@v2
         with:
           file_glob: true

diff --git a/docs/source/for-dummies/environment-setup.rst b/docs/source/for-dummies/environment-setup.rst
@@ -74,6 +74,10 @@ to install dependencies of `icefall`_:
 
    pip install k2==1.24.4.dev20231220+cpu.torch2.0.0 -f https://k2-fsa.github.io/k2/cpu.html
 
+   # For users from China
+   # 中国国内用户，如果访问不了 huggingface, 请使用
+   # pip install k2==1.24.4.dev20231220+cpu.torch2.0.0 -f https://k2-fsa.github.io/k2/cpu-cn.html
+
    # Install the latest version of lhotse
 
    pip install git+https://github.com/lhotse-speech/lhotse

diff --git a/docs/source/installation/index.rst b/docs/source/installation/index.rst
@@ -206,6 +206,9 @@ We will install `k2`_ from pre-compiled wheels by following
 .. code-block:: bash
 
   (test-icefall) kuangfangjun:~$ pip install k2==1.24.3.dev20230725+cuda11.6.torch1.13.0 -f https://k2-fsa.github.io/k2/cuda.html
+  # For users from China
+  # 中国国内用户，如果访问不了 huggingface, 请使用
+  # pip install k2==1.24.3.dev20230725+cuda11.6.torch1.13.0 -f https://k2-fsa.github.io/k2/cuda-cn.html
 
   Looking in indexes: https://pypi.tuna.tsinghua.edu.cn/simple
   Looking in links: https://k2-fsa.github.io/k2/cuda.html

diff --git a/egs/librispeech/ASR/zipformer/zipformer.py b/egs/librispeech/ASR/zipformer/zipformer.py
@@ -788,7 +788,7 @@ def forward(
         selected_attn_weights = attn_weights[0:1]
         if torch.jit.is_scripting() or torch.jit.is_tracing():
             pass
-        elif not self.training and random.random() < float(self.const_attention_rate):
+        elif self.training and random.random() < float(self.const_attention_rate):
             # Make attention weights constant.  The intention is to
             # encourage these modules to do something similar to an
             # averaging-over-time operation.

diff --git a/egs/librispeech/ASR/zipformer_adapter/export-onnx.py b/egs/librispeech/ASR/zipformer_adapter/export-onnx.py
@@ -27,11 +27,13 @@
 
 2. Export the model to ONNX
 
-./zipformer/export-onnx.py \
+./zipformer_adapter/export-onnx.py \
   --tokens $repo/data/lang_bpe_500/tokens.txt \
   --use-averaged-model 0 \
   --epoch 99 \
   --avg 1 \
+  --use-adapters 1 \
+  --adapter-dim 32 \
   --exp-dir $repo/exp \
   --num-encoder-layers "2,2,3,4,3,2" \
   --downsampling-factor "1,2,4,8,4,2" \
@@ -131,7 +133,7 @@ def get_parser():
     parser.add_argument(
         "--exp-dir",
         type=str,
-        default="zipformer/exp",
+        default="zipformer_adapter/exp",
         help="""It specifies the directory where all training related
         files, e.g., checkpoints, log, etc, are saved
         """,