JaheimLee

I have the same problem in Ubuntu 20.04 using pip in my conda env. The log is: ``` Using pip 22.2.1 from /data/miniconda3/envs/k2/lib/python3.9/site-packages/pip (python 3.9) Looking in indexes: https://pypi.tuna.tsinghua.edu.cn/simple Collecting...

By the way, I found a similar issue [here](https://github.com/microsoft/onnxruntime/issues/9170). But I don't know how to manually set the link.

> find /data/miniconda3/envs/k2 -name "libmkl_intel_thread*" The output is: ``` (k2) lijinghui@idc-op-dev-gpu-001:/data/lijinghui$ find /data/miniconda3/envs/k2 -name "libmkl_intel_ilp64*" /data/miniconda3/envs/k2/lib/libmkl_intel_ilp64.so /data/miniconda3/envs/k2/lib/libmkl_intel_ilp64.so.2 (k2) lijinghui@idc-op-dev-gpu-001:/data/lijinghui$ find /data/miniconda3/envs/k2 -name "libmkl_core*" /data/miniconda3/envs/k2/lib/libmkl_core.so /data/miniconda3/envs/k2/lib/libmkl_core.so.2 (k2) lijinghui@idc-op-dev-gpu-001:/data/lijinghui$ find /data/miniconda3/envs/k2 -name...

> Could you use > > ```shell > export LIBRARY_PATH=/data/miniconda3/envs/k2/lib:$LIBRARY_PATH > ``` > > and run `pip install --verbose kaldifeat` again? > > **Caution**: It is `LIBRARY_PATH`, not `LD_LIBRARY_PATH`. I...

> Could you use `pip install --verbose kaldifeat` to get more verbose logs? > > By the way, you can use `conda install -c kaldifeat kaldifeat` Conda install works well....

> pooled_output 就是 [CLS]，参见[代码](https://github.com/huggingface/transformers/blob/b5e2b183af5e40e33a4dc7659e697d137259d56e/src/transformers/models/bert/modeling_bert.py#L636) 他的pool操作是cls，dense，最后接具体任务的classifier。我想明确你们是否是cls直接到具体任务的classifier，中间有没有那个dense。看你们论文描述像是没有的

same problem on windows

conda install > export LD_LIBRARY_PATH=$LD_LIBRARY_PATH:/data/miniconda3/envs/k2_py38/lib/ My OS is Ubuntu18.04. And my Python version is 3.9.13. I use conda like: ``` conda install pytorch=1.12.0 torchvision torchaudio cudatoolkit=10.2 -c pytorch conda install...

The above env 'k2_py38' is not correct, just ignore it

> ``` > -c k2-fsa \ > -c k2-fsa-sherpa \ > -c kaldifeat \ > -c kaldi_native_io \ > k2 \ > sherpa \ > kaldifeat \ > kaldi_native_io >...