Training Help

[jkurian49]

Hi! I am running into a segmentation fault when trying to run train.sh, so I wanted to ask what the preprocessing steps are in case I am missing something. Besides installing the dependences and running extractor.py and prepro.py, are there any steps required before running train.sh?

[wtnan2003]

Same Questions @stonyhu

[CrystalWang1225]

Same Questions @stonyhu

It seems that the longformer is giving me a seg fault. Are you experiencing the same thing?

[CrystalWang1225]

Same Questions @stonyhu

I have solved the issue by reinstalling the transformer. I looked into the longformer package and it looks like the transformer importing isnt successful

[stonyhu]

@CrystalWang1225 Actually, you can implement your own local self-attention network by naive pytorch. But you can also implement it based on longformer that provides the custom CUDA kernel to accelerate computation and reduce gpu memory cost.

[klong121]

Hello @CrystalWang1225 , could you please share your method for reinstalling transformer? I reinstall transformer by pip uninstall transformers;pip install transformers==2.5.1 but still run into segmentation fault. I also try to install transformer from source by the following code, but segmentation fault still occurs.

git clone https://github.com/huggingface/transformers.git
cd transformers
pip install -e .

Thanks in advance!

[CrystalWang1225]

@klong121 So I downgraded transformers to pip install transformers==3.5.0 The train.pyis working for us for now. These are our requirements we have on our CUDA:

absl-py==0.11.0
appdirs==1.4.4
audioread==2.1.9
boto3==1.16.35
botocore==1.19.35
cachetools==4.2.0
certifi==2020.12.5
cffi==1.14.4
chardet==3.0.4
click==7.1.2
decorator==4.4.2
distlib==0.3.1
essentia==2.1b6.dev234
filelock==3.0.12
future==0.18.2
google-auth==1.24.0
google-auth-oauthlib==0.4.2
grpcio==1.34.0
idna==2.10
imageio==2.9.0
importlib-metadata==2.0.0
jmespath==0.10.0
joblib==0.17.0
librosa==0.8.0
llvmlite==0.35.0
Markdown==3.3.3
numba==0.52.0
numpy==1.19.4
oauthlib==3.1.0
packaging==20.8
pandas==1.1.5
Pillow==8.0.1
pooch==1.3.0
protobuf==3.14.0
pyasn1==0.4.8
pyasn1-modules==0.2.8
pycparser==2.20
pyparsing==2.4.7
python-dateutil==2.8.1
pytorch-lightning==0.6.0
pytz==2020.4
PyYAML==5.3.1
regex==2020.11.13
requests==2.25.0
requests-oauthlib==1.3.0
resampy==0.2.2
rsa==4.6
s3transfer==0.3.3
sacremoses==0.0.43
scikit-learn==0.23.2
scipy==1.5.4
sentencepiece==0.1.91
six==1.15.0
SoundFile==0.10.3.post1
tensorboard==2.4.0
tensorboard-plugin-wit==1.7.0
tensorboardX==2.1
test-tube==0.7.5
threadpoolctl==2.1.0
tokenizers==0.9.3
torch==1.3.1
torchvision==0.4.2
tqdm==4.54.1
transformers==3.5.1
urllib3==1.26.2
virtualenv==20.0.35
Werkzeug==1.0.1
zipp==3.3.1

[klong121]

Thanks for your help! I downgraded the transformers and the code is working for me now. @CrystalWang1225

[aryanna384]

Hi, I am confused about the meaning of BOS_POSE in pose.py which used while training, is the initialization of the skeleton? How does it get? BOS_POSE = np.array([ -0.00647944, -0.540766, -0.00655532, -0.3829, -0.0862469, -0.388304, -0.107649, -0.137507, -0.0648178, 0.0423361, 0.0791047, -0.382726, 0.0882645, -0.148286, 0.0699688, 0.0641147, 0.0056175, 0.05319, -0.0464647, 0.0476973, -0.0465082, 0.467254, -0.031224, 0.854397, 0.0576743, 0.0585055, 0.0394117, 0.478282, 0.0208931, 0.854322, -0.0158758, -0.568058, 0.0147899, -0.568106, -0.0465137, -0.568036, 0.0391742, -0.562693, 0.0209782, 0.936187, 0.0424097, 0.930442, 0.0148087, 0.886806, -0.0249588, 0.935932, -0.0434218, 0.935805, -0.0280986, 0.886854 ])

THANKS VERY MUCH!