Hello,
I am newbie of pytorch-pretrained-Bert.
After successfully converted from init-checkpoint of tensorflow to pytorch bin,
I found an error when I do run_squad.
Guessing I should've included some configuration ahead, could anyone can help?
See below.
File "run_squad.py", line 37, in <module>
from pytorch_pretrained_bert.file_utils import PYTORCH_PRETRAINED_BERT_CACHE, WEIGHTS_NAME, CONFIG_NAME
ImportError: No module named pytorch_pretrained_bert.file_utils
Did you install pytorch-pretrained-bert as indicated in the README?
pip install pytorch_pretrained_bert
You don't have to convert the checkpoints yourself, there are already converted.
Try reading the installation and usage sections of the README.
Of cause I installed,
More precisely, error code is slightly changed.
Traceback (most recent call last):
File "run_squad.py", line 37, in <module>
from pytorch_pretrained_bert.file_utils import PYTORCH_PRETRAINED_BERT_CACHE, WEIGHTS_NAME, CONFIG_NAME
ImportError: cannot import name WEIGHTS_NAME
Hmm you are right, the examples are compatible with master only now that we have a new token serialization. I guess we'll have to do a new release (0.6.2) today so everybody is on the same page.
Let me do that.
Actually, we'll wait for the merge of #506.
In the meantime you can install from source and it should work.
Oh it's done immediately when I installed from source. Thanks.
Ok great.
Just a side note on writing messages in github: you should add triple-quotes like this: ``` before and after the command line, errors and code you are pasting. This way it's easier to read.
Ex:
```
pip install -e .
```
will display like:
pip install -e .
Good point(triple quotes).
I didn't know what to do, but now I have it all.
Thanks.
Actually, we'll wait for the merge of #506.
In the meantime you can install from source and it should work.
how to "install from source"?
@YanZhangADS
You can install from source with this command below
git clone https://github.com/huggingface/pytorch-pretrained-BERT.git
cd pytorch-pretrained-BERT
python setup.py install
Same problem with "ImportError: cannot import name WEIGHTS_NAME". However, after building 0.6.1 from source, I get:
from pytorch_pretrained_bert.optimization import BertAdam, warmup_linear
ImportError: cannot import name 'warmup_linear'
I don't need the warmup, so I removed the import, but letting you guys know that this is an import error as well. Thanks!
Thanks for that @dumitrescustefan, we're working on it in #518.
I'm closing this issue for now as we start to deviate from the original discussion.
I just built from source. I'm still getting the same error as in original issue.
The version 0.4.0 doesn't give this issue.
pip install pytorch_pretrained_bert==0.4.0
Most helpful comment
@YanZhangADS
You can install from source with this command below