2

conda by default installing transformers 2.x however pip installs 4.x by default which is what I want but via conda.

If I install by specifying the latest distribution file from conda-forge… conda install https://anaconda.org/conda-forge/transformers/4.16.2/download/noarch/transformers-4.16.2-pyhd8ed1ab_0.tar.bz2

then it complains that environment is not consistent and list the package causing the issue which is PyTorch version 1.11.

I removed pytorch and then installed as listed above then installation go through. I then tried installing datasets… conda install datasets

Now it complains that environment is inconsistent due to transformers 4.16.2 that I installed.

Not sure whats wrong and how to install pytorch, transformers and datasets together with no issue. Do I need specific versions of these to make it work, could not find any such guideline on huggingface docs or support pages.

thanks.

Gonçalo Peres
  • 11,752
  • 3
  • 54
  • 83
Sunil Surendra Singh
  • 21
  • 1
  • 4
  • 2
    Could you make a reproducing example in a new environment? We don't know what else you have installed that could be causing conflicts. And making a new environment fixes this sort of issue in many cases. – Matt Thompson Apr 05 '22 at 18:30
  • Tried in a new blank environment and first package I tried to install transformers ... – Sunil Surendra Singh Apr 09 '22 at 06:58
  • continued.. `conda install -c huggingface transformers`. This time it picked up `transformers version 4.x` and `python version 3.8x`. Now, if I first install `python 3.9.x` (which is default with miniconda) and then try to install transformer then it falls back to version 2.x. This tell me that in order to have version 4.x I need python version 3.8 or lower. This is not listed anywhere in huggingface installation page though. – Sunil Surendra Singh Apr 09 '22 at 07:06
  • That's surprising; it's built as a `noarch` package which has the consequence of typically being compatible with all versions of Python. I suspect something is pulled down incorrectly in its dependencies. The documentation only gives the instructions you've followed and makes no mention of Python 3.9 incompatibility - therefore I suggest raising this as an issue with the maintainers directly. Make sure to include details logs, etc. and whatever else they may ask for in their issue prompt. – Matt Thompson Apr 12 '22 at 03:22

1 Answers1

3

Just tried installing transformers in the venv that I will be working on as follows

conda install transformers

And it installed the version transformers-4.18.0.

If one wants to install specifically from the channel huggingface, then do the following

conda install -c huggingface transformers

Last case scenario, install it with pip, as using pip can potentially wreck one's installation (pip and conda do not manage dependencies in the same way).

pip install transformers

Or, by specifying the version

pip install transformers==4.18.0

Or directly from the source

pip install git+https://github.com/huggingface/transformers

If one has transformers already installed and wants to install a different version than the one we currently have, one should pass -Iv (as suggested here)

pip install -Iv transformers==4.18.0

To check if transformers was properly installed, run the following

python -c "from transformers import pipeline; print(pipeline('sentiment-analysis')('we love you'))"

It will download a pretrained model, then print out the label and score.

For more information on transformers installation, consult this page.

Gonçalo Peres
  • 11,752
  • 3
  • 54
  • 83