"NameError: name 'torch' is not defined" but torch is installed and imported

[giulianoformisano]

After creating a virtual environment, I tried to install and import m3inference:

pip install m3inference
import m3inference

But I get the following error, how could I fix it?

NameError                                 Traceback (most recent call last)
<ipython-input-9-50ee37ff85fa> in <module>
----> 1 import m3inference

3 frames
/usr/local/lib/python3.8/dist-packages/m3inference/full_model.py in M3InferenceModel()
     10 
     11 class M3InferenceModel(nn.Module):
---> 12     def __init__(self, device='cuda' if torch.cuda.is_available() else 'cpu'):
     13         super(M3InferenceModel, self).__init__()
     14 

NameError: name 'torch' is not defined

I tried to install and import torch before doing the same with m3inference.

Thanks!

[computermacgyver]

Can you please follow these directions to install pytorch
https://pytorch.org/get-started/locally/
?

If it still does not work, please provide the specific versions of packages you have installed using pip freeze

[zijwang]

@giulianoformisano could you try to verify whether you installed torch correctly? You could try import torch in Python console inside your virtualenv.

[giulianoformisano]

Thanks a lot, unfortunately, I couldn't fix the issue following the instructions (https://pytorch.org/get-started/locally/).

Please find all my installed packages:

appnope==0.1.3
asttokens==2.2.1
backcall==0.2.0
certifi==2022.12.7
charset-normalizer==2.1.1
comm==0.1.2
contourpy==1.0.6
cycler==0.11.0
debugpy==1.6.4
decorator==5.1.1
entrypoints==0.4
executing==1.2.0
fonttools==4.38.0
idna==3.4
ipykernel==6.19.4
ipython==8.7.0
jedi==0.18.2
jupyter_client==7.4.8
jupyter_core==5.1.1
kiwisolver==1.4.4
m3inference==1.1.5
matplotlib==3.6.2
matplotlib-inline==0.1.6
nest-asyncio==1.5.6
numpy==1.24.1
packaging==22.0
pandas==1.5.2
parso==0.8.3
pexpect==4.8.0
pickleshare==0.7.5
Pillow==9.3.0
platformdirs==2.6.0
prompt-toolkit==3.0.36
psutil==5.9.4
ptyprocess==0.7.0
pure-eval==0.2.2
pycld2==0.41
Pygments==2.13.0
pyparsing==3.0.9
python-dateutil==2.8.2
pytz==2022.7
pyzmq==24.0.1
rauth==0.7.3
requests==2.28.1
seaborn==0.12.1
six==1.16.0
stack-data==0.6.2
torch==1.13.1
torchvision==0.14.1
tornado==6.2
tqdm==4.64.1
traitlets==5.8.0
typing_extensions==4.4.0
urllib3==1.26.13
wcwidth==0.2.5

[wdwgonzales]

@computermacgyver @zijwang

Everything was working fine earlier this year. When I tried running the script this time around, I encountered the same problem as @giulianoformisano. Some context: I work with an M1 mac and torch seems to be working just fine for me.

(base) ➜  scripts python m3twitter.py --id=19854920 --auth auth.txt --skip-cache
Traceback (most recent call last):
  File "/Users/wdwg/Desktop/scripts/m3twitter.py", line 4, in <module>
    from m3inference import M3Twitter
  File "/Users/wdwg/opt/anaconda3/lib/python3.9/site-packages/m3inference/__init__.py", line 1, in <module>
    from .m3inference import M3Inference
  File "/Users/wdwg/opt/anaconda3/lib/python3.9/site-packages/m3inference/m3inference.py", line 14, in <module>
    from .full_model import M3InferenceModel
  File "/Users/wdwg/opt/anaconda3/lib/python3.9/site-packages/m3inference/full_model.py", line 11, in <module>
    class M3InferenceModel(nn.Module):
  File "/Users/wdwg/opt/anaconda3/lib/python3.9/site-packages/m3inference/full_model.py", line 12, in M3InferenceModel
    def __init__(self, device='cuda' if torch.cuda.is_available() else 'cpu'):
NameError: name 'torch' is not defined

I tried to fix it by adding import torch, torchvision in the downloaded scripts. I've also run the following command, to update pytorch:

conda install pytorch torchvision torchaudio -c pytorch-nightly

After that, I added map_location=torch.device('mps') as an argument wherever there is torch.load(...) in the scripts, specifically in the m3inference.py file. MPS was selected because of a recent update. I tried "cpu" but it didn't work.

Anyway, running the program with the above adjustments partially fixed the problem. Now, I am stuck again with a segmentation error.

(base) ➜  scripts python m3twitter.py --id=19854920 --auth auth.txt --skip-cache
12/31/2022 17:46:16 - INFO - m3inference.m3inference -   Version 1.1.5
12/31/2022 17:46:16 - INFO - m3inference.m3inference -   Running on cpu.
12/31/2022 17:46:16 - INFO - m3inference.m3inference -   Will use full M3 model.
12/31/2022 17:46:17 - INFO - m3inference.m3inference -   Model full_model exists at /Users/wdwg/m3/models/full_model.mdl.
12/31/2022 17:46:17 - INFO - m3inference.utils -   Checking MD5 for model full_model at /Users/wdwg/m3/models/full_model.mdl
12/31/2022 17:46:17 - INFO - m3inference.utils -   MD5s match.
12/31/2022 17:46:18 - INFO - m3inference.m3inference -   Loaded pretrained weight at /Users/wdwg/m3/models/full_model.mdl
/Users/wdwg/Desktop/scripts
12/31/2022 17:46:18 - INFO - m3inference.m3twitter -   skip_cache is True. Fetching data from Twitter for id 19854920.
12/31/2022 17:46:18 - INFO - m3inference.m3twitter -   GET /users/show.json?id=19854920
12/31/2022 17:46:18 - INFO - m3inference.dataset -   1 data entries loaded.
Predicting...:   0%|                                                    | 0/1 [00:00<?, ?it/s][1]    
27886 segmentation fault  python m3twitter.py --id=19854920 --auth auth.txt --skip-cache
(base) ➜  scripts /Users/wdwg/opt/anaconda3/lib/python3.9/multiprocessing/resource_tracker.py:216: UserWarning: resource_tracker: There appear to be 12 leaked semaphore objects to clean up at shutdown
  warnings.warn('resource_tracker: There appear to be %d '

I use this package heavily (and cite it whenever I can). It would be a shame for it not to work in future projects. :(

[wdwgonzales]

Update:

I switched environments (switched to pytorch-nightly) using conda activate torch-nightly, and made the same modifications as detailed above, and it worked?

(torch-nightly) ➜  scripts python m3twitter.py --screen-name=barackobama --auth auth.txt
12/31/2022 18:13:52 - INFO - m3inference.m3inference -   Version 1.1.5
12/31/2022 18:13:52 - INFO - m3inference.m3inference -   Running on cpu.
12/31/2022 18:13:52 - INFO - m3inference.m3inference -   Will use full M3 model.
12/31/2022 18:13:53 - INFO - m3inference.m3inference -   Model full_model exists at /Users/wdwg/m3/models/full_model.mdl.
12/31/2022 18:13:53 - INFO - m3inference.utils -   Checking MD5 for model full_model at /Users/wdwg/m3/models/full_model.mdl
12/31/2022 18:13:53 - INFO - m3inference.utils -   MD5s match.
12/31/2022 18:13:54 - INFO - m3inference.m3inference -   Loaded pretrained weight at /Users/wdwg/m3/models/full_model.mdl
/Users/wdwg/Desktop/scripts
12/31/2022 18:13:54 - INFO - m3inference.m3twitter -   Results not in cache. Fetching data from Twitter for barackobama.
12/31/2022 18:13:54 - INFO - m3inference.m3twitter -   GET /users/show.json?screen_name=barackobama
12/31/2022 18:13:55 - INFO - m3inference.dataset -   1 data entries loaded.
Predicting...:   0%|                                                                                                      | 0/1 [00:00<?, ?it/s][W NNPACK.cpp:64] Could not initialize NNPACK! Reason: Unsupported hardware.
Predicting...: 100%|██████████████████████████████████████████████████████████████████████████████████████████████| 1/1 [00:07<00:00,  7.58s/it]
{'input': {'description': 'Dad, husband, President, citizen.',
           'id': '813286',
           'img_path': '/Users/wdwg/m3/cache/813286_224x224.jpg',
           'lang': 'en',
           'name': 'Barack Obama',
           'screen_name': 'BarackObama'},
 'output': {'age': {'19-29': 0.0003,
                    '30-39': 0.0003,
                    '<=18': 0.0004,
                    '>=40': 0.9991},
            'gender': {'female': 0.0004, 'male': 0.9996},
            'org': {'is-org': 0.0046, 'non-org': 0.9954}}}

[computermacgyver]

Interesting. Thanks @wdwgonzales . It sounds like the segmentation fault from #26 is solved by using the latest (nightly) build and adding map_location=torch.device('mps') as an argument wherever there is torch.load(...). That is probably something we can do automatically if we detect the computer has an M1/arm64 architecture.

What computer architecture and OS are you using @giulianoformisano ? Also what version of Python?

[giulianoformisano]

Thanks a lot for your inputs! I followed @wdwgonzales, but the procedure didn't fix the issue.

I am currently using an M2, python versions 3.9.6 and 3.8.8. I also tried the same procedure on Windows Server 2012, yet getting the same error.

[MuraliRamRavipati]

downgrading torch to 1.12.1 and torchvision to 0.13.1 worked for me

[giulianoformisano]

@MuraliRamRavipati thank you so much! Downgrading torch to 1.12.1 and torchvision to 0.13.1 fixed the issue.

I created a new environment (python 3.8.8) and followed MuraliRamRavipati's suggestion.