Cannot import name 'QuantStub' from 'torch.ao.quantization'

fastai
1

I was able to creaet a fastai 2.5.x / Python 3.9 conda env using:

  • conda env create -n fastbook -c defaults python=3.9
  • conda config --env --add channels defaults
  • conda config --env --add channels pytorch
  • conda config --env --add channels fastai
  • mamba install ipywidgets matplotlib pandas scikit-learn sentencepiece
  • pip install -r requirements.txt

The last command downgraded the mamba-installed torch 1.11 to 1.10, in case this matters.

When trying to import fastbook I get the following error:

----> 1 import fastbook

File ~/miniconda3/envs/fastbook/lib/python3.9/site-packages/fastbook/__init__.py:3, in <module>
      1 __version__ = "0.0.18"
      2 import matplotlib as mpl, pkgutil, requests, time
----> 3 from fastai.vision.all import *
      4 from fastdownload import download_url
      5 from pandas.api.types import CategoricalDtype

File ~/miniconda3/envs/fastbook/lib/python3.9/site-packages/fastai/vision/all.py:1, in <module>
----> 1 from . import models
      2 from ..basics import *
      3 from ..callback.all import *

File ~/miniconda3/envs/fastbook/lib/python3.9/site-packages/fastai/vision/models/__init__.py:1, in <module>
----> 1 from . import xresnet
      2 from . import unet
      3 from .tvm import *

File ~/miniconda3/envs/fastbook/lib/python3.9/site-packages/fastai/vision/models/xresnet.py:13, in <module>
     11 # Cell
     12 from ...torch_basics import *
---> 13 try: from torchvision.models.utils import load_state_dict_from_url
     14 except ModuleNotFoundError: from torch.hub import load_state_dict_from_url
     16 # Cell

File ~/miniconda3/envs/fastbook/lib/python3.9/site-packages/torchvision/__init__.py:7, in <module>
      5 from torchvision import datasets
      6 from torchvision import io
----> 7 from torchvision import models
      8 from torchvision import ops
      9 from torchvision import transforms

File ~/miniconda3/envs/fastbook/lib/python3.9/site-packages/torchvision/models/__init__.py:18, in <module>
     16 from . import feature_extraction
     17 from . import optical_flow
---> 18 from . import quantization
     19 from . import segmentation
     20 from . import video

File ~/miniconda3/envs/fastbook/lib/python3.9/site-packages/torchvision/models/quantization/__init__.py:1, in <module>
----> 1 from .mobilenet import *
      2 from .resnet import *
      3 from .googlenet import *

File ~/miniconda3/envs/fastbook/lib/python3.9/site-packages/torchvision/models/quantization/mobilenet.py:1, in <module>
----> 1 from .mobilenetv2 import QuantizableMobileNetV2, mobilenet_v2, __all__ as mv2_all
      2 from .mobilenetv3 import QuantizableMobileNetV3, mobilenet_v3_large, __all__ as mv3_all
      4 __all__ = mv2_all + mv3_all

File ~/miniconda3/envs/fastbook/lib/python3.9/site-packages/torchvision/models/quantization/mobilenetv2.py:5, in <module>
      3 from torch import Tensor
      4 from torch import nn
----> 5 from torch.ao.quantization import QuantStub, DeQuantStub
      6 from torchvision.models.mobilenetv2 import InvertedResidual, MobileNetV2, model_urls
      8 from ..._internally_replaced_utils import load_state_dict_from_url

ImportError: cannot import name 'QuantStub' from 'torch.ao.quantization' (/home/maye/miniconda3/envs/fastbook/lib/python3.9/site-packages/torch/ao/quantization/__init__.py)

Anybody seen this and got around it?

2

what’s your torchvision version? perhaps need update or downgrade it…

3

Same version, but pip removed pytorch, weirdly…
Some funny stuff is happening there, with different channels and dependency breakages between pip and conda.
I’ll try to summarize what I currently have found is a complete set of instructions to make import fastbook work:

  1. conda create -n fastbook -c defaults python=3.9 -y
  2. conda activate fastbook
    2a: conda config --env --add channels defaults (if not yet set, my default is conda-forge)
    2b: conda config --env --add channels pytorch
    2c: conda config --env --add channels fastai
    (don’t add fastchan as a channel, and don’t install from there, it leads to mamba not finding dependencies for everything)
  3. mamba install pytorch torchvision ipywidgets matplotlib pandas scikit-learn sentencepiece -y
  4. pip install fastai graphviz azure-cognitiveservices-search-imagesearch
    (this has removed pytorch which will lead to above error, so it needs to be back-installed:)
  5. mamba install pytorch -y

and finally, import fastbook succeeds. I could run the first learning example from fastbook with that (around 1 minute on my Linux PC with NVIDIA GPU).

1 Like