Conda - ModuleNotFoundError: No module named 'torch'

Question

Steps to reproduce:

I am using Anaconda on Windows to set up environment for this repo.

conda create --name pytorch-yolo

Then I install all dependencies with conda install --file requirements.txt

Which returns

PackagesNotFoundError: The following packages are not available from current channels:
- torch[version='>=1.2']

So I install pytorch with conda install pytorch torchvision cudatoolkit=10.2 -c pytorch

When I am trying to evaluate the model from the following repo: https://github.com/eriklindernoren/PyTorch-YOLOv3 with python3 test.py --weights_path weights/yolov3.weights inside conda environment, it returns the following error ModuleNotFoundError: No module named 'torch'

If I run a Jupyter lab or notebook, or even go with python inside conda terminal, torch is easily imported.

What I've already tried:

Almost everything from this thread: No module named "Torch"

Specifically, creating a new conda environment with python set to 3.8.2, installing torch to base and then to the environment.

I'm also using pyenv to set global python to 3.8.2, but that doesn't help at all.

conda list shows I have pytorch installed

Can't wrap my head around this issue.

Answer 1

You are probably using the wrong python binary. Can you try python test.py --weights_path weights/yolov3.weights?

I am not familiar with Windows terminal, but you can get the path to the binaries by using the where command (which for Linux):

(pytorch-yolo) C:\Users\RemiChauvenne>where python3
C:\Users\RemiChauvenne\AppData\Local\Microsoft\WindowsApps\python3.exe

(pytorch-yolo) C:\Users\RemiChauvenne>where python
C:\Users\RemiChauvenne\miniconda3\envs\pytorch-yolo\python.exe
C:\Users\RemiChauvenne\AppData\Local\Microsoft\WindowsApps\python.exe

We can see that python3 does not go to the python binary inside the conda environment, whereas python is correctly linked to pytorch-yolo.