ModuleNotFoundError: No module named 'sparseml.pytorch.utils.quantization' about sparseml HOT 10 CLOSED

I'm still waiting for help! pls don't forget me :)

from sparseml.

Hi @bfineran
after running your suggested command, error not arise again, but during processing, its show this massage:
torch~=1.7.0 not found and is required by YOLOv5, attempting auto-update...
and then back to torch 1.7 again!(download and install it) and if I run my command again, previous error come back!! :(
can you solve this inconsistency?

also, it is said that cant find images directory for training!
( I have worked with YOLOv5 earlier, and I do thing as it want, i.e creating folder next to yolo root contain image and label for train and test, and adresing in data.yaml)

from sparseml.

Hi @RasoulZamani, we're looking into this now to see if we can replicate it. Will have a response by the end of the day.

Thanks,
Mark

from sparseml.

hi @RasoulZamani this issue was due to setup_integration.py pointing to an old branch of the yolov5 repo. The repo has been updated as well as the setup script. Could you try re pulling the repo and checking out release/0.11 and re-running?

(in yolov5)

git checkout master
git pull
git checkout release/0.11

from sparseml.

Hi dear @bfineran
thanks for your guidance. I do what you said and error disappear - ! , but another error arise :( :

CUDA error: no kernel image is available for execution on the device

when I google this new error , people suggest me :
pip install --pre torch torchvision -f https://download.pytorch.org/whl/nightly/cu110/torch_nightly.html -U
but when I run this, another error arise:
ImportError: required max torch version 1.9.100, found 1.11.0

so now what can I do?

from sparseml.

Hi @RasoulZamani can you try the following command to install a version of torch 1.9.1 compiled for cuda 11?

pip install torch==1.9.1+cu111 -f https://download.pytorch.org/whl/cu111/torch_stable.html

from sparseml.

Hi @RasoulZamani, unfortunately, we are currently hard coding the torch version to 1.7 in the yolov5 integration since that is the last official one that has been tested with that flow. Additionally, PyTorch has not compiled a cuda 11 compatible build for version 1.7. Until we're able to upgrade the flows and test (in process right now and we hope to have it completed in the next two weeks), the safest option is downgrading the cuda version to 10 to enable PyTorch 1.7.

An alternative, experimental pathway, for now, would be to edit the requirements.txt file on your local setup to checkout torch 1.9 instead of 1.7. The risk here is whether or not the quantization flows will continue to work correctly. If you run into any issues with this please let us know and we'll be happy to jump on and support.

Thanks,
Mark

from sparseml.

thanks @markurtz and @bfineran for your answers.

from sparseml.

Hi @RasoulZamani

As it has been a few weeks with no further conversation, I am going to go ahead and close this comment. Please re-open if you have a follow-up. Also, I invite you to "star" our SparseML repo if you like and have not already! We always like seeing community support!
https://github.com/neuralmagic/sparseml/

Thank you!
Jeannie / Neural Magic

from sparseml.

pip install torch==1.7.1+cu110 torchvision==0.8.2+cu110 torchaudio==0.7.2 -f https://download.pytorch.org/whl/torch_stable.html

from sparseml.