Hi, I have a Windows PC with latest Nvidia GPU and wanted to go through the lessons locally. I followed all the install instructions (Miniconda, mamba, “-c pytorch -c nvidia -c fastai” etc.) for Windows and I installed fastbook using this command: mamba install -c fastai fastbook. When I execute the learner.fine_tune() method: learn = cnn_learner(dls, resnet34, metrics=error_rate) learn.fine_tune(1)
I see NaN for training and validation loss. I couldn’t find any posts about people running into this issue on fastai v2 - am I missing something here?
PS. I see pytorch 1.9 warning about my GPU.
*C:\Users\shekh\miniconda3\envs\fastbook\lib\site-packages\torch\cuda_init_.py:106: UserWarning: * GeForce RTX 3080 with CUDA capability sm_86 is not compatible with the current PyTorch installation. The current PyTorch install supports CUDA capabilities sm_37 sm_50 sm_60 sm_61 sm_70 sm_75 compute_37. If you want to use the GeForce RTX 3080 GPU with PyTorch, please check the instructions at Start Locally | PyTorch
Update: WSL-2 did not work for me. I had a hard time getting GPU acceleration even though torch.cuda.is_available() was True.
I bit the bullet and now running a dual boot setup on my PC.
Hi I just loaded everything per instructions in the thread, spent all yesterday trying all kinds of permuatioan. Finally everything loaded fine, but failure occurs at Running Your First Notebook
Could not do one pass in your dataloader, there is something wrong in it
C:\Users\alpha\miniconda3\lib\site-packages\torch\nn\functional.py:718: UserWarning: Named tensors and all their associated APIs are an experimental feature and subject to change. Please do not use them for anything important until they are released as stable. (Triggered internally at …\c10/core/TensorImpl.h:1156.) return torch.max_pool2d(input, kernel_size, stride, padding, dilation, ceil_mode)
Perhaps my s/w versions need to be different? What are the latest working versions pls?
Are you using NVIDIA 30X0 GPU? If not, please ignore my post.
Otherwise, you need to use Cuda >= 11.1 (which you are) and torch version compatible with that. See here
Sorry, I’m not sure. I’m new to both torch library and fastai- so I don’t want to mislead you.
What I know is that the torch install you do with pip/conda is most likely built with older cuda version (10.2). So, you have to specify the cudatoolkit version if you’re using more recent version (> 11.0)
I am on Windows 11 Pro Version 21H2 OS Build 22000.71 from the dev branch since WSL2 wasn’t available outside of dev branch when I started ML. My CUDA toolkit is 11.4 from running nvidia-smi which is the latest supported on my GTX1080Ti.
I have not benchmarked in a while but it used to be about 15-20% slower than running in either Windows or Linux but that is OK with me. MSFT said they’d have native CUDA support with official release of Windows 11.
I am asking around to see if anyone has free AWS/other cloud credits. Do you know if fast.ai has any such offers for the students?
It looks like you’re running into some issues with your local Windows 10 install while trying to go through the lessons. Based on the error you mentioned, it seems that your GPU may not be compatible with the current PyTorch installation. Did you check the instructions at Start Locally | PyTorch 8 as suggested in the warning message? It might help you to resolve this issue.
Also, I wanted to mention that if you’re looking for a cheap windows 10 key, there are a few options out there that you can check out. Just be sure to do your research and buy from a reputable source to avoid any issues.