It’s very difficult for me to give much/any help to those who try this on windows, you guys really need to find someone who has windows access and knows it well to be of any guidance.
As you can see from the outputs C:\Users\Andrew\Anaconda3\fastai\old needs to be in sys.path and it’s not. The quick one off hack would be to start the notebook with:
where you tell python explicitly where to find fastai. That’s if haven’t installed fastai and want to you use it directly from the clone. If you have installed it, then first check under those folders in sys.path and see whether the fastai folder is there physically. If it is then it should just work.
But the correct solution is to figure out the windows symlinks, but I’m not your man for that. You need to look here instead.
Googling the error, suggests that perhaps this is an incompatibility of pytorch versions? maybe this is relevant? make sure you are running torch-0.3.1 for fastai-0.7.x
@stas one more req. to add to the list - make sure your graphics card is up to date. I had a old Geforce GTX 560 that was on the approved CUDA list, but was no longer supported by PyTorch! Folks should be ok with any Geforce GTX 10series (1000 +) cards. I picked up a GTX 1050 and had lesson 1 run in 5 mins. Otherwise get a error like this when trying to train models:
Found GPU0 GeForce GTX 950M which is of cuda capability 5.0.
PyTorch no longer supports this GPU because it is too old.
As explained in this link…
So to wrap-up my lessons learned:
Update to windows 10 (its still a free download) from windows 7/8
Follow the instructions correctly and create a good symlink using admin privileges (mklink /d fastai …\old\fastai)
don’t use an outdated graphics card even if its on the CUDA supported list
I added the link to the CUDA-supported cards in general. I wasn’t aware that it doesn’t map 1:1 to pytorch supported cards. Do pytorch folks provide their narrowed-down list somewhere, or documented the exclusions? I
hi there, just want to share my finding when running 0.7.0 on crestle.
If you have only install fastai (0.7.0) without install torch (0.3.1) before, then it used to using torch 0.3.1 (automatically). but crestle has recently upgraded torch to 0.4.0, and most importantly, the related packages.
so you may need to install torch 0.3.1 and torchtext 0.2.3, if you didn’t install torch before.
@stas, thanks for the hint the pytorch version was correct but
“** torchtext**” version was not compatible it has to be “torchtext==0.2.3.” for pytorch 0.3.1.
Installed fastai v1 first, in a dedicated env (it seems to work). Then I installed fastai 0.7 following the procedure described (which is the usual procedure).
As I try to import fastai modules, it complains what follows:
I am using clouderizer.com Fast.ai project template for the course and to run the template I use google colab like a backend and I am not able to import the library properly. I have already posted my query here and the solution I received was for the case when we run the complete notebook on google colab itself or I may not be able to understand it because these things are new to me. If someone in the forum is using clouderizer and has found a solution, please help!.
Thanks in advance.
Edit: Finally found a solution. Click on edit project, go to the setup section, in the setup script column, add pip install torchtext==0.2.3. Now I am able to run the notebooks without any problem.
thanks stas!
tried that in kaggle kernels but got error: AttributeError: ‘torch.dtype’ object has no attribute ‘type’
but also seems somewhat random, i managed to get it work once, then it showed error, some other people also reported errors.
ML lesson 1 worked with !pip install fastai==0.7.0, however DL v2 lessons are not running on kernels reliably. the more i look forward to v3 course. noticed v3 notebooks started to appear on github repo!