I’ve been trying to start the mit deep learning lab1. I know it comes with the notebook you can run in google colab, but I’d like to be able to play with this locally and hopefully experiment easier. I had a ton of problems getting this to work, but I finally did it. Here are some notes.
- Install python 3.8 from the python website not from windows app marketplace. I was finding that the windows app one seems to be sandboxed and cannot fully find stuff in PATH.
- I grabbed tensorflow nightly as I could not get the default to work at this time.
- I used cuda 11.0 update 2. (This is the one I have working other cuda 11s might work). This was ” cuda_11.0.3_451.82_win10.exe ”
- I used cuDNN ” cudnn-11.0-windows-x64-v184.108.40.206.zip ” use ” cudnn-11.2-windows-x64-v220.127.116.11.zip ” now. Had a version error later during running a NN but not during install.
- nVidia cuDNN check that path is to do the bin folder. Settings -> System -> About -> Advanced System Settings. Advanced tab -> Environment variables.To compare my PATH looks like:C:toolscudabin (this should be the only one you need to manually add)C:Program FilesNVIDIA GPU Computing ToolkitCUDAv11.0binC:Program FilesNVIDIA GPU Computing ToolkitCUDAv11.0extrasCUPTIlib64C:Program FilesNVIDIA GPU Computing ToolkitCUDAv11.0includeC:Program FilesNVIDIA GPU Computing ToolkitCUDAv11.0libnvvpC:Program FilesNVIDIA CorporationNsight Compute 2020.1.2
- If you do all of this it still wont work, but pay attention to the errors! there is a dll in C:Program FilesNVIDIA GPU Computing ToolkitCUDAv11.0bin that it cannot find.
- This dll is infact xyz_10.dll in the folder but python wants to grab xyz_11.dll. I renamed this to _11.dll and all works. Note xyz is some random name as I for get the exact file name now, but there was only one.
Also I am running this on a non AVX supported cpu (X58 era W3690) with at GTX 1060 6GB. Hope this helps I found a lot of similar posts in various sites with abandonment or no answers and almost gave up. Reposted this as I forgot to join before submitting and this was marked as spam.