3060 mobile graphics
Win10 64
Problem: 8 batches before would burst ram, can only be run 6 batches, today suddenly run to 32 batches, where the solution is changed, can affect the so big?
Problems related to background:
The original environment: cuda11.1 (drivers and runtime) + cudnn
Then use conda install pytorch torchvision torchaudio cudatoolkit pytorch=11.1 - c - c nvidia installation pytorch
In early may, pytorch network, also run the batchsize to six times, "efficientB0" can only be run before batchsize=6, highest can now to batchsize=32
Change:
Then install tensorflow2.5 these days, and then install cuda11.0 (runtime) + corresponding cudnn,
And found that the original pyotch program can be run to 32 batches,