RuntimeError: no valid convolution algorithms available in CuDNN

In one belt GPU Of Windows Run an open source project on , Install the dependent libraries in turn 、 Solve the problem of insufficient virtual memory space 、 After solving the problem of memory out of bounds , Thought it was going to be finished , I didn't expect another fatal problem ：

Baidu searched , Many people say that it is caused by insufficient video memory . So when running the program, it passes nvidia-smi Check the usage of video memory , It is found that the existing occupancy has been increasing before the error occurs , Until it is basically close to the upper limit of video memory ：

So move the same program to Linux Server ( Single graphics card video memory 24G) Configure to run , yes OK Of . 

So to solve this problem, in addition to changing the size of the video memory and graphics card , Another idea is to reduce the data sent to video memory in a single time . actually , In the above problematic code , Have already put wokers The quantity is set to 0, also batch_size It's also 1, So maybe something else went wrong . It is still positioning .