Is there a pytorch implementation for projected SGD optimizer? Specifically, the algorithm is similar to SGD except there is an additional step where the updated parameters (w_{t+1}) are projected to a hyper-sphere with a specified radius R (||w|| < R).