Policy Gradient with PyTorch