weixin_45330042的博客

原创 几行代码让你搞懂torch.no_grad

一篇讲得比较好的文章首先明确一点，no_grad与detach有异曲同工之妙，都是逃避autograd的追踪。接下来我们做个实验：a = torch.ones(2,requires_grad=True)b = a*2print(a, a.grad, a.requires_grad )b.sum().backward(retain_graph = True )print(a, a.grad, a.requires_grad )with torch.no_grad(): a = a .