torch.no_grad() สำหรับการอนุมาน

You only need gradients while training. When you just want predictions, tracking the graph is wasted work, so PyTorch lets you turn it off. 🛑

Wrap code in a torch.no_grad() block and autograd stops recording inside it. No graph is built, and no gradients are stored.

Building the graph costs memory to remember every step for a backward pass that will never come. During inference that overhead is pure waste.

Inside no_grad your forward pass uses less memory and runs a little faster. On big models this lets you fit larger batches when only predicting.

Tensors created inside the block have requires_grad false. They are plain results you can print, save, or turn into NumPy without complaint.

Validation and test loops should always run under no_grad. You are scoring the model, not training it, so there is no reason to track gradients.

For inference, set model.eval() and wrap calls in no_grad together. eval fixes layers like dropout; no_grad stops the graph. They solve different problems.

no_grad only affects code inside the with block. Once you leave it, autograd switches tracking back on automatically for your next training step.

You can also tag a whole function with @torch.no_grad(). Every tensor op inside that function then runs without building a graph.

Need to free just one tensor from the graph instead of a whole block? Call .detach() on it to get a copy that carries no gradient history.

Reach for no_grad any time you are not learning: inference, metrics, or saving outputs. It is a tiny change that quietly saves memory and speed.

Wrap inference in torch.no_grad() to skip graph building, saving memory and time. Pair it with model.eval() whenever you predict instead of train. 🚀