pytorch with torch.no_ grad

【pytorch series 】 with torch.no_grad(): Usage details _ Great black mountain monastic blog -CSDN Blog _torch.no_grad(): stay pytorch Written in the network ,with torch.no_grad(): Very common . First , About python Medium with：with Statement is suitable for accessing resources , Make sure that the necessary... Is executed regardless of whether an exception occurs during use “ clear ” operation , Release resources , For example, the file is automatically closed after use ／ Automatic acquisition and release of locks in threads . for example ：file = open("１.txt")data = file.read()file.close() The problems are as follows ：（１） An exception occurred while reading the file , But nothing was done ;（２） May forget to close the file handle ; improvement ：try: https://blog.csdn.net/sazass/article/details/116668755

First , About python Medium with：

with Statement is suitable for accessing resources , Make sure that the necessary... Is executed regardless of whether an exception occurs during use “ clear ” operation , Release resources , For example, the file is automatically closed after use ／ Automatic acquisition and release of locks in threads .

About torch.no_grad()
First of all, from the requires_grad Speak up ：

requires_grad
stay pytorch in ,tensor There is one requires_grad Parameters , If set to True, In the case of reverse propagation , The tensor Will automatically take the derivative .tensor Of requires_grad The default attribute of is False, If a node （ Leaf variables ： self-created tensor）requires_grad Set to True, Then all nodes that depend on it requires_grad All for True（ Even if other dependent tensor Of requires_grad = False）

When requires_grad Set to False when , Back propagation does not automatically take the derivative , Therefore, it greatly saves video memory or memory .

with torch.no_grad The role of
Under this module , All calculated tensor Of requires_grad Are automatically set to False.

Even if one tensor（ Name it x） Of requires_grad = True, stay with torch.no_grad Calculation , from x Get new tensor（ Name it w- Scalar ）requires_grad Also for the False, And grad_fn Also for the None, That's not right w Derivation .