What is volatile variable in Pytorch

For versions of Pytorch previous to 0.4.0, Variable and Tensor were two different entities. For variables, you could specify two flags: volatile and require_grad. Both of them were used for fine grained exclusion of subgraphs from gradient computation.

The difference between volatile and requires_grad is in how the flag is applied to the outputs of an operation. If there is even a single volatile = True Variable as input to an operation, its output is also going to be marked as volatile. For requires_grad, you need all the inputs to that operation to be flagged requires_grad = False, so that the output is also flagged in the same way.

From Pytorch 0.4.0, Tensors and Variables have merged, and the volatile flag is deprecated.

Basically, set the input to a network to volatile if you are doing inference only and won't be running backpropagation in order to conserve memory.

From the docs:

Volatile is recommended for purely inference mode, when you’re sure you won’t be even calling .backward(). It’s more efficient than any other autograd setting - it will use the absolute minimal amount of memory to evaluate the model. volatile also determines that requires_grad is False.

Edit: The volatile keyword has been deprecated as of pytorch version 0.4.0