My pytorch version is ‘0.2.0_4’. I was trying to release some of the memory controlled by the pytorch memory allocator by empty_cache() according to the document.
But the function torch.cuda.empty_cache() was not implemented. Is this function some legacy or a feature to be released in the future? I also realize some other functions are not there either, like get_device_capability().
Doc version master (0.4.0a0+c7cb6a7 ).
You should be looking at 0.2.0 document if you are using 0.2.0. http://pytorch.org/docs/0.2.0/
Thanks! That’s what I’m thinking. The default installation of pytorch by using conda will install 0.2.0. But I also like some of the features provided by 0.4.0. Actually I was using 0.2.0 doc, if you want to work with CUDA semantics, then the link will jump to 0.4.0 doc. Some of the functions in the doc work fine, but some of them are not there.
Is there an easy way to install 0.4.0?
Here are instructions to install from source: https://github.com/pytorch/pytorch#from-source have fun!
But a few things are currently unstable on master, you should try out the
v0.3.0 branch if you are seeing issues.
v0.3.0 has almost feature parity with master, except for some major refactoring.
Thanks! I will install that on one machine to see how it works. Do you happen to know how much behind is the conda default installation to the source version usually?
0.3 will soon be released. Check here: https://github.com/pytorch/pytorch/issues/3649.
Master is versioned at 0.4 only because the refactors. Feature-wise not much is added compared to 0.3.