You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
torch_nonzero and tensor$nonzero() now return 1-based indexes. (#432)
Breaking change: torch_arange returns in the closed interval [start, end] instead of the half open [start, end). This makes it behave similar to R's seq. (#506)
New features
torch_split now accepts a list of sizes as well as a fixed size. (#429)
New arguments worker_globals and worker_packages allowing to easily pass objects to workers in parallel dataloaders (#449).
We now call R garbage collector when there's no memory available on GPU, this can help in a few cases when the laziness of the garbage collector allows too many tensors to be on memory even though they are no longer referenced in R. (#456)
Implemented nn_group_norm and fixed a bug in nnf_group_norm (#474)
Added backend functions allowing us to query which optimizations LibTorch was compiled with (#476)
Fixed backward compatibility issue when loading models saved in older versions of torch. This bug was introduced in #452 and is now fixed and we also added a regression test. (#458)
Fixed default argument of nn_init_trunc_normal_ initializer function. (#535)
Documentation
Added vignette on reading models from Python (#469)
Internal changes
Removed the PerformanceReporter from tests to get easier to read stack traces. (#449)
Internal change in the R7 classes so R7 objects are simple external pointer instead of environments. This might cause breaking change if you relied on saving any kind of state in the Tensor object. (#452)
Internal refactoring making Rcpp aware of some XPtrTorch* types so making it simpler to return them from Rcpp code. This might cause a breaking change if you are relying on torch_dtype() being an R6 class. (#451)
Internal changes to auto unwrap arguments from SEXP's in Rcpp. This will make easier to move the dispatcher system to C++ in the future, but already allows us to gain ~30% speedups in small operations. (#454)