v0.13.0

Breaking changes

• lantern is now distributed over a different URL (https://torch-cdn.mlverse.org).
  For most users this shouldn't have any effect, unless you need special authorization
  to access some URL's. (#1162)

New features

• Added support for a private $finalize_deep_clone() method for nn_module which
  allows to run some code after cloning a module.
• A compare_proxy method for the torch_tensor type was added
  it allows to compare torch tensors using testthat::expect_equal().
• Converting torch tensor to R array works when tensor has 'cuda' device (#1130)

Bug fixes

• Fix a bug on using input projection initialization bias in nnf_multi_head_attention_forward (#1154 @cregouby)
• Bugfix: calling $detach() on a tensor now preserves attributes (#1136)
• Make sure deep cloning of tensor and nn_module preserves class attributes and the requires_grad field. (#1129)
• Fixed that parameters and buffers of children of nn_modules were not cloned
• Cloned objects no longer reference the object from which they were cloned
• Fixed bug where nn_module's patched clone method was invalid after a call to
  the internal create_nn_module_callable()
• Printing of grad_fn now appends a new line at the end.
• Make sure deep cloning preserve state dict attributes. (#1129)
• Added separate setter and unsetter for the autocast context instead of only allowing local_autocast(). (#1142)
• Fixed a bug in torch_arange() causing it to return 1:(n-1) values when specific request dtype = torch_int64() (#1160)

v0.12.0

Breaking changes

• New torch_save serialization format. It's ~10x faster and since it's based on safetensors, files can be read with any safetensors implementation. (#1071)
• Updated to LibTorch 2.0.1. (#1085)
• torch_load no longer supports device=NULL to load weights in the same device they were saved. (#1085)
• Lantern binaries and torch pre-built binaries are now built on Ubuntu 20.04. (#1124)

New features

• Added support for CUDA 11.8. (#1089)
• Added support for iterable datasets. (#1095)

Bug fixes

• fix printer of torch device (add new line at the end)
• as.array now moves tensors to the cpu before copying data into R. (#1080)
• Fixed segfault caused by comparing a dtype with a NULL. (#1090)
• Fixed incorrect naming of complex data type names, such as torch_cfloat64. (#1091)
• Fixed name of the out_features attribute in the nn_linear module. (#1097)
• Fixed issues when loading the state dict of optimizers and learning rate schedulers. (#1100)
• Fixed bug when cloning nn_modules with empty state dicts. (#1108)
• distr_multivariate_normal now correctly handles precision matrix's. (#1110)
• Moved length.torch_tensor implementation to R7 to avoid problems when a torch dataset has the torch_tensor class. (#1111)
• Fixed problem when deep cloning a nn_module. (#1123)

v0.11.0

Breaking changes

• load_state_dict() for optimizers now default to cloning the tensors in the state dict, so they don't keep references to objects in the dict. (#1041)

New features

• Added nn_utils_weight_norm (#1025)
• Added support for reading from ordered state dicts serialized with PyTorch. (#1031)
• Added jit_ops allowing to access JIT operators. (#1023)
• Added with_device and local_device to allow temporarily modify the default device tensors get initialized. (#1034)
• nnf_gelu() and nn_gelu() gained the approximate argument. (#1043)
• Implemented != for torch devices. (#1042)
• Allows setting the dtype with a string. (#1045)
• You can now create a named list of modules using nn_module_dict(). (#1046)
• Faster load_state_dict(), also using less memory. It' possible to use the legacy implementation if required, see PR. (#1051)
• Export helpers for handling RNG state, and temprarily modifying it. (#1057)
• Added support for converting half tensors into R with as.numeric(). (#1056)
• Added new torch_tensor_from_buffer() and buffer_from_torch_tensor() that allow low level creation of torch tensors. (#1061, #1062)

Documentation

• Improved documentation for LBFGS optimizer. (#1035)
• Added a message asking the user to restart the session after a manual installation with install_torch(). (#1055)

Bug fixes

• Fixed bug related to handling of non-persistent buffers. They would get added to the state_dict() even if they should not. (#1036)
• Fixed a typo in the optim_adamw class name.
• Fixed nn_cross_entropy_loss class name. (#1043)
• Fixed bug in LBFGS w/ line search. (#1048)
• Correctly handle the installation when RemoteSha is a package version. (#1058)

Internal

• Started building LibLantern on macOS 11 instead of macOS12 for maximum compatibility. (#1026)
• Added CXXSTD to Makevars to enable C+11 compilation options.
• Refactored codepath for TensorOptions and now all tensors initialization are handled by the same codepath. (#1033)
• Added internal argument .refer_to_state_dict to the load_state_dict() nn_module() method. Allows loading the state dict into the model keeping parmaters as references to that state dict. (#1036)

v0.10.0

Breaking changes

• Updated to LibTorch v1.13.1 (#977)

New features

• Provide pre-built binaries for torch using a GH Action workflow. (#975)
• Added nn_silu() and nnf_silu(). (#985)
• Added support for deep cloning nn_modules. (#986)
• Added local_no_grad() and local_enable_grad() as alternatives for the with_ functions. (#990)
• Added optim_adamw optimizer. (#991)
• Added support for automatic mixed precision (#996)
• Added functionality to temporarily modify the torch seed. (#999)
• Support for creating torch tensors from raw vectors and back. (#1003)

Bug fixes

• Dataloaders now preserve the batch dimension when batch_size=1 is used. (#994)

Internal

• Large refactoring of the build system. (#964)
• Use native symbol registration instead of dynamic lookup. (#976)
• Returning lists of tensors to R is now much faster. (#993)

v0.9.1

Breaking changes

• torch_where now returns 1-based indices when it's called with the condition argument only. (#951, @skeydan)

New features

• Added support for nonABI builds on CUDA 11.6. (#919)
• The torch_fft_fftfreq() function is now exported. (#950, @skeydan)

Bug fixes

• Fixed bug that caused distr_normal$sample() not being able to generate reproducible results after setting seeds. (#938)
• torch_cat error message now correctly reflects 1-based indexing. (#952, @skeydan)

Internal

• Fixed warnings in R CMD Check generated by the unsafe use of sprintf. (#959, @shikokuchuo)
• Import, not suggest glue (#960)

v0.9.0

Breaking changes

• Updated to LibTorch v1.12.1. (#889, #893, #899)
• torch_bincount is now 1-based indexed. (#896)
• torch_movedim() and $movedim() are now both 1-based indexed. (#905)

New features

• Added cuda_synchronize() to allow synchronization of CUDA operations. (#887)
• Added support for M1 Macs, including creating Tensors in the MPS device. (#890)
• Added support for CUDA 11.6 on Linux. (#902)
• Added cuda_empty_cache() to allow freeing memory from the caching allocator to the system. (#903)
• Added $is_sparse() method to check wether a Tensor is sparse or not. (#903)
• dataset_subset now adds a class to the modified dataset that is the same as the original dataset classes postfixed with _subset. (#904)
• Added torch_serialize() to allow creating a raw vector from torch objects. (#908)

Bug fixes

• Fixed bug in torch_arange that was causing the end value not getting included in the result. (#885, @skeydan)
• Fixed bug in window functions by setting a default dtype. (#886, @skeydan)
• Fixed bug when using install_torch(reinstall = TRUE). (#883)
• The dims argument in torch_tile() is no longer modified, as it's not meant to be the a 1-based dimension. (#905)
• nn_module$state_diict() now detaches output tensors by default. (#916)

Internal

• Re-implemented the $ method for R7 classes in C/C++ to improve speed when calling methods. (#873)
• Re-implemented garbage collection logic when calling it from inside a backward() call. This improves speed because we no longer need to call GC
  everytime backward is called. (#873)
• We now use a thread pool instead of launching a new thread for backward calls. (#883)
• Implemented options to allow configuring the activation of garbage collection when allocating more CUDA memory. (#883)
• Some nnf_ functions have been updated to use a single torch_ kernel instead of the custom implementation. (#896)
• Improved performance of dataloaders. (#900)
• We now let LibTorch query the default generator, this allows one to use torch_bernoulli() with device="gpu". (#906)

v0.8.1

Breaking changes

• We now prompt the user before installing torch additional dependencies in interactive environments. This was requested by CRAN maintainers. (#864)

New features

• Dataloaders can now handle logical values. (#858, @ryan-heslin)
• We now provide builds for Pre CXX11 ABI version of LibTorch. They can be used by setting the environment variable PRECXX11ABI=1. This can be useful in environments with older versions of GLIBC. (#870)

Bug fixes

• Fixed the way errors are passed from dataloaders workers to the main process. Now using new rlang error chaining. (#864)

Internal

• We can now call GC even if from a backward call (ie, from a different thread) which allows for better memory management. (#853)
• Fix HTML5 Manual information as resquested by CRAN (#869)

v0.8.0

Breaking changes

• Serialization is now much faster because we avoid base64 encoding the serialized tensors. As a result, files serialized with newer versions of torch can't be opened with older versions of torch. Set options(torch.serialization_version = 1) if you want your file to be readable by older versions. (#803)
• Deprecated support for CUDA 10.2 on Windows. (#835)
• linalg_matrix_rank and linalg_pinv gained atol and rtol arguments while deprecating tol and rcond. (#835)

New features

• Improved auto-detection of CUDA version on Windows. (#798, @SvenVw)
• Improved parallel dataloaders performance by using a socket conection to transfer data between workers and the main process. (#803)
• keep_graph now defaults to the value of create_graph when calling $backward(). We also renamed it to retain_graph to match PyTorch. (#811)
• Optimizers created with optimizer now carry the classname in the generator and in instances. Optimizer generators now have the class torch_optimizer_generator. The class of torch optimizers has been renamed from torch_Optimizer to torch_optimizer. (#814)
• New utility function nn_prune_head() to prune top layer(s) of a network (#819 @cregouby)
• torch_kron() is now exported (#818).
• Added nn_embedding_bag. (#827, @egillax)
• nn_multihead_attention now supports the batch_first option. (#828, @jonthegeek)
• It's now possible to modify the gradient of a tensor using the syntax x$grad <- new_grad. (#832)
• sampler() is now exported allowing to create custom samplers that can be passed to dataloader(). (#833)
• Creating nn_modules without a initialize method is now supported. (#834)
• Added lr_reduce_on_plateau learning rate scheduler. (#836, @egillax)
• torch_tensor(NULL) no longer fails. It now returns a tensor with no dimensions and no data. (#839)
• Improved complex numbers handling, including better printing and support for casting from and to R. (#844)

Bug fixes

• Fixed bug in weight decay handling in the Adam optimizer. (#824, @egillax)
• Fixed bug in nn_l1_loss. (#825, @sebffischer)

Documentation

• Nice error message when embed_dim is not divisible by num_heads in nn_multihead_attention. (#828)

Internal

• Updated to LibTorch v1.11.0. (#835)
• Moved error message translations into R, this makes easier to add new ones and update the existing. (#841)

v0.7.2

Bug fix

• Fixed vignette building on Windows.

v0.7.1

New features

• Added cuda_runtime_version() to query the CUDA Tolkit version that torch is using. (#790)