NEWS.md

torch (development version)

• 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)

torch 0.8.1

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)

torch 0.8.0

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)

torch 0.7.2

Bug fix

• Fixed vignette building on Windows.

torch 0.7.1

New features

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

torch 0.7.0

Breaking changes

• torch_sort and Tensor$sort now return 1-indexed results. (#709, @mohamed-180)
• Support for LibTorch 1.10.2. See also release notes for the PyTorch v1.10. (#758, #763, #775, @hsbadr).
• Changed default dim from 1 to 2 in nnf_cosine_similarity. (#769)
• The default value for arguments of various functions have changed. A bug in the code generation was truncating the default values specially if they were float values that needed more than 6 digit precision. (#770)

New features

• jit_save_for_mobile allows to save a traced model in bytecode form, to be loaded by a LiteModuleLoader. (#713)
• Exported is_torch_tensor to check wether an object is a tensor or not. (#730, @rdinnager)
• Adds cuda_get_device_properties(device) that allows one to query device capability and other properties. (#734, @rdinnager)
• Implemented call_torch_function() to allow calling potentially unexported torch core functions. (#743, @rdinnager)
• Now when installing torch all of LibTorch and Lantern headers will be installed within the inst directory. This will allow for packages extending torch to bind directly to its C++ library. (#718)
• dataset_subset will use the .getbatch method of the wrapped dataset if one is available. (#742, @egillax)
• Added nn_flatten and nn_unflatten modules. (#773)
• Added cuda_memory_stats() and cuda_memory_summary() to verify the amount of memory torch is using from the GPU. (#774)
• Added backends_cudnn_version() to query the CuDNN version found by torch. (#774)

Bug fixes

• Fixed a bug in .validate_sample for the Distribution class that would incorrectly check for tensors. (#739, @hsbadr)
• Fixed memory leak when applying custom autograd_functions. (#750)
• Fixed a bug that caused autograd_grad to deadlock when used with custom autograd functions. (#771)
• Fixed a bug in torch_max and torch_min that would fail with length=2 Tensors. (#772)

Documentation

• Improved the 'Loading data' vignette and datasets documentation. (#780, @jnolis)

Internal

• Refactored the internal Lantern types and Rcpp types and made clearer which are the exported types that can be used in the C++ extensions. (#718)
• Simplified concurrency related constructs in autograd. (#755, @yitao-li)
• R and C++ code cleanup, styling, and formatting. (#753, @hsbadr)
• Dataloaders are slightly faster with a new transpose function. (#783)
• torch_tensor is now a C++ only function slighly increasing performance in a few situations. (#784)

torch 0.6.1

New features

• Fixed valgrind errors on CRAN by requiring a more recent version of knitr.
• Updated LibTorch to version 1.9.1 (#725 @hsbadr)
• We now check if lantern DLL's are loaded before calling any lantern function. This avoids segfaults when Lantern is not installed. (#723).

torch 0.6.0

Breaking changes

• nn_sequential is now a bare nn_module, allowing to easily inherit from it. This is a breaking change if you used the name argument. The name behavior can be achieved by subclassing; see the tests in the PR. (#699)

New features

• Additional info is showed when printing tensors like if it requires grad and the grad fn. (#668, #669, #673, @mohamed-180)
• We can now subset nn_sequential modules using [. (#678, @mohamed-180)
• We now allow padding='same' and padding='valid' when using convolutions. (#679)
• nnf_cross_entropy now uses the ATen cross_entropy operation directly instead of doing logsoftmax + NLLLoss. (#680)
• Inherited classes are now persisted by subclasses. This is specially useful if you subclass nn_sequential and still want that the specific S3 methods still work. (#701)

Bug fixes

• Fixed bug when indexing with numeric vectors. (#693, @mohamed-180)
• Fixed bug when indexing tensors with ellipsis and a tensor. (#696)

Documentation

• Improved optimizer documentation by adding a 'Warning' regarding the creation and usage order. (#698)

torch 0.5.0

Breaking changes

• Droped support for CUDA 10.1 (#610)
• torch_manual_seed() now matches PyTorch's behavior so we can more easily compare implementations. Since this is a breaking change we added the torch.old_seed_behavior=TRUE option so users can stick to the old behavior. (#639)
• Indexing with vectors has a now the same behavior as R indexing, making it easier to understand. Users can still use the old behavior by using torch_index or torch_index_put. (#649)

New features

• Added support for ScriptModule. Loaded JIT modules now operate as nn_modules. (#593)
• Added a jit_compile function that allows compiling arbitrary TorchScript code into script function that can be serialized and executed. (#601)
• Added jit_trace support for nn_module created from R. (#604)
• Updated LibTorch to version 1.9.0 (#610)
• Added Linear Algebra functions (#612)
• Added contrib_sort_vertices to efficiently sort vertices on CUDA. (#619)
• Allows querying the graph from traced modules. (#623)
• Added with_detect_anomaly to debug autograd errors. (#628)
• Implemented traced_module$graph_for() to allow inspecting the optimized jit graph. (#643)
• Added slc to allow dynamically creating slices when indexing tensors. (#648)

Bug fixes

• Fixed a bug when using a .getbatch method that didn't return a torch_tensor. (#615)
• Fixed warning when using %/% caused by a call to deprecated torch_floor_divide (#616)
• Improved CUDA version auto-detection (#644)

Internal changes

• Improved R <-> JIT types conversion. (#593)
• Added Dockerfiles for CUDA 11.1 (#597)
• A warning is raised when an incompatible dataset is passed to a parallel dataloader. (#626)
• Additionally to calling gc when CUDA memory is exhausted we now call R_RunPendingFinalizers. This should improve memory usage, because we will now delete tensors earlier. (#654)
• Fix rchk issues (#667)

torch 0.4.0

Breaking changes

• torch_multinomial now returns 1-based indexes to comply with 1-based indexing across torch. (#588)

New features

• Added parameter to multihead attention module to allow output of unaveraged attention weights. (@jonathanbratt #542)
• We now allow jit_trace functions with more than 1 argument. (#544)
• Added Multivariate normal distribution (#552)
• Export the torch_diff function and added docs for it. (#565)
• Added a device argument to torch_load() allowing one to select to which device parameters should be loaded. (#578)
• Added distr_categorical() (#576)
• Added distr_mixture_same_family() (#576)
• Improve handling of optimizers state and implement load_state_dict() and state_dict() for optimizers. (#585)
• Added the ability to save R lists containing torch_tensors using torch_save. This allows us to save the state of optimizers and modules using torch_save(). (#586)

Bug fixes

• Fixed bug in nn_multihead_attention when q,k,v inputs not all the same. (@jonathanbratt #540)
• Fixed $copy_ so it correctly respects the src requires_grad() when reloading saved models with torch_load(). (#545)
• Fixed nn_init_xavier_normal_() and nn_init_xavier_uniform_() standard deviation calculation. (#557)
• Fixed bug in torch_tensordot() when called when infering dimensions. (#563)
• Dataset's .getbatch now takes an integer vector as input instead of a list(). (#572)
• Fixed bug with tensor$size() when indexing with negative numbers. (#570)
• Fixed bug in the log_prob of distr_bernoulli() (#581)

Internal changes

• Better handling optional Tensor arguments by using an explicit XPtrTorchOptionalTensor class. (#565)
• Tensors in the R side that point to the same C++ Tensor are now guaranteed to be the same object. This allows to easily determine unique model parameters. (#582)

torch 0.3.0

Breaking changes

• 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)
• Added nn_layer_norm. (#435)
• Allow timeout=360 as install_torch() parameter for large file download (@cregouby #438)
• Added install_torch_from_file() and get_install_libs_url()for setup cases where direct download is not possible (@cregouby #439)
• Added mean.torch_tensor (#448)
• 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)
• Added normal distribution (#462)
• Added bernoulli distribution (#484)
• as.list for nn_modules (#492)
• Enumerate support in Bernoulli distribution (#490)
• Added Poisson Distriibution (#495)
• Allow optional .getbatch in datasets/dataloaders (#498)
• nn_lstm, nn_gru and nn_gru can now use cudnn accelerations when available (#503).
• Added Gamma distribution (#489)
• We now respect the TORCH_HOME env var to automatically install torch. (#522)
• Implement comparison operator != for torch dtypes. (#524)
• Added Chi-square distribution. (#518)
• Added optimizer function allowing to easily implement custom optimizers. (#527)

Bug fixes

• Fixed bug in optim_lbfgs that would make model objects exponentially big. (#431)
• Correctly handle NaNs in L-BFGS optimizer (#433)
• The default collate function now respects the data type when converting to a tensor (if the dataset returns an R object) (#434)
• Fixed torch_normal. (#450)
• 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 bug when using RNN's on the GPU (#460)
• Found and fixed some memory leaks, specially when creating datatypes from strings and when saving models with torch_save. (#454)
• Fixed bug in nnf_pad when using mode='circular'. (#471)
• Bugfixes in nn_multihead_attention (#496)
• Fixed bug when using packed sequences with nn_lstm (#500)
• Fixed bug in the to method of nn_module that would reset the requires_grad attribute of parameters. (#501)
• Added strong_wolfe option to optim_lbfgs. (#517)
• 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)
• Added a Windows GPU CI workflow (#508).
• Update to LibTorch v1.8 (#513)
• Moved some parts of the dispatcher to C++ to make it faster. (#520)

torch 0.2.1

Breaking changes

• Made torch_one_hot and nnf_one_hot use 1-based indexing. (#410)
• nn_module$eval() and nn_module$train() now return a callable nn_module instead of a nn_Module. (#425)

New features

• Added a custom CPU allocator to call gc when torch might need more memory (#402)
• Updated to LibTorch 1.7.1 (#412)
• Allow listing all nested modules in a nn_module (#417)
• Allow modifying the requires_grad attribute using the $<- operator (#419)
• Added length method for the nn_sequential container. (#423)
• Added support for CUDA 11 on linux (#424)

Bug fixes

• Fix support for cuda 9.2 (#398)
• Fixed GPU CI that was skipping tests. (#398)
• Fixed a memory leak when printing tensors (#402)
• Fixed a memory leak when passing integer vectors to lantern. (#402)
• Fixed a few more memory leaks related to autograd context (#405)
• Fixed nnf_normalize and x$norm() as they were not able to be called (#409)

Documentation

• Small improvement to nn_module documentation (#399).
• The getting started section has been removed from the pkgdown website in favor of the new guide in the landing page (#401)
• Updated the landing page to include a getting started tutorial (#400)

torch 0.2.0

Breaking changes

• Dataloaders now returns a coro::exhausted intead of raising stop_iteration_error when the dataloader exceeds. (#366)
• Fixed bug that would happen with functions that need to transform tensors from 0-based to 1-based in the GPU. (#317)
• Fixed torch_argsort and x$argsort to return 1-based indexes (#342)
• Fixed torch_argmax, torch_argmin, x$argmax() and x$argmin() return 1-based indexes. (#389)

New features

• Added $element_size() method (@dirkschumacher #322)
• Added $bool() method (@dirkschumacher #323)
• torch__addr and torch__addr_ have been removed as they are no longer available in LibTorch 1.7.
• We now check the MD5 hashes of downloaded LibTorch binaries. (@dirkschumacher #325)
• Added a Distribution abstract class (@krzjoa #333)
• Updated to LibTorch 1.7 (#337)
• We now warn when converting long tensors to R and there's a chance of an integer overflow. (#347)
• Allow private and active methods in nn_module's and dataset's. (#349)
• Added nn_batch_norm3d (@mattwarkentin #354)
• Added nn_lstm and nn_gru modules. (#362)
• Added distribution constraints (@krzjoa #364)
• Dataloaders now use the num_workers argument to load data in parallel (#366)
• Added Exponential Family classs to distributions (#373)
• Added Dockerfile and docker compose file with GPU support, with a how-to guide. (#380 #386)
• Added R 3.6 to the CI system and fixed compilation from source with it on Windows (#387)
• Initial support for JIT tracing (#377)
• Added LBFGS optimizer (#392)
• Improved the nn_module UI by improving autocomplete support and adding a print method (#391)

Bug fixes

• Fixed bug when trying to print the grad_fn of a Tensor that doesn't have one. See (#321)
• Refactored the optimizers code to avoid duplication of parameter checks, etc. (@dirkschumacher #328)
• Fixed torch_norm so it can be called with a dim argument. (#345)
• Fixed crash when calling torch_hann_window with an invalid NULL window_length. (#351)
• Fixed torch_stft calls for LibTorch 1.7 (added the return_complex argument) (#355)
• Fixed bug when strides were NULL in some pooling operations. (#361)
• Use nvcc --version instead of nvidia-smi to find the CUDA version as nvidia-smi reports the latest supported version and not the installed one. (#363)
• Corrected URL to download LibTorch under Linux with CUDA 10.2 (#367)
• Fixed handling of integer tensors when indexing tensors (#385)
• Fixed bug when passing length zero vectors to lantern/libtorch. (#388)

torch 0.1.1

Bug fixes

• Fixed bug that made RandomSampler(replacement = TRUE) to never take the last element in the dataset. (84861fa)
• Fixed torch_topk and x$topk so the returned indexes are 1-based (#280)
• Fixed a bug (#275) that would cause 1 - torch_tensor(1, device = "cuda") to fail because 1 was created in the CPU. (#279)
• We now preserve names in the dataloader output (#286)
• torch_narrow, Tensor$narrow() and Tensor$narrow_copy are now indexed starting at 1. (#294)
• Tensor$is_leaf is now an active method. (#295)
• Fixed bug when passing equations to torch_einsum. (#296)
• Fixed nn_module_list() to correctly name added modules, otherwise they are not returned when doing state_dict() on it. (#300)
• Fixed bug related to random number seeds when using in-place methods. (#303)
• Fixed nn_batchnorm* so it returns the same results as PyTorch (#302)
• Fixed a bug that made nn_module$parameter when there were shared parameters between layers. (#306)
• Fixed $max and $min to return 1-based indexes. (#315)

New features

• Expanded the utils_data_default_collate to support converting R objects to torch tensors when needed. (#269)
• Added an as.matrix method for torch Tensors. (#282)
• By default we now truncate the output of print(totrch_tensor(1:40)) if it spans for more than 30 lines. This is useful for not spamming the console or taking very long to print when you print a very large tensor. (#283)
• Added the Adadelta optimizer (@krzjoa #284)
• Added support for GPU's on Windows (#281)
• Added the Adagrad optimizer (@krzjoa #289)
• Added RMSprop optimizer (@krzjoa #290)
• Added the Rprop optimizer (@krzjoa #297)
• Added gradient clipping utilities (#299)
• Added nnf_contrib_sparsemax and nn_contrib_sparsemax. (#309)
• Added ASGD optimizer (@krzjoa #307)
• Getters and setters for the number of threads used by torch (#311)

torch 0.1.0

• Added many missing losses (#252)
• Implemented the $<- and [[<- operators for the nn_module class. (#253)
• Export nn_parameter, nn_buffer, and is_* auxiliary functions.
• Added a new serialization vignette.
• Added a few learning rate schedulers (#258)

torch 0.0.2

• Added a NEWS.md file to track changes to the package.
• Auto install when loading the package for the first time.