NEWS

mlr3torch 0.3.3 (2026-01-31)

Feat: Improve lazy_tensor printing.
Fix: Improve consistency in as_lazy_tensor() when converting 1D tensors to lazy tensors.
Various minor bug fixes

t_opt("adamw") now actually uses AdamW and not Adam.
Caching: Cache directory is now created, even if its parent directory does not exist.
Add mlr3torch to mlr_reflections$loaded_packages to fix errors when using mlr3torch in parallel.

FT Transformer can now be (un-)marshaled after being trained on categorical data (#412).
Parameters (batch)-sampler now work (#420, thanks @tdhock)

The output dimension of neural networks for binary classification tasks is now expected to be 1 and not 2 as before. The behavior of nn("head") was also changed to match this. This means that for binary classification tasks, t_loss("cross_entropy") now generates nn_bce_with_logits_loss instead of nn_cross_entropy_loss. This also came with a reparametrization of the t_loss("cross_entropy") loss (thanks to @tdhock, #374).

Added po("nn_identity")
Added po("nn_fn") for calling custom functions in a network.
Added the FT Transformer model for tabular data.
Added encoders for numericals and categoricals
nn("block") (which allows to repeat the same network segment multiple times) now has an extra argument trafo, which allows to modify the parameter values per layer.

The context for callbacks now includes the network prediction (y_hat).
The lr_one_cycle callback now infers the total number of steps.
Progress callback got argument digits for controlling the precision with which validation/training scores are logged.

TorchIngressToken now also can take a Selector as argument features.
Added function lazy_shape() to get the shape of a lazy tensor.
Better error messages for MLP and TabResNet learners.
TabResNet learner now supports lazy tensors.
The LearnerTorch base class now supports the private method $.ingress_tokens(task, param_vals) for generating the torch::dataset.
Shapes can now have multiple NAs and not only the batch dimension can be missing. However, most nn() operators still expect only one missing values and will throw an error if multiple dimensions are unknown.
Training now does not fail anymore when encountering a missing value during validation but uses NA instead.
It is now possible to specify parameter groups for optimizers via the param_groups parameter.

LearnerTorchModel can now be parallelized and trained with encapsulation activated.
jit_trace now works in combination with batch normalization.
Ensures compatibility with R6 version 2.6.0

Removed some optimizers for which no fast ('ignite') variant exists.
The default optimizer is now AdamW instead of Adam.
The private LearnerTorch$.dataloader() method now operates no longer on the task but on the dataset generated by the private LearnerTorch$.dataset() method.
The shuffle parameter during model training is now initialized to TRUE to sidestep issues where data is sorted.

Optimizers now use the faster ('ignite') version of the optimizers, which leads to considerable speed improvements.
The jit_trace parameter was added to LearnerTorch, which when set to TRUE can lead to significant speedups. This should only be enabled for 'static' models, see the torch tutorial for more information.
Added parameter num_interop_threads to LearnerTorch.
The tensor_dataset parameter was added, which allows to stack all batches at the beginning of training to make loading of batches afterwards faster.
Use a faster default image loader.

Torch learners can now be used with AutoTuner.
Early stopping now not uses epochs - patience for the internally tuned values instead of the trained number of epochs as it was before.
The dataset of a learner must no longer return the tensors on the specified device, which allows for parallel dataloading on GPUs.
PipeOpBlock should no longer create ID clashes with other PipeOps in the graph (#260).

fix(preprocessing): regarding the construction of some PipeOps such as po("trafo_resize") which failed in some cases.
fix(ci): tests were not run in the CI
fix(learner): LearnerTabResnet now works correctly
Fix that tests were not run in the CI
feat: added the nn() helper function to simplify the creation of neural network layers