NEWS
mlr3automl 0.1.0.9000
- fix: The fastai learner now requires the Python
torch package below version 2.14. Version 2.14 attaches logging handlers whose stream is a read-only property, which reticulate::py_capture_output() cannot redirect, so training failed with "property 'stream' of '_StderrHandler' object has no setter". install_python_learners() installs a compatible version when "fastai" is among the learners.
- fix: The default configuration of the tabpfn learner in the initial design now matches TabPFN-3. The number of estimators changed from 4 to 8 and the softmax temperature from 1.0 to 0.9, which were the defaults of TabPFN-2.
- fix: The tabpfn learner now sets
auto_scale_n_estimators = FALSE. TabPFN-3 otherwise raises the number of estimators on its own when the task has more features than a single ensemble member sees, which overrides the tuned value. This requires mlr3extralearners 1.6.0.9000 or later, which is now the minimum version.
- fix: Tuning on subspaces no longer fails when a learner has internally tuned parameters, e.g.
xgboost.nrounds or ft_transformer.epochs. The subspaces are now derived from the search space of the tuning instance, which no longer holds the internally tuned parameters.
- fix: The torch learners (mlp, resnet, and ft_transformer) now train on the CPU when
devices does not include "cuda". Previously they passed device = "auto" to mlr3torch, which selects the GPU whenever one is available, so they ignored devices = "cpu" and the resource accounting of the workers.
- fix: The torch learners (mlp, resnet, and ft_transformer) now train with a batch size of 256 instead of 32 on the GPU. A batch of 32 rows leaves the GPU mostly idle because kernel launches and host to device transfers dominate the step time.
- fix: The torch learners (mlp, resnet, and ft_transformer) are no longer removed from the search space by
memory_limit when they train on the GPU. Their memory estimates are fitted on host memory measurements, which do not describe the memory the model allocates on the GPU.
- fix: The gpu compute profile is now exempt from the large data set rules. Its number of workers is fixed by the number of GPUs, so reducing it does not free any resources, and its worker previously received the CPU cores and the memory that are freed on the cpu profiles.
- fix: The number of workers of a compute profile is now reduced to a quarter rounded up instead of rounded down, so a profile no longer loses more than the intended factor, e.g. 7 workers are reduced to 2 instead of 1.
- feat: New learner id
"tabfm" adds the TabFM tabular foundation model from mlr3extralearners to the search space, along with the new auto learners classif.auto_tabfm and regr.auto_tabfm. It runs via reticulate and is registered for "cuda" only, because it predicts in context and runs the backbone over the training rows once per estimator, which is too slow to be useful on the CPU. Construct AutoTabFM$new(devices = c("cpu", "cuda")) and re-register it in mlr_auto to run it on the CPU anyway.
- feat: The auto learners gained the
n_cpu and n_gpu parameters that override the per-learner resource requirements. When the requirements are mixed, "cuda" is part of devices, and the workers are distributed over the mirai compute profiles "mlr3automl_cpu" and "mlr3automl_gpu" with rush::rush_plan(profiles = c(mlr3automl_cpu = 7, mlr3automl_gpu = 1)), the search space is partitioned into a cpu and a gpu subspace and tuned with mlr3mbo::TunerADBOSubspaces. The workers of a profile only evaluate points of the subspace of that profile.
- BREAKING CHANGE: With
devices = c("cpu", "cuda"), the boosting learners (xgboost, lightgbm, and catboost) now train on the CPU by default because their default n_gpu requirement is 0. Set e.g. n_gpu = c(xgboost = 1) to train them on the GPU again.
mlr3automl 0.1.0 (2026-09-03)