Added
- Add
TabPFNRegressor(presets=...), named bundles ofinference_configvalues. The"zero_inflated"preset is for nonnegative targets with many exact zeros: it sets thelog_fixed_atomtarget transform and thenonnegative_and_cappedreadout. On the models up to TabPFN-3.5 it setslog_quantile_gapwith the same readout instead.fitsuggests the preset, once per process, for nonnegative targets with more than 20% zeros and at least 50 widely spread, non-count positives. (#1413) - With GPU preprocessing, a table held in host memory goes through the selective steps (quantile, squashing, safepower, robust) a block at a time, and the fingerprint hashes it where it is. Any other step that cannot run by blocks raises on such a table. (#1426)
- Add the
"grouped"preset forgroup_coldata, on both estimators:presets="grouped"setsAPPEND_GROUP_AGGREGATES=TrueandPOOL_GROUP_PREDICTIONS="mean".TabPFNClassifiergets thepresetsargument for it. Both modes are computed over the rows passed to predict, so predict all rows of a group in one call. (#1434)
Changed
- A CPU preprocessing pipeline that writes nothing into a float32 input no longer copies it at predict, nor at fit inside
overwrite_fit_input(). (#1425) - With GPU preprocessing, random sign flips and the feature shuffle write into a table they may consume, and run on a table held in host memory without moving it to the GPU. (#1438)
- With GPU preprocessing, the quantile, squashing and safepower steps transform a table they may consume in place, a block of columns at a time. Quantile and safepower can then differ from a fit over all columns by float rounding. (#1440)
- When every CPU preprocessing step only selects, reorders or encodes (as with GPU preprocessing), ensemble members fitted inside
overwrite_fit_input()read the caller's table and hold only the categorical columns they encode, rather than a copy each. At predict, a test table without categorical columns is read the same way. (#1442)