github PriorLabs/TabPFN v10.1.0

one hour ago

Breaking Changes

  • fit now accepts sample_weight, so scikit-learn meta-estimators pass weights to it, and fit raises for models without SAMPLE_WEIGHT_MODE. BaggingClassifier(TabPFNClassifier()) and BaggingRegressor(TabPFNRegressor()) raise in fit with any bootstrap and max_samples, because they pass their row counts as weights instead of fitting on a row subset. CalibratedClassifierCV(TabPFNClassifier()).fit(X, y, sample_weight=w) raises instead of warning that the estimator ignores the weights. No released model supports sample weights yet. (#1415)
  • A missing value in a group_col column now raises in fit and predict. Before, such a row was treated as belonging to no known group. (#1418)

Added

  • Add the log_quantile_gap, log_fixed_atom and log1p_target regression target transforms for nonnegative targets. log_quantile_gap and log_fixed_atom decode a range to exactly 0 for zero-inflated targets. log1p_target encodes log1p(y) of the raw target and decodes negative codes to 0. (#1409)
  • Add InferenceConfig.REGRESSION_READOUT. With "nonnegative_and_capped", the regression outputs of a nonnegative target are read from the ensemble's averaged distribution of max(Y, 0) given Y <= y_max, with y_max the largest training target. (#1410)
  • Add RegressorTuningConfig.calibrate_mean. It multiplies the predicted mean by a constant fitted on the tuning holdout rows. Its only mean_calibration_mode, "tweedie", picks the constant that minimizes the Tweedie deviance (power mean_calibration_power, default 1.5). (#1411)
  • RegressorTuningConfig.calibrate_mean weights each tuning holdout row by sample_weight * label_reliability. For an exposure, pass the rate as the target and the exposure as sample_weight. (#1412)
  • TabPFNClassifier.fit and TabPFNRegressor.fit take experimental sample_weight and label_reliability arguments, applied by models whose inference config sets SAMPLE_WEIGHT_MODE. For every other model fit raises when they are given. (#1415)
  • Two inference_config flags add transductive modes for group_col data: APPEND_GROUP_AGGREGATES appends per-group aggregates of the features, and POOL_GROUP_PREDICTIONS gives the rows of a group their mean or geometric-mean prediction when every group has one label. A missing value in a group_col column is now refused. (#1418)
  • sample_weight and label_reliability can be combined with tuning_config: each tuning fold is fit with the weights of its rows, and a holdout row counts in the tuned metric in proportion to sample_weight * label_reliability. (#1428)
  • Add InferenceConfig.AVERAGE_BEFORE_SOFTMAX, so a checkpoint can set whether the ensemble averages logits or probabilities. average_before_softmax now defaults to "auto", which uses the checkpoint's value. (#1432)

Changed

  • Lower peak host and GPU memory when fitting and predicting on large tables: preprocessing works in place by blocks, a float32 input stays float32 where the values are not changed, and the model input can stay in host memory for a model whose forward takes x_host. Add tabpfn.preprocessing.clean.overwrite_fit_input() to let fit write into an array the caller will not read again. (#1422)
  • With ENABLE_GPU_PREPROCESSING, the robust preset runs on the GPU in float32. (#1424)
  • With APPEND_GROUP_AGGREGATES, the fit selects the per-group aggregates faster: on n_preprocessing_jobs threads, and for numeric columns on the GPU when ENABLE_GPU_PREPROCESSING is set and the estimator runs on a CUDA device. The default of one job on CPU is unchanged. (#1429)

Fixed

  • TabPFNRegressor.predict with fit_mode="fit_with_cache" now predicts large test sets in blocks of rows, so peak memory no longer grows with the number of test rows. The block size is set by TABPFN_MAX_REGRESSOR_PREDICT_ROWS (default 65,536; 0 disables it). (#1417)

Don't miss a new TabPFN release

NewReleases is sending notifications on new releases.