You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
This commit was created on GitHub.com and signed with GitHub’s verified signature.
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)