Fast and flexible AutoML with learning guarantees.
PredictionOutput that are not in the best_export_outputs.warm_start support to adanet Estimators.AutoEnsembleTPUEstimator.adanet.experimental Keras ModelFlow APIs.model_dir.adanet.ensemble.MeanEnsembler with a basic implementation for taking the mean of logits of subnetworks. This also supports including the mean of last_layer (helpful if subnetworks have same configurations) in the predictions and export_outputs of the EstimatorSpec.adanet.Evaluator is changing. The Evaluator.evaluate_adanet_losses(sess, adanet_losses) function is being replaced with Evaluator.evaluate(sess, ensemble_metrics). The ensemble_metrics parameter contains all computed metrics for each candidate ensemble as well as the adanet_loss. Code which overrides evaluate_adanet_losses must migrate over to use the new evaluate method (we suspect that such cases are very rare).adanet.Evaluator before Estimator#evaluate, Estimator#predict, and Estimator#export_saved_model. This can have the effect of changing the best candidate chosen at the final round. When the user passes an Evaluator, we run it to establish the best candidate during evaluation, predict, and export_saved_model. Previously they used the adanet_loss moving average collected during training. While the previous ensemble would have been established by the Evaluator, the current set of candidate ensembles that were not done training would be considered according to the adanet_loss. Now when a user passes an Evaluator that, for example, uses a hold-out set, AdaNet runs it before making predictions or exporting a SavedModel to use the best new candidate according to the hold-out set.tf.keras.metrics.Metrics during evaluation.OutOfRangeError raised during bagging.max_steps and steps are both None.TPUEmbedding.max_iteration_steps=None.adanet.AutoEnsembleSubestimator for training subestimators on different training data partitions and implement ensemble methods like bootstrap aggregating (a.k.a bagging).AutoEnsembleEstimator during distributed training.AutoEnsembleEstimator's candidate_pool argument to be a lambda in order to create Estimators lazily.adanet.subnetwork.Builder#prune_previous_ensemble for abstract class. This behavior is now specified using adanet.ensemble.Strategy subclasses.adanet.AutoEnsembleEstimator. This bug incremented the global_step by n+1 for n canned Estimators like DNNEstimator.adanet.TPUEstimator with adanet.Estimator feature parity.adanet.AutoEnsembleEstimator constructor to specify human-readable candidate names.tf.estimator.Estimator subclasses.adanet.ensemble which contains interfaces and examples of ways to learn ensembles using AdaNet. Users can now extend AdaNet to use custom ensemble-learning methods.scalar, image, histogram, and audio summaries on TPU during training.tf.train.SessionRunHook support to handle more edge cases.adanet.subnetwork package using from adanet.core import subnetwork will no longer work, because the package was moved to the adanet/subnetwork directory. Most users should already be using adanet.subnetwork or from adanet import subnetwork, and should not be affected.adanet.TPUEstimator.tf.train.SessionRunHook instances for training with adanet.subnetwork.TrainOpSpec.shared field to adanet.Subnetwork to deprecate, replace, and be more flexible than persisted_tensors.Tensors between iterations including Python primitives, objects, and lambdas for greater flexibility. Eliminating reliance on a MetaGraphDef proto also eliminates I/O allowing for faster training, and better future-proofing.adanet.Estimator.adanet.AutoEnsembleEstimator for learning to ensemble tf.estimator.Estimator instances.adanet.subnetwork.Builder's build_subnetwork method.adanet.subnetwork.Builder, so not passing var_list to the optimizer.minimize will lead to the same behavior as passing it in by default.tf.summary inside adanet.subnetwork.Builder is now equivalent to using the adanet.Summary object.global_step from within an adanet.subnetwork.Builder will return the iteration_step variable instead, so that the step starts at zero at the beginning of each iteration. One subnetwork incrementing the step will not affect other subnetworks.tf.name_scope("") hack.adanet.Estimator models.