tensorflow_model_server_linux.zip 0.2.6

TF-DF with TF-Serving binary for Tensorflow 2.9.1

0.2.5

Features

• Adds the contrib module for contributed, non-core functionality.
• Adds contrib.scikit_learn_model_converter, which facilitates converting
  Scikit-Learn tree-based models into TF-DF models.
• Discard hessian splits with score lower than the parents. This change has
  little effect on the model quality, but it can reduce its size.
• Add internal flag hessian_split_score_subtract_parent to subtract the
  parent score in the computation of an hessian split score.
• Add support for hyper-parameter optimizers (also called tuner).
• Add text pretty print of trees with tree.pretty() or str(tree).
• Add support for loading YDF models with file prefixes. Newly created models
  have a random prefix attached to them. This allows combining multiple models
  in Keras.
• Add support for discretized numerical features.

0.2.3

Features

• Honest Random Forests (also work with Gradient Boosted Tree and CART).
• Can train Random Forests with example sampling without replacement.
• Add support for Focal Loss with Gradient Boosted Trees.
• Add support for MacOS.

Fixes

• Incorrect default evaluation of categorical split with uplift tasks. This
  was making uplift models with missing categorical values perform worst, and
  made the inference of uplift model possibly slower.
• Fix pd_dataframe_to_tf_dataset on Pandas dataframe not containing arrays.

0.2.2

Features

• Surface the validation_interval_in_trees,
  keep_non_leaf_label_distribution and 'random_seed' hyper-parameters.
• Add the batch_size argument in the pd_dataframe_to_tf_dataset utility.
• Automatically determine the number of threads if num_threads=None.
• Add constructor argument try_resume_training to facilitate resuming
  training.
• Check that the training dataset is well configured for TF-DF e.g. no repeat
  operation, has a large enough batch size, etc. The check can be disabled
  with check_dataset=False.
• When a model is created manually with the model builder, and if the dataspec
  is not provided, tries to adapt the dataspec so that the model looks as if
  it was trained with the global imputation strategy for missing values (i.e.
  missing_value_policy: GLOBAL_IMPUTATION). This makes manually created models
  more likely to be compatible with the fast inference engines.
• TF-DF models fit method now passes the validation_data to the Yggdrasil
  learners. This is used for example for early stopping in the case of GBT
  model.
• Add the "loss" parameter of the GBT model directly in the model constructor.
• Control the amount of training logs displayed in the notebook (if using
  notebook) or in the console with the verbose constructor argument and
  fit parameter of the model.

Fixes

• num_candidate_attributes is not ignored anymore when
  num_candidate_attributes_ratio=-1.
• Use the median bucket split value strategy in the discretized numerical
  splitters (local and distributed).
• Surface the max_num_scanned_rows_to_accumulate_statistics parameter to
  control how many examples are scanned to determine the feature statistics
  when training from a file dataset with fit_on_dataset_path.

0.2.1

Features

• Compatibility with TensorFlow 2.7.0.

0.2.0

Features

• Add advanced option predict_single_probability_for_binary_classification
  to generate prediction tensors of shape [batch_size, 2] for binary
  classification model.
• Add support for weighted training.
• Add support for permutation variable importance in the GBT learner with the
  compute_permutation_variable_importance parameter.
• Support for tf.int8 and tf.int16 values.
• Support for distributed gradient boosted trees learning. Currently, the TF
  ParameterServerStrategy distribution strategy is only available in
  monolithic TF-DF builds. The Yggdrasil Decision Forest GRPC distribute
  strategy can be used instead.
• Support for training from dataset stored on disk in CSV and RecordIO format
  (instead of creating a tensorflow dataset). This option is currently more
  efficient for distributed training (until the ParameterServerStrategy
  support per-worker datasets).
• Add max_vocab_count argument to the model constructor. The existing
  max_vocab_count argument in FeatureUsage objects take precedence.

Fixes

• Missing filtering of unique values in the categorical-set training feature
  accumulator. Was responsible for a small (e.g. ~0.5% on SST2 dataset) drop
  of accuracy compared to the C++ API.
• Fix broken support for max_vocab_count in a FeatureUsage with type
  CATEGORICAL_SET.

0.1.9

Features

• Disable tree pruning in the CART algorithm if the validation dataset is empty (i.e. validation_ratio=0).
• Migration to Tensorflow 2.6. You will see an undefined symbol error if you install this version with a TensorFlow version different than 2.6. Previous versions were compiled for TF 2.5.

Fixes

• Fix failure from
  Github Issue #45
  where the wrong field was accessed for leaf node distributions.
• Fix saving of categorical features specification in the Builder.

0.1.9rc1

Pre-release of 0.1.9

Major change : Tensorflow 2.6 compatibility

This release is currently being tested and will be updated to be the latest version in PyPI soon, in the meantime users who need the fixes below can install this version directly from the wheels below, i.e. pip install tensorflow_decision_forests-0.1.9-cp39-cp39-manylinux_2_12_x86_64.manylinux2010_x86_64.whl for python 3.9.

Fixes

• Fix failure from Github Issue #45 where the wrong field was accessed for leaf node distributions.

• Fix incorrect handling of CART pruning when validation set is empty. Previously, the whole tree would be erroneously pruned. Now, pruning is disabled if the validation set is not specified.

• Fix saving of categorical features specification in the Builder.

• Migration to Tensorflow 2.6. You will see an undefined symbol error if you install this version with a TensorFlow version different than 2.6. Previous versions were compiled for TF 2.5.

0.1.8

Features

• Model can be composed with the functional Keras API before being trained.
• Makes all the Yggdrasil structural variable importances available.
• Makes getting the variable importance instantaneous.
• Surface the name argument in the model classes constructors.
• Add a postprocessing model constructor argument to easy apply
  post-processing on the model predictions without relying on the Keras
  Functional API.
• Add extract_all_trees method in the model inspector to efficiently exact
  all the trees.
• Add num_threads constructor argument to control the number of training
  threads without using the advanced configuration.
• By default, remove the temporary directory used to train the model when the
  model python object is garbage collected.
• Add the import_dataspec constructor argument to the model builder to
  import the feature definition and dictionaries (instead of relying on
  automatic discovery).

Changes

• When saving a model in a directory already containing a model, only the
  assets directory is entirely removed before the export (instead of the
  entire model directory).

Fixes

• Wrong label shape in the model inspector's objective field for
  pre-integerized labels.

0.1.7

Features

• Add more of characters to the non-recommended list of feature name
  characters.
• Make the inference op multi-thread compatible.
• Print an explicit error and some instructions when training a model with a
  Pandas dataframe.
• pd_dataframe_to_tf_dataset can automatically rename feature to make them
  compatible with SavedModel export signatures.
• model.save(...) can override an existing model.
• The link function of GBT model can be removed. For example, a binary
  classification GBT model trained with apply_link_function=False will output
  logits.