tfdf.keras.GradientBoostedTreesModel

Gradient Boosted Trees learning algorithm.

Inherits From: GradientBoostedTreesModel, CoreModel, InferenceCoreModel

Used in the notebooks

Used in the guide Used in the tutorials

A Gradient Boosted Trees (GBT), also known as Gradient Boosted Decision Trees (GBDT) or Gradient Boosted Machines (GBM), is a set of shallow decision trees trained sequentially. Each tree is trained to predict and then "correct" for the errors of the previously trained trees (more precisely each tree predict the gradient of the loss relative to the model output).

Usage example:

import tensorflow_decision_forests as tfdf
import pandas as pd

dataset = pd.read_csv("project/dataset.csv")
tf_dataset = tfdf.keras.pd_dataframe_to_tf_dataset(dataset, label="my_label")

model = tfdf.keras.GradientBoostedTreesModel()
model.fit(tf_dataset)

print(model.summary())

Hyper-parameter tuning:

import tensorflow_decision_forests as tfdf
import pandas as pd

dataset = pd.read_csv("project/dataset.csv")
tf_dataset = tfdf.keras.pd_dataframe_to_tf_dataset(dataset, label="my_label")

tuner = tfdf.tuner.RandomSearch(num_trials=20)

# Hyper-parameters to optimize.
tuner.discret("max_depth", [4, 5, 6, 7])

model = tfdf.keras.GradientBoostedTreesModel(tuner=tuner)
model.fit(tf_dataset)

print(model.summary())

task Task to solve (e.g. Task.CLASSIFICATION, Task.REGRESSION, Task.RANKING, Task.CATEGORICAL_UPLIFT, Task.NUMERICAL_UPLIFT).
features Specify the list and semantic of the input features of the model. If not specified, all the available features will be used. If specified and if exclude_non_specified_features=True, only the features in features will be used by the model. If "preprocessing" is used, features corresponds to the output of the preprocessing. In this case, it is recommended for the preprocessing to return a dictionary of tensors.
exclude_non_specified_features If true, only use the features specified in features.
preprocessing Functional keras model or @tf.function to apply on the input feature before the model to train. This preprocessing model can consume and return tensors, list of tensors or dictionary of tensors. If specified, the model only "sees" the output of the preprocessing (and not the raw input). Can be used to prepare the features or to stack multiple models on top of each other. Unlike preprocessing done in the tf.dataset, the operation in "preprocessing" are serialized with the model.
postprocessing Like "preprocessing" but applied on the model output.
training_preprocessing Functional keras model or @tf.function to apply on the input feature, labels, and sample_weight before model training.
ranking_group Only for task=Task.RANKING. Name of a tf.string feature that identifies queries in a query/document ranking task. The ranking group is not added automatically for the set of features if exclude_non_specified_features=false.
uplift_treatment Only for task=Task.CATEGORICAL_UPLIFT or task=Task.NUMERICAL_UPLIFT. Name of an integer feature that identifies the treatment in an uplift problem. The value 0 is reserved for the control treatment.
temp_directory Temporary directory used to store the model Assets after the training, and possibly as a work directory during the training. This temporary directory is necessary for the model to be exported after training e.g. model.save(path). If not specified, temp_directory is set to a temporary directory using tempfile.TemporaryDirectory. This directory is deleted when the model python object is garbage-collected.
verbose Verbosity mode. 0 = silent, 1 = small details, 2 = full details.
hyperparameter_template Override the default value of the hyper-parameters. If None (default) the default parameters of the library are used. If set, default_hyperparameter_template refers to one of the following preconfigured hyper-parameter sets. Those sets outperforms the default hyper-parameters (either generally or in specific scenarios). You can omit the version (e.g. remove "@v5") to use the last version of the template. In this case, the hyper-parameter can change in between releases (not recommended for training in production).

  • better_default@v1: A configuration that is generally better than the default parameters without being more expensive. The parameters are: growing_strategy="BEST_FIRST_GLOBAL".
  • benchmark_rank1@v1: Top ranking hyper-parameters on our benchmark slightly modified to run in reasonable time. The parameters are: growing_strategy="BEST_FIRST_GLOBAL", categorical_algorithm="RANDOM", split_axis="SPARSE_OBLIQUE", sparse_oblique_normalization="MIN_MAX", sparse_oblique_num_projections_exponent=1.0.
advanced_arguments Advanced control of the model that most users won't need to use. See AdvancedArguments for details.
num_threads Number of threads used to train the model. Different learning algorithms use multi-threading differently and with different degree of efficiency. If None, num_threads will be automatically set to the number of processors (up to a maximum of 32; or set to 6 if the number of processors is not available). Making num_threads significantly larger than the number of processors can slow-down the training speed. The default value logic might change in the future.
name The name of the model.
max_vocab_count Default maximum size of the vocabulary for CATEGORICAL and CATEGORICAL_SET features stored as strings. If more unique values exist, only the most frequent values are kept, and the remaining values are considered as out-of-vocabulary. The value max_vocab_count defined in a FeatureUsage (if any) takes precedence.
try_resume_training If true, the model training resumes from the checkpoint stored in the temp_directory directory. If temp_directory does not contain any model checkpoint, the training start from the beginning. Resuming training is useful in the following situations: (1) The training was interrupted by the user (e.g. ctrl+c or "stop" button in a notebook). (2) the training job was interrupted (e.g. rescheduling), ond (3) the hyper-parameter of the model were changed such that an initially completed training is now incomplete (e.g. increasing the number of trees). Note: Training can only be resumed if the training datasets is exactly the same (i.e. no reshuffle in the tf.data.Dataset).
check_dataset If set to true, test if the dataset is well configured for the training: (1) Check if the dataset does contains any repeat operations, (2) Check if the dataset does contain a batch operation, (3) Check if the dataset has a large enough batch size (min 100 if the dataset contains more than 1k examples or if the number of examples is not available) If set to false, do not run any test.
tuner If set, automatically optimize the hyperparameters of the model using this tuner. If the model is trained with distribution (i.e. the model definition is wrapper in a TF Distribution strategy, the tuning is distributed.
discretize_numerical_features If true, discretize all the numerical features before training. Discretized numerical features are faster to train with, but they can have a negative impact on the model quality. Using discretize_numerical_features=True is equivalent as setting the feature semantic DISCRETIZED_NUMERICAL in the feature argument. See the definition of DISCRETIZED_NUMERICAL for more details.
num_discretize_numerical_bins Number of bins used when disretizing numerical features. The value num_discretized_numerical_bins defined in a FeatureUsage (if any) takes precedence.
multitask If set, train a multi-task model, that is a model with multiple outputs trained to predict different labels. If set, the tf.dataset label (i.e. the second selement of the dataset) should be a dictionary of label_key:label_values. Only one of multitask and task can be set.
adapt_subsample_for_maximum_training_duration Control how the maximum training duration (if set) is applied. If false, the training stop when the time is used. If true, the size of the sampled datasets used train individual trees are adapted dynamically so that all the trees are trained in time. Default: False.
allow_na_conditions If true, the tree training evaluates conditions of the type X is NA i.e. X is missing. Default: False.
apply_link_function If true, applies the link function (a.k.a. activation function), if any, before returning the model prediction. If false, returns the pre-link function model output. For example, in the case of binary classification, the pre-link function output is a logic while the post-link function is a probability. Default: True.
categorical_algorithm How to learn splits on categorical attributes.
  • CART: CART algorithm. Find categorical splits of the form "value \in mask". The solution is exact for binary classification, regression and ranking. It is approximated for multi-class classification. This is a good first algorithm to use. In case of overfitting (very small dataset, large dictionary), the "random" algorithm is a good alternative.
  • ONE_HOT: One-hot encoding. Find the optimal categorical split of the form "attribute == param". This method is similar (but more efficient) than converting converting each possible categorical value into a boolean feature. This method is available for comparison purpose and generally performs worse than other alternatives.
  • RANDOM: Best splits among a set of random candidate. Find the a categorical split of the form "value \in mask" using a random search. This solution can be seen as an approximation of the CART algorithm. This method is a strong alternative to CART. This algorithm is inspired from section "5.1 Categorical Variables" of "Random Forest", 2001. Default: "CART".
  • categorical_set_split_greedy_sampling For categorical set splits e.g. texts. Probability for a categorical value to be a candidate for the positive set. The sampling is applied once per node (i.e. not at every step of the greedy optimization). Default: 0.1.
    categorical_set_split_max_num_items For categorical set splits e.g. texts. Maximum number of items (prior to the sampling). If more items are available, the least frequent items are ignored. Changing this value is similar to change the "max_vocab_count" before loading the dataset, with the following exception: With max_vocab_count, all the remaining items are grouped in a special Out-of-vocabulary item. With max_num_items, this is not the case. Default: -1.
    categorical_set_split_min_item_frequency For categorical set splits e.g. texts. Minimum number of occurrences of an item to be considered. Default: 1.
    compute_permutation_variable_importance If true, compute the permutation variable importance of the model at the end of the training using the validation dataset. Enabling this feature can increase the training time significantly. Default: False.
    dart_dropout Dropout rate applied when using the DART i.e. when forest_extraction=DART. Default: None.
    early_stopping Early stopping detects the overfitting of the model and halts it training using the validation dataset. If not provided directly, the validation dataset is extracted from the training dataset (see "validation_ratio" parameter):
  • NONE: No early stopping. All the num_trees are trained and kept.
  • MIN_LOSS_FINAL: All the num_trees are trained. The model is then truncated to minimize the validation loss i.e. some of the trees are discarded as to minimum the validation loss.
  • LOSS_INCREASE: Classical early stopping. Stop the training when the validation does not decrease for early_stopping_num_trees_look_ahead trees. Default: "LOSS_INCREASE".
  • early_stopping_initial_iteration 0-based index of the first iteration considered for early stopping computation. Increasing this value prevents too early stopping due to noisy initial iterations of the learner. Default: 10.
    early_stopping_num_trees_look_ahead Rolling number of trees used to detect validation loss increase and trigger early stopping. Default: 30.
    focal_loss_alpha EXPERIMENTAL. Weighting parameter for focal loss, positive samples weighted by alpha, negative samples by (1-alpha). The default 0.5 value means no active class-level weighting. Only used with focal loss i.e. loss="BINARY_FOCAL_LOSS" Default: 0.5.
    focal_loss_gamma EXPERIMENTAL. Exponent of the misprediction exponent term in focal loss, corresponds to gamma parameter in https://arxiv.org/pdf/1708.02002.pdf. Only used with focal loss i.e. loss="BINARY_FOCAL_LOSS" Default: 2.0.
    forest_extraction How to construct the forest:
  • MART: For Multiple Additive Regression Trees. The "classical" way to build a GBDT i.e. each tree tries to "correct" the mistakes of the previous trees.
  • DART: For Dropout Additive Regression Trees. A modification of MART proposed in http://proceedings.mlr.press/v38/korlakaivinayak15.pdf. Here, each tree tries to "correct" the mistakes of a random subset of the previous trees. Default: "MART".
  • goss_alpha Alpha parameter for the GOSS (Gradient-based One-Side Sampling; "See LightGBM: A Highly Efficient Gradient Boosting Decision Tree") sampling method. Default: 0.2.
    goss_beta Beta parameter for the GOSS (Gradient-based One-Side Sampling) sampling method. Default: 0.1.
    growing_strategy How to grow the tree.
  • LOCAL: Each node is split independently of the other nodes. In other words, as long as a node satisfy the splits "constraints (e.g. maximum depth, minimum number of observations), the node will be split. This is the "classical" way to grow decision trees.
  • BEST_FIRST_GLOBAL: The node with the best loss reduction among all the nodes of the tree is selected for splitting. This method is also called "best first" or "leaf-wise growth". See "Best-first decision tree learning", Shi and "Additive logistic regression : A statistical view of boosting", Friedman for more details. Default: "LOCAL".
  • honest In honest trees, different training examples are used to infer the structure and the leaf values. This regularization technique trades examples for bias estimates. It might increase or reduce the quality of the model. See "Generalized Random Forests", Athey et al. In this paper, Honest trees are trained with the Random Forest algorithm with a sampling without replacement. Default: False.
    honest_fixed_separation For honest trees only i.e. honest=true. If true, a new random separation is generated for each tree. If false, the same separation is used for all the trees (e.g., in Gradient Boosted Trees containing multiple trees). Default: False.
    honest_ratio_leaf_examples For honest trees only i.e. honest=true. Ratio of examples used to set the leaf values. Default: 0.5.
    in_split_min_examples_check Whether to check the min_examples constraint in the split search (i.e. splits leading to one child having less than min_examples examples are considered invalid) or before the split search (i.e. a node can be derived only if it contains more than min_examples examples). If false, there can be nodes with less than min_examples training examples. Default: True.
    keep_non_leaf_label_distribution Whether to keep the node value (i.e. the distribution of the labels of the training examples) of non-leaf nodes. This information is not used during serving, however it can be used for model interpretation as well as hyper parameter tuning. This can take lots of space, sometimes accounting for half of the model size. Default: True.
    l1_regularization L1 regularization applied to the training loss. Impact the tree structures and lead values. Default: 0.0.
    l2_categorical_regularization L2 regularization applied to the training loss for categorical features. Impact the tree structures and lead values. Default: 1.0.
    l2_regularization L2 regularization applied to the training loss for all features except the categorical ones. Default: 0.0.
    lambda_loss Lambda regularization applied to certain training loss functions. Only for NDCG loss. Default: 1.0.
    loss The loss optimized by the model. If not specified (DEFAULT) the loss is selected automatically according to the \"task\" and label statistics. For example, if task=CLASSIFICATION and the label has two possible values, the loss will be set to BINOMIAL_LOG_LIKELIHOOD. Possible values are:
  • DEFAULT: Select the loss automatically according to the task and label statistics.
  • BINOMIAL_LOG_LIKELIHOOD: Binomial log likelihood. Only valid for binary classification.
  • SQUARED_ERROR: Least square loss. Only valid for regression.
  • POISSON: Poisson log likelihood loss. Mainly used for counting problems. Only valid for regression.
  • MULTINOMIAL_LOG_LIKELIHOOD: Multinomial log likelihood i.e. cross-entropy. Only valid for binary or multi-class classification.
  • LAMBDA_MART_NDCG5: LambdaMART with NDCG5.
  • XE_NDCG_MART: Cross Entropy Loss NDCG. See arxiv.org/abs/1911.09798.
  • BINARY_FOCAL_LOSS: Focal loss. Only valid for binary classification. See https://arxiv.org/pdf/1708.02002.pdf.
  • POISSON: Poisson log likelihood. Only valid for regression.
  • MEAN_AVERAGE_ERROR: Mean average error a.k.a. MAE. Default: "DEFAULT".
  • max_depth Maximum depth of the tree. max_depth=1 means that all trees will be roots. max_depth=-1 means that tree depth is not restricted by this parameter. Values <= -2 will be ignored. Default: 6.
    max_num_nodes Maximum number of nodes in the tree. Set to -1 to disable this limit. Only available for growing_strategy=BEST_FIRST_GLOBAL. Default: None.
    maximum_model_size_in_memory_in_bytes Limit the size of the model when stored in ram. Different algorithms can enforce this limit differently. Note that when models are compiled into an inference, the size of the inference engine is generally much smaller than the original model. Default: -1.0.
    maximum_training_duration_seconds Maximum training duration of the model expressed in seconds. Each learning algorithm is free to use this parameter at it sees fit. Enabling maximum training duration makes the model training non-deterministic. Default: -1.0.
    mhld_oblique_max_num_attributes For MHLD oblique splits i.e. split_axis=MHLD_OBLIQUE. Maximum number of attributes in the projection. Increasing this value increases the training time. Decreasing this value acts as a regularization. The value should be in [2, num_numerical_features]. If the value is above the total number of numerical features, the value is capped automatically. The value 1 is allowed but results in ordinary (non-oblique) splits. Default: None.
    mhld_oblique_sample_attributes For MHLD oblique splits i.e. split_axis=MHLD_OBLIQUE. If true, applies the attribute sampling controlled by the "num_candidate_attributes" or "num_candidate_attributes_ratio" parameters. If false, all the attributes are tested. Default: None.
    min_examples Minimum number of examples in a node. Default: 5.
    missing_value_policy Method used to handle missing attribute values.
  • GLOBAL_IMPUTATION: Missing attribute values are imputed, with the mean (in case of numerical attribute) or the most-frequent-item (in case of categorical attribute) computed on the entire dataset (i.e. the information contained in the data spec).
  • LOCAL_IMPUTATION: Missing attribute values are imputed with the mean (numerical attribute) or most-frequent-item (in the case of categorical attribute) evaluated on the training examples in the current node.
  • RANDOM_LOCAL_IMPUTATION: Missing attribute values are imputed from randomly sampled values from the training examples in the current node. This method was proposed by Clinic et al. in "Random Survival Forests" (https://projecteuclid.org/download/pdfview_1/euclid.aoas/1223908043). Default: "GLOBAL_IMPUTATION".
  • num_candidate_attributes Number of unique valid attributes tested for each node. An attribute is valid if it has at least a valid split. If num_candidate_attributes=0, the value is set to the classical default value for Random Forest: sqrt(number of input attributes) in case of classification and number_of_input_attributes / 3 in case of regression. If num_candidate_attributes=-1, all the attributes are tested. Default: -1.
    num_candidate_attributes_ratio Ratio of attributes tested at each node. If set, it is equivalent to num_candidate_attributes = number_of_input_features x num_candidate_attributes_ratio. The possible values are between ]0, and 1] as well as -1. If not set or equal to -1, the num_candidate_attributes is used. Default: -1.0.
    num_trees Maximum number of decision trees. The effective number of trained tree can be smaller if early stopping is enabled. Default: 300.
    pure_serving_model Clear the model from any information that is not required for model serving. This includes debugging, model interpretation and other meta-data. The size of the serialized model can be reduced significatively (50% model size reduction is common). This parameter has no impact on the quality, serving speed or RAM usage of model serving. Default: False.
    random_seed Random seed for the training of the model. Learners are expected to be deterministic by the random seed. Default: 123456.
    sampling_method Control the sampling of the datasets used to train individual trees.
  • NONE: No sampling is applied. This is equivalent to RANDOM sampling with \"subsample=1\".
  • RANDOM (default): Uniform random sampling. Automatically selected if "subsample" is set.
  • GOSS: Gradient-based One-Side Sampling. Automatically selected if "goss_alpha" or "goss_beta" is set.
  • SELGB: Selective Gradient Boosting. Automatically selected if "selective_gradient_boosting_ratio" is set. Only valid for ranking. Default: "RANDOM".
  • selective_gradient_boosting_ratio Ratio of the dataset used to train individual tree for the selective Gradient Boosting (Selective Gradient Boosting for Effective Learning to Rank; Lucchese et al; http://quickrank.isti.cnr.it/selective-data/selective-SIGIR2018.pdf) sampling method. Default: 0.01.
    shrinkage Coefficient applied to each tree prediction. A small value (0.02) tends to give more accurate results (assuming enough trees are trained), but results in larger models. Analogous to neural network learning rate. Fixed to 1.0 for DART models. Default: 0.1.
    sorting_strategy How are sorted the numerical features in order to find the splits
  • AUTO: Selects the most efficient method among IN_NODE, FORCE_PRESORT, and LAYER.
  • IN_NODE: The features are sorted just before being used in the node. This solution is slow but consumes little amount of memory.
  • FORCE_PRESORT: The features are pre-sorted at the start of the training. This solution is faster but consumes much more memory than IN_NODE.
  • PRESORT: Automatically choose between FORCE_PRESORT and IN_NODE. . Default: "PRESORT".
  • sparse_oblique_max_num_projections For sparse oblique splits i.e. split_axis=SPARSE_OBLIQUE. Maximum number of projections (applied after the num_projections_exponent). Oblique splits try out max(p^num_projections_exponent, max_num_projections) random projections for choosing a split, where p is the number of numerical features. Increasing "max_num_projections" increases the training time but not the inference time. In late stage model development, if every bit of accuracy if important, increase this value. The paper "Sparse Projection Oblique Random Forests" (Tomita et al, 2020) does not define this hyperparameter. Default: None.
    sparse_oblique_normalization For sparse oblique splits i.e. split_axis=SPARSE_OBLIQUE. Normalization applied on the features, before applying the sparse oblique projections.
  • NONE: No normalization.
  • STANDARD_DEVIATION: Normalize the feature by the estimated standard deviation on the entire train dataset. Also known as Z-Score normalization.
  • MIN_MAX: Normalize the feature by the range (i.e. max-min) estimated on the entire train dataset. Default: None.
  • sparse_oblique_num_projections_exponent For sparse oblique splits i.e. split_axis=SPARSE_OBLIQUE. Controls of the number of random projections to test at each node. Increasing this value very likely improves the quality of the model, drastically increases the training time, and doe not impact the inference time. Oblique splits try out max(p^num_projections_exponent, max_num_projections) random projections for choosing a split, where p is the number of numerical features. Therefore, increasing this num_projections_exponent and possibly max_num_projections may improve model quality, but will also significantly increase training time. Note that the complexity of (classic) Random Forests is roughly proportional to num_projections_exponent=0.5, since it considers sqrt(num_features) for a split. The complexity of (classic) GBDT is roughly proportional to num_projections_exponent=1, since it considers all features for a split. The paper "Sparse Projection Oblique Random Forests" (Tomita et al, 2020) recommends values in [1/4, 2]. Default: None.
    sparse_oblique_projection_density_factor Density of the projections as an exponent of the number of features. Independently for each projection, each feature has a probability "projection_density_factor / num_features" to be considered in the projection. The paper "Sparse Projection Oblique Random Forests" (Tomita et al, 2020) calls this parameter lambda and recommends values in [1, 5]. Increasing this value increases training and inference time (on average). This value is best tuned for each dataset. Default: None.
    sparse_oblique_weights For sparse oblique splits i.e. split_axis=SPARSE_OBLIQUE. Possible values:
  • BINARY: The oblique weights are sampled in {-1,1} (default).
  • CONTINUOUS: The oblique weights are be sampled in [-1,1]. Default: None.
  • split_axis What structure of split to consider for numerical features.
  • AXIS_ALIGNED: Axis aligned splits (i.e. one condition at a time). This is the "classical" way to train a tree. Default value.
  • SPARSE_OBLIQUE: Sparse oblique splits (i.e. random splits on a small number of features) from "Sparse Projection Oblique Random Forests", Tomita et al., 2020.
  • MHLD_OBLIQUE: Multi-class Hellinger Linear Discriminant splits from "Classification Based on Multivariate Contrast Patterns", Canete-Sifuentes et al., 2029 Default: "AXIS_ALIGNED".
  • subsample Ratio of the dataset (sampling without replacement) used to train individual trees for the random sampling method. If \"subsample\" is set and if \"sampling_method\" is NOT set or set to \"NONE\", then \"sampling_method\" is implicitly set to \"RANDOM\". In other words, to enable random subsampling, you only need to set "\"subsample\". Default: 1.0.
    uplift_min_examples_in_treatment For uplift models only. Minimum number of examples per treatment in a node. Default: 5.
    uplift_split_score For uplift models only. Splitter score i.e. score optimized by the splitters. The scores are introduced in "Decision trees for uplift modeling with single and multiple treatments", Rzepakowski et al. Notation: p probability / average value of the positive outcome, q probability / average value in the control group.
  • KULLBACK_LEIBLER or KL: - p log (p/q)
  • EUCLIDEAN_DISTANCE or ED: (p-q)^2
  • CHI_SQUARED or CS: (p-q)^2/q Default: "KULLBACK_LEIBLER".
  • use_hessian_gain Use true, uses a formulation of split gain with a hessian term i.e. optimizes the splits to minimize the variance of "gradient / hessian. Available for all losses except regression. Default: False.
    validation_interval_in_trees Evaluate the model on the validation set every "validation_interval_in_trees" trees. Increasing this value reduce the cost of validation and can impact the early stopping policy (as early stopping is only tested during the validation). Default: 1.
    validation_ratio Fraction of the training dataset used for validation if not validation dataset is provided. The validation dataset, whether provided directly or extracted from the training dataset, is used to compute the validation loss, other validation metrics, and possibly trigger early stopping (if enabled). When early stopping is disabled, the validation dataset is only used for monitoring and does not influence the model directly. If the "validation_ratio" is set to 0, early stopping is disabled (i.e., it implies setting early_stopping=NONE). Default: 0.1.
    activity_regularizer Optional regularizer function for the output of this layer.
    autotune_steps_per_execution Settable property to enable tuning for steps_per_execution
    compute_dtype The dtype of the layer's computations.

    This is equivalent to Layer.dtype_policy.compute_dtype. Unless mixed precision is used, this is the same as Layer.dtype, the dtype of the weights.

    Layers automatically cast their inputs to the compute dtype, which causes computations and the output to be in the compute dtype as well. This is done by the base Layer class in Layer.call, so you do not have to insert these casts if implementing your own layer.

    Layers often perform certain internal computations in higher precision when compute_dtype is float16 or bfloat16 for numeric stability. The output will still typically be float16 or bfloat16 in such cases.

    distribute_reduction_method The method employed to reduce per-replica values during training.

    Unless specified, the value "auto" will be assumed, indicating that the reduction strategy should be chosen based on the current running environment. See reduce_per_replica function for more details.

    distribute_strategy The tf.distribute.Strategy this model was created under.
    dtype The dtype of the layer weights.

    This is equivalent to Layer.dtype_policy.variable_dtype. Unless mixed precision is used, this is the same as Layer.compute_dtype, the dtype of the layer's computations.

    dtype_policy The dtype policy associated with this layer.

    This is an instance of a tf.keras.mixed_precision.Policy.

    dynamic Whether the layer is dynamic (eager-only); set in the constructor.
    input Retrieves the input tensor(s) of a layer.

    Only applicable if the layer has exactly one input, i.e. if it is connected to one incoming layer.

    input_spec InputSpec instance(s) describing the input format for this layer.

    When you create a layer subclass, you can set self.input_spec to enable the layer to run input compatibility checks when it is called. Consider a Conv2D layer: it can only be called on a single input tensor of rank 4. As such, you can set, in __init__():

    self.input_spec = tf.keras.layers.InputSpec(ndim=4)
    

    Now, if you try to call the layer on an input that isn't rank 4 (for instance, an input of shape (2,), it will raise a nicely-formatted error:

    ValueError: Input 0 of layer conv2d is incompatible with the layer:
    expected ndim=4, found ndim=1. Full shape received: [2]
    

    Input checks that can be specified via input_spec include:

    • Structure (e.g. a single input, a list of 2 inputs, etc)
    • Shape
    • Rank (ndim)
    • Dtype

    For more information, see tf.keras.layers.InputSpec.

    jit_compile Specify whether to compile the model with XLA.

    XLA is an optimizing compiler for machine learning. jit_compile is not enabled by default. Note that jit_compile=True may not necessarily work for all models.

    For more information on supported operations please refer to the XLA documentation. Also refer to known XLA issues for more details.

    layers

    learner Name of the learning algorithm used to train the model.
    learner_params Gets the dictionary of hyper-parameters passed in the model constructor.

    Changing this dictionary will impact the training.

    losses List of losses added using the add_loss() API.

    Variable regularization tensors are created when this property is accessed, so it is eager safe: accessing losses under a tf.GradientTape will propagate gradients back to the corresponding variables.

    class MyLayer(tf.keras.layers.Layer):
      def call(self, inputs):
        self.add_loss(tf.abs(tf.reduce_mean(inputs)))
        return inputs
    l = MyLayer()
    l(np.ones((10, 1)))
    l.losses
    [1.0]
    inputs = tf.keras.Input(shape=(10,))
    x = tf.keras.layers.Dense(10)(inputs)
    outputs = tf.keras.layers.Dense(1)(x)
    model = tf.keras.Model(inputs, outputs)
    # Activity regularization.
    len(model.losses)
    0
    model.add_loss(tf.abs(tf.reduce_mean(x)))
    len(model.losses)
    1
    inputs = tf.keras.Input(shape=(10,))
    d = tf.keras.layers.Dense(10, kernel_initializer=&#x27;ones')
    x = d(inputs)
    outputs = tf.keras.layers.Dense(1)(x)
    model = tf.keras.Model(inputs, outputs)
    # Weight regularization.
    model.add_loss(lambda: tf.reduce_mean(d.kernel))
    model.losses
    [<tf.Tensor: shape=(), dtype=float32, numpy=1.0>]

    metrics Return metrics added using compile() or add_metric().

    inputs = tf.keras.layers.Input(shape=(3,))
    outputs = tf.keras.layers.Dense(2)(inputs)
    model = tf.keras.models.Model(inputs=inputs, outputs=outputs)
    model.compile(optimizer="Adam", loss="mse", metrics=["mae"])
    [m.name for m in model.metrics]
    []
    x = np.random.random((2, 3))
    y = np.random.randint(0, 2, (2, 2))
    model.fit(x, y)
    [m.name for m in model.metrics]
    [&#x27;loss', 'mae']
    inputs = tf.keras.layers.Input(shape=(3,))
    d = tf.keras.layers.Dense(2, name=&#x27;out')
    output_1 = d(inputs)
    output_2 = d(inputs)
    model = tf.keras.models.Model(
       inputs=inputs, outputs=[output_1, output_2])
    model.add_metric(
       tf.reduce_sum(output_2), name=&#x27;mean', aggregation='mean')
    model.compile(optimizer="Adam", loss="mse", metrics=["mae", "acc"])
    model.fit(x, (y, y))
    [m.name for m in model.metrics]
    [&#x27;loss', 'out_loss', 'out_1_loss', 'out_mae', 'out_acc', 'out_1_mae',
    &#x27;out_1_acc', 'mean']

    metrics_names Returns the model's display labels for all outputs.

    inputs = tf.keras.layers.Input(shape=(3,))
    outputs = tf.keras.layers.Dense(2)(inputs)
    model = tf.keras.models.Model(inputs=inputs, outputs=outputs)
    model.compile(optimizer="Adam", loss="mse", metrics=["mae"])
    model.metrics_names
    []
    x = np.random.random((2, 3))
    y = np.random.randint(0, 2, (2, 2))
    model.fit(x, y)
    model.metrics_names
    [&#x27;loss', 'mae']
    inputs = tf.keras.layers.Input(shape=(3,))
    d = tf.keras.layers.Dense(2, name=&#x27;out')
    output_1 = d(inputs)
    output_2 = d(inputs)
    model = tf.keras.models.Model(
       inputs=inputs, outputs=[output_1, output_2])
    model.compile(optimizer="Adam", loss="mse", metrics=["mae", "acc"])
    model.fit(x, (y, y))
    model.metrics_names
    [&#x27;loss', 'out_loss', 'out_1_loss', 'out_mae', 'out_acc', 'out_1_mae',
    &#x27;out_1_acc']

    name_scope Returns a tf.name_scope instance for this class.
    non_trainable_weights List of all non-trainable weights tracked by this layer.

    Non-trainable weights are not updated during training. They are expected to be updated manually in call().

    num_training_examples Number of training examples.
    num_validation_examples Number of validation examples.
    output Retrieves the output tensor(s) of a layer.

    Only applicable if the layer has exactly one output, i.e. if it is connected to one incoming layer.

    run_eagerly Settable attribute indicating whether the model should run eagerly.

    Running eagerly means that your model will be run step by step, like Python code. Your model might run slower, but it should become easier for you to debug it by stepping into individual layer calls.

    By default, we will attempt to compile your model to a static graph to deliver the best execution performance.

    steps_per_execution Settable steps_per_execution variable. Requires a compiled model. </td> </tr><tr> <td>submodules` Sequence of all sub-modules.

    Submodules are modules which are properties of this module, or found as properties of modules which are properties of this module (and so on).

    a = tf.Module()
    b = tf.Module()
    c = tf.Module()
    a.b = b
    b.c = c
    list(a.submodules) == [b, c]
    True
    list(b.submodules) == [c]
    True
    list(c.submodules) == []
    True

    supports_masking Whether this layer supports computing a mask using compute_mask.
    trainable

    trainable_weights List of all trainable weights tracked by this layer.

    Trainable weights are updated via gradient descent during training.

    training_model_id Identifier of the model.
    variable_dtype Alias of Layer.dtype, the dtype of the weights.
    weights Returns the list of all layer variables/weights.

    Methods

    add_loss

    Add loss tensor(s), potentially dependent on layer inputs.

    Some losses (for instance, activity regularization losses) may be dependent on the inputs passed when calling a layer. Hence, when reusing the same layer on different inputs a and b, some entries in layer.losses may be dependent on a and some on b. This method automatically keeps track of dependencies.

    This method can be used inside a subclassed layer or model's call function, in which case losses should be a Tensor or list of Tensors.

    Example:

    class MyLayer(tf.keras.layers.Layer):
      def call(self, inputs):
        self.add_loss(tf.abs(tf.reduce_mean(inputs)))
        return inputs
    

    The same code works in distributed training: the input to add_loss() is treated like a regularization loss and averaged across replicas by the training loop (both built-in Model.fit() and compliant custom training loops).

    The add_loss method can also be called directly on a Functional Model during construction. In this case, any loss Tensors passed to this Model must be symbolic and be able to be traced back to the model's Inputs. These losses become part of the model's topology and are tracked in get_config.

    Example:

    inputs = tf.keras.Input(shape=(10,))
    x = tf.keras.layers.Dense(10)(inputs)
    outputs = tf.keras.layers.Dense(1)(x)
    model = tf.keras.Model(inputs, outputs)
    # Activity regularization.
    model.add_loss(tf.abs(tf.reduce_mean(x)))
    

    If this is not the case for your loss (if, for example, your loss references a Variable of one of the model's layers), you can wrap your loss in a zero-argument lambda. These losses are not tracked as part of the model's topology since they can't be serialized.

    Example:

    inputs = tf.keras.Input(shape=(10,))
    d = tf.keras.layers.Dense(10)
    x = d(inputs)
    outputs = tf.keras.layers.Dense(1)(x)
    model = tf.keras.Model(inputs, outputs)
    # Weight regularization.
    model.add_loss(lambda: tf.reduce_mean(d.kernel))
    

    Args
    losses Loss tensor, or list/tuple of tensors. Rather than tensors, losses may also be zero-argument callables which create a loss tensor.
    **kwargs Used for backwards compatibility only.

    build

    Builds the model based on input shapes received.

    This is to be used for subclassed models, which do not know at instantiation time what their inputs look like.

    This method only exists for users who want to call model.build() in a standalone way (as a substitute for calling the model on real data to build it). It will never be called by the framework (and thus it will never throw unexpected errors in an unrelated workflow).

    Args
    input_shape Single tuple, TensorShape instance, or list/dict of shapes, where shapes are tuples, integers, or TensorShape instances.

    Raises
    ValueError

    1. In case of invalid user-provided data (not of type tuple, list, TensorShape, or dict).
    2. If the model requires call arguments that are agnostic to the input shapes (positional or keyword arg in call signature).
    3. If not all layers were properly built.
    4. If float type inputs are not supported within the layers.

    In each of these cases, the user should build their model by calling it on real tensor data.

    build_from_config

    Builds the layer's states with the supplied config dict.

    By default, this method calls the build(config["input_shape"]) method, which creates weights based on the layer's input shape in the supplied config. If your config contains other information needed to load the layer's state, you should override this method.

    Args
    config Dict containing the input shape associated with this layer.

    call

    View source

    Inference of the model.

    This method is used for prediction and evaluation of a trained model.

    Args
    inputs Input tensors.
    training Is the model being trained. Always False.

    Returns
    Model predictions.

    call_get_leaves

    View source

    Computes the index of the active leaf in each tree.

    The active leaf is the leave that that receive the example during inference.

    The returned value "leaves[i,j]" is the index of the active leave for the i-th example and the j-th tree. Leaves are indexed by depth first exploration with the negative child visited before the positive one (similarly as "iterate_on_nodes()" iteration). Leaf indices are also available with LeafNode.leaf_idx.

    Args
    inputs Input tensors. Same signature as the model's "call(inputs)".

    Returns
    Index of the active leaf for each tree in the model.

    capabilities

    View source

    Lists the capabilities of the learning algorithm.

    collect_data_step

    View source

    Collect examples e.g. training or validation.

    compile

    View source

    Configure the model for training.

    Unlike for most Keras model, calling "compile" is optional before calling "fit".

    Args
    metrics List of metrics to be evaluated by the model during training and testing.
    weighted_metrics List of metrics to be evaluated and weighted by sample_weight or class_weight during training and testing.
    **kwargs Other arguments passed to compile.

    Raises
    ValueError Invalid arguments.

    compile_from_config

    Compiles the model with the information given in config.

    This method uses the information in the config (optimizer, loss, metrics, etc.) to compile the model.

    Args
    config Dict containing information for compiling the model.

    compute_loss

    Compute the total loss, validate it, and return it.

    Subclasses can optionally override this method to provide custom loss computation logic.

    Example:

    class MyModel(tf.keras.Model):
    
      def __init__(self, *args, **kwargs):
        super(MyModel, self).__init__(*args, **kwargs)
        self.loss_tracker = tf.keras.metrics.Mean(name='loss')
    
      def compute_loss(self, x, y, y_pred, sample_weight):
        loss = tf.reduce_mean(tf.math.squared_difference(y_pred, y))
        loss += tf.add_n(self.losses)
        self.loss_tracker.update_state(loss)
        return loss
    
      def reset_metrics(self):
        self.loss_tracker.reset_states()
    
      @property
      def metrics(self):
        return [self.loss_tracker]
    
    tensors = tf.random.uniform((10, 10)), tf.random.uniform((10,))
    dataset = tf.data.Dataset.from_tensor_slices(tensors).repeat().batch(1)
    
    inputs = tf.keras.layers.Input(shape=(10,), name='my_input')
    outputs = tf.keras.layers.Dense(10)(inputs)
    model = MyModel(inputs, outputs)
    model.add_loss(tf.reduce_sum(outputs))
    
    optimizer = tf.keras.optimizers.SGD()
    model.compile(optimizer, loss='mse', steps_per_execution=10)
    model.fit(dataset, epochs=2, steps_per_epoch=10)
    print('My custom loss: ', model.loss_tracker.result().numpy())
    

    Args
    x Input data.
    y Target data.
    y_pred Predictions returned by the model (output of model(x))
    sample_weight Sample weights for weighting the loss function.

    Returns
    The total loss as a tf.Tensor, or None if no loss results (which is the case when called by Model.test_step).

    compute_mask

    Computes an output mask tensor.

    Args
    inputs Tensor or list of tensors.
    mask Tensor or list of tensors.

    Returns
    None or a tensor (or list of tensors, one per output tensor of the layer).

    compute_metrics

    Update metric states and collect all metrics to be returned.

    Subclasses can optionally override this method to provide custom metric updating and collection logic.

    Example:

    class MyModel(tf.keras.Sequential):
    
      def compute_metrics(self, x, y, y_pred, sample_weight):
    
        # This super call updates `self.compiled_metrics` and returns
        # results for all metrics listed in `self.metrics`.
        metric_results = super(MyModel, self).compute_metrics(
            x, y, y_pred, sample_weight)
    
        # Note that `self.custom_metric` is not listed in `self.metrics`.
        self.custom_metric.update_state(x, y, y_pred, sample_weight)
        metric_results['custom_metric_name'] = self.custom_metric.result()
        return metric_results
    

    Args
    x Input data.
    y Target data.
    y_pred Predictions returned by the model (output of model.call(x))
    sample_weight Sample weights for weighting the loss function.

    Returns
    A dict containing values that will be passed to tf.keras.callbacks.CallbackList.on_train_batch_end(). Typically, the values of the metrics listed in self.metrics are returned. Example: {'loss': 0.2, 'accuracy': 0.7}.

    compute_output_shape

    Computes the output shape of the layer.

    This method will cause the layer's state to be built, if that has not happened before. This requires that the layer will later be used with inputs that match the input shape provided here.

    Args
    input_shape Shape tuple (tuple of integers) or tf.TensorShape, or structure of shape tuples / tf.TensorShape instances (one per output tensor of the layer). Shape tuples can include None for free dimensions, instead of an integer.

    Returns
    A tf.TensorShape instance or structure of tf.TensorShape instances.

    count_params

    Count the total number of scalars composing the weights.

    Returns
    An integer count.

    Raises
    ValueError if the layer isn't yet built (in which case its weights aren't yet defined).

    evaluate

    Returns the loss value & metrics values for the model in test mode.

    Computation is done in batches (see the batch_size arg.)

    Args
    x Input data. It could be:

    • A Numpy array (or array-like), or a list of arrays (in case the model has multiple inputs).
    • A TensorFlow tensor, or a list of tensors (in case the model has multiple inputs).
    • A dict mapping input names to the corresponding array/tensors, if the model has named inputs.
    • A tf.data dataset. Should return a tuple of either (inputs, targets) or (inputs, targets, sample_weights).
    • A generator or keras.utils.Sequence returning (inputs, targets) or (inputs, targets, sample_weights). A more detailed description of unpacking behavior for iterator types (Dataset, generator, Sequence) is given in the Unpacking behavior for iterator-like inputs section of Model.fit.
    y Target data. Like the input data x, it could be either Numpy array(s) or TensorFlow tensor(s). It should be consistent with x (you cannot have Numpy inputs and tensor targets, or inversely). If x is a dataset, generator or keras.utils.Sequence instance, y should not be specified (since targets will be obtained from the iterator/dataset).
    batch_size Integer or None. Number of samples per batch of computation. If unspecified, batch_size will default to 32. Do not specify the batch_size if your data is in the form of a dataset, generators, or keras.utils.Sequence instances (since they generate batches).
    verbose "auto", 0, 1, or 2. Verbosity mode. 0 = silent, 1 = progress bar, 2 = single line. "auto" becomes 1 for most cases, and to 2 when used with ParameterServerStrategy. Note that the progress bar is not particularly useful when logged to a file, so verbose=2 is recommended when not running interactively (e.g. in a production environment). Defaults to 'auto'.
    sample_weight Optional Numpy array of weights for the test samples, used for weighting the loss function. You can either pass a flat (1D) Numpy array with the same length as the input samples (1:1 mapping between weights and samples), or in the case of temporal data, you can pass a 2D array with shape (samples, sequence_length), to apply a different weight to every timestep of every sample. This argument is not supported when x is a dataset, instead pass sample weights as the third element of x.
    steps Integer or None. Total number of steps (batches of samples) before declaring the evaluation round finished. Ignored with the default value of None. If x is a tf.data dataset and steps is None, 'evaluate' will run until the dataset is exhausted. This argument is not supported with array inputs.
    callbacks List of keras.callbacks.Callback instances. List of callbacks to apply during evaluation. See callbacks.
    max_queue_size Integer. Used for generator or keras.utils.Sequence input only. Maximum size for the generator queue. If unspecified, max_queue_size will default to 10.
    workers Integer. Used for generator or keras.utils.Sequence input only. Maximum number of processes to spin up when using process-based threading. If unspecified, workers will default to 1.
    use_multiprocessing Boolean. Used for generator or keras.utils.Sequence input only. If True, use process-based threading. If unspecified, use_multiprocessing will default to False. Note that because this implementation relies on multiprocessing, you should not pass non-pickleable arguments to the generator as they can't be passed easily to children processes.
    return_dict If True, loss and metric results are returned as a dict, with each key being the name of the metric. If False, they are returned as a list.
    **kwargs Unused at this time.

    See the discussion of Unpacking behavior for iterator-like inputs for Model.fit.

    Returns
    Scalar test loss (if the model has a single output and no metrics) or list of scalars (if the model has multiple outputs and/or metrics). The attribute model.metrics_names will give you the display labels for the scalar outputs.

    Raises
    RuntimeError If model.evaluate is wrapped in a tf.function.

    export

    Create a SavedModel artifact for inference (e.g. via TF-Serving).

    This method lets you export a model to a lightweight SavedModel artifact that contains the model's forward pass only (its call() method) and can be served via e.g. TF-Serving. The forward pass is registered under the name serve() (see example below).

    The original code of the model (including any custom layers you may have used) is no longer necessary to reload the artifact -- it is entirely standalone.

    Args
    filepath str or pathlib.Path object. Path where to save the artifact.

    Example:

    # Create the artifact
    model.export("path/to/location")
    
    # Later, in a different process / environment...
    reloaded_artifact = tf.saved_model.load("path/to/location")
    predictions = reloaded_artifact.serve(input_data)
    

    If you would like to customize your serving endpoints, you can use the lower-level keras.export.ExportArchive class. The export() method relies on ExportArchive internally.

    fit

    View source

    Trains the model.

    Local training

    It is recommended to use a Pandas Dataframe dataset and to convert it to a TensorFlow dataset with pd_dataframe_to_tf_dataset():

      pd_dataset = pandas.Dataframe(...)
      tf_dataset = pd_dataframe_to_tf_dataset(dataset, label="my_label")
      model.fit(pd_dataset)
    

    The following dataset formats are supported:

    1. "x" is a tf.data.Dataset containing a tuple "(features, labels)". "features" can be a dictionary a tensor, a list of tensors or a dictionary of tensors (recommended). "labels" is a tensor.

    2. "x" is a tensor, list of tensors or dictionary of tensors containing the input features. "y" is a tensor.

    3. "x" is a numpy-array, list of numpy-arrays or dictionary of numpy-arrays containing the input features. "y" is a numpy-array.

    1. The dataset need to be read exactly once. If you use a TensorFlow dataset, make sure NOT to add a "repeat" operation.
    2. The algorithm does not benefit from shuffling the dataset. If you use a TensorFlow dataset, make sure NOT to add a "shuffle" operation.
    3. The dataset needs to be batched (i.e. with a "batch" operation). However, the number of elements per batch has not impact on the model. Generally, it is recommended to use batches as large as possible as its speeds-up reading the dataset in TensorFlow.

    Input features do not need to be normalized (e.g. dividing numerical values by the variance) or indexed (e.g. replacing categorical string values by an integer). Additionally, missing values can be consumed natively.

    Distributed training

    Some of the learning algorithms will support distributed training with the ParameterServerStrategy.

    In this case, the dataset is read asynchronously in between the workers. The distribution of the training depends on the learning algorithm.

    Like for non-distributed training, the dataset should be read exactly once. The simplest solution is to divide the dataset into different files (i.e. shards) and have each of the worker read a non overlapping subset of shards.

    Currently (to be changed), the validation dataset (if provided) is simply feed to the model.evaluate() method. Therefore, it should satisfy Keras' evaluate API. Notably, for distributed training, the validation dataset should be infinite (i.e. have a repeat operation).

    See https://www.tensorflow.org/decision_forests/distributed_training for more details and examples.

    Here is a single example of distributed training using PSS for both dataset reading and training distribution.

      def dataset_fn(context, paths, training=True):
        ds_path = tf.data.Dataset.from_tensor_slices(paths)
    
    
        if context is not None:
          # Train on at least 2 workers.
          current_worker = tfdf.keras.get_worker_idx_and_num_workers(context)
          assert current_worker.num_workers > 2
    
          # Split the dataset's examples among the workers.
          ds_path = ds_path.shard(
              num_shards=current_worker.num_workers,
              index=current_worker.worker_idx)
    
        def read_csv_file(path):
          numerical = tf.constant([math.nan], dtype=tf.float32)
          categorical_string = tf.constant([""], dtype=tf.string)
          csv_columns = [
              numerical,  # age
              categorical_string,  # workclass
              numerical,  # fnlwgt
              ...
          ]
          column_names = [
            "age", "workclass", "fnlwgt", ...
          ]
          label_name = "label"
          return tf.data.experimental.CsvDataset(path, csv_columns, header=True)
    
        ds_columns = ds_path.interleave(read_csv_file)
    
        def map_features(*columns):
          assert len(column_names) == len(columns)
          features = {column_names[i]: col for i, col in enumerate(columns)}
          label = label_table.lookup(features.pop(label_name))
          return features, label
    
        ds_dataset = ds_columns.map(map_features)
        if not training:
          dataset = dataset.repeat(None)
        ds_dataset = ds_dataset.batch(batch_size)
        return ds_dataset
    
      strategy = tf.distribute.experimental.ParameterServerStrategy(...)
      sharded_train_paths = [list of dataset files]
      with strategy.scope():
        model = DistributedGradientBoostedTreesModel()
        train_dataset = strategy.distribute_datasets_from_function(
          lambda context: dataset_fn(context, sharded_train_paths))
    
        test_dataset = strategy.distribute_datasets_from_function(
          lambda context: dataset_fn(context, sharded_test_paths))
    
      model.fit(sharded_train_paths)
      evaluation = model.evaluate(test_dataset, steps=num_test_examples //
        batch_size)
    

    Args
    x Training dataset (See details above for the supported formats).
    y Label of the training dataset. Only used if "x" does not contains the labels.
    callbacks Callbacks triggered during the training. The training runs in a single epoch, itself run in a single step. Therefore, callback logic can be called equivalently before/after the fit function.
    verbose Verbosity mode. 0 = silent, 1 = small details, 2 = full details.
    validation_steps Number of steps in the evaluation dataset when evaluating the trained model with model.evaluate(). If not specified, evaluates the model on the entire dataset (generally recommended; not yet supported for distributed datasets).
    validation_data Validation dataset. If specified, the learner might use this dataset to help training e.g. early stopping.
    sample_weight Training weights. Note: training weights can also be provided as the third output in a tf.data.Dataset e.g. (features, label, weights).
    steps_per_epoch [Parameter will be removed] Number of training batch to load before training the model. Currently, only supported for distributed training.
    class_weight For binary classification only. Mapping class indices (integers) to a weight (float) value. Only available for non-Distributed training. For maximum compatibility, feed example weights through the tf.data.Dataset or using the weight argument of pd_dataframe_to_tf_dataset.
    **kwargs Extra arguments passed to the core keras model's fit. Note that not all keras' model fit arguments are supported.

    Returns
    A History object. Its History.history attribute is not yet implemented for decision forests algorithms, and will return empty. All other fields are filled as usual for Keras.Mode.fit().

    fit_on_dataset_path

    View source

    Trains the model on a dataset stored on disk.

    This solution is generally more efficient and easier than loading the dataset with a tf.Dataset both for local and distributed training.

    Usage example

    Local training

    model = keras.GradientBoostedTreesModel()
    model.fit_on_dataset_path(
      train_path="/path/to/dataset.csv",
      label_key="label",
      dataset_format="csv")
    model.save("/model/path")
    

    Distributed training

    with tf.distribute.experimental.ParameterServerStrategy(...).scope():
      model = model = keras.DistributedGradientBoostedTreesModel()
    model.fit_on_dataset_path(
      train_path="/path/to/dataset@10",
      label_key="label",
      dataset_format="tfrecord+tfe")
    model.save("/model/path")
    

    Args
    train_path Path to the training dataset. Supports comma separated files, shard and glob notation.
    label_key Name of the label column.
    weight_key Name of the weighing column.
    valid_path Path to the validation dataset. If not provided, or if the learning algorithm does not supports/needs a validation dataset, valid_path is ignored.
    dataset_format Format of the dataset. Should be one of the registered dataset format (see User Manual for more details). The format "csv" is always available but it is generally only suited for small datasets.
    max_num_scanned_rows_to_accumulate_statistics Maximum number of examples to scan to determine the statistics of the features (i.e. the dataspec, e.g. mean value, dictionaries). (Currently) the "first" examples of the dataset are scanned (e.g. the first examples of the dataset is a single file). Therefore, it is important that the sampled dataset is relatively uniformly sampled, notably the scanned examples should contains all the possible categorical values (otherwise the not seen value will be treated as out-of-vocabulary). If set to None, the entire dataset is scanned. This parameter has no effect if the dataset is stored in a format that already contains those values.
    try_resume_training If true, tries to resume training from the model checkpoint stored in the temp_directory directory. If temp_directory does not contain any model checkpoint, start the training from the start. Works in the following three situations: (1) The training was interrupted by the user (e.g. ctrl+c). (2) the training job was interrupted (e.g. rescheduling), ond (3) the hyper-parameter of the model were changed such that an initially completed training is now incomplete (e.g. increasing the number of trees).
    input_model_signature_fn A lambda that returns the (Dense,Sparse,Ragged)TensorSpec (or structure of TensorSpec e.g. dictionary, list) corresponding to input signature of the model. If not specified, the input model signature is created by build_default_input_model_signature. For example, specify input_model_signature_fn if an numerical input feature (which is consumed as DenseTensorSpec(float32) by default) will be feed differently (e.g. RaggedTensor(int64)).
    num_io_threads Number of threads to use for IO operations e.g. reading a dataset from disk. Increasing this value can speed-up IO operations when IO operations are either latency or cpu bounded.

    Returns
    A History object. Its History.history attribute is not yet implemented for decision forests algorithms, and will return empty. All other fields are filled as usual for Keras.Mode.fit().

    from_config

    Creates a layer from its config.

    This method is the reverse of get_config, capable of instantiating the same layer from the config dictionary. It does not handle layer connectivity (handled by Network), nor weights (handled by set_weights).

    Args
    config A Python dictionary, typically the output of get_config.

    Returns
    A layer instance.

    get_build_config

    Returns a dictionary with the layer's input shape.

    This method returns a config dict that can be used by build_from_config(config) to create all states (e.g. Variables and Lookup tables) needed by the layer.

    By default, the config only contains the input shape that the layer was built with. If you're writing a custom layer that creates state in an unusual way, you should override this method to make sure this state is already created when TF-Keras attempts to load its value upon model loading.

    Returns
    A dict containing the input shape associated with the layer.

    get_compile_config

    Returns a serialized config with information for compiling the model.

    This method returns a config dictionary containing all the information (optimizer, loss, metrics, etc.) with which the model was compiled.

    Returns
    A dict containing information for compiling the model.

    get_config

    View source

    Not supported by TF-DF, returning empty directory to avoid warnings.

    get_layer

    Retrieves a layer based on either its name (unique) or index.

    If name and index are both provided, index will take precedence. Indices are based on order of horizontal graph traversal (bottom-up).

    Args
    name String, name of layer.
    index Integer, index of layer.

    Returns
    A layer instance.

    get_metrics_result

    Returns the model's metrics values as a dict.

    If any of the metric result is a dict (containing multiple metrics), each of them gets added to the top level returned dict of this method.

    Returns
    A dict containing values of the metrics listed in self.metrics.
    Example {'loss': 0.2, 'accuracy': 0.7}.

    get_weight_paths

    Retrieve all the variables and their paths for the model.

    The variable path (string) is a stable key to identify a tf.Variable instance owned by the model. It can be used to specify variable-specific configurations (e.g. DTensor, quantization) from a global view.

    This method returns a dict with weight object paths as keys and the corresponding tf.Variable instances as values.

    Note that if the model is a subclassed model and the weights haven't been initialized, an empty dict will be returned.

    Returns
    A dict where keys are variable paths and values are tf.Variable instances.

    Example:

    class SubclassModel(tf.keras.Model):
    
      def __init__(self, name=None):
        super().__init__(name=name)
        self.d1 = tf.keras.layers.Dense(10)
        self.d2 = tf.keras.layers.Dense(20)
    
      def call(self, inputs):
        x = self.d1(inputs)
        return self.d2(x)
    
    model = SubclassModel()
    model(tf.zeros((10, 10)))
    weight_paths = model.get_weight_paths()
    # weight_paths:
    # {
    #    'd1.kernel': model.d1.kernel,
    #    'd1.bias': model.d1.bias,
    #    'd2.kernel': model.d2.kernel,
    #    'd2.bias': model.d2.bias,
    # }
    
    # Functional model
    inputs = tf.keras.Input((10,), batch_size=10)
    x = tf.keras.layers.Dense(20, name='d1')(inputs)
    output = tf.keras.layers.Dense(30, name='d2')(x)
    model = tf.keras.Model(inputs, output)
    d1 = model.layers[1]
    d2 = model.layers[2]
    weight_paths = model.get_weight_paths()
    # weight_paths:
    # {
    #    'd1.kernel': d1.kernel,
    #    'd1.bias': d1.bias,
    #    'd2.kernel': d2.kernel,
    #    'd2.bias': d2.bias,
    # }
    

    get_weights

    Retrieves the weights of the model.

    Returns
    A flat list of Numpy arrays.

    load_own_variables

    Loads the state of the layer.

    You can override this method to take full control of how the state of the layer is loaded upon calling keras.models.load_model().

    Args
    store Dict from which the state of the model will be loaded.

    load_weights

    View source

    No-op for TensorFlow Decision Forests models.

    load_weights is not supported by TensorFlow Decision Forests models. To save and restore a model, use the SavedModel API i.e. model.save(...) and tf_keras.models.load_model(...). To resume the training of an existing model, create the model with try_resume_training=True (default value) and with a similar temp_directory argument. See documentation of try_resume_training for more details.

    Args
    *args Passed through to base keras.Model implemenation.
    **kwargs Passed through to base keras.Model implemenation.

    make_inspector

    View source

    Creates an inspector to access the internal model structure.

    Usage example:

    inspector = model.make_inspector()
    print(inspector.num_trees())
    print(inspector.variable_importances())
    

    Args
    index Index of the sub-model. Only used for multitask models.

    Returns
    A model inspector.

    make_predict_function

    View source

    Prediction of the model (!= evaluation).

    make_test_function

    View source

    Predictions for evaluation.

    make_train_function

    Creates a function that executes one step of training.

    This method can be overridden to support custom training logic. This method is called by Model.fit and Model.train_on_batch.

    Typically, this method directly controls tf.function and tf.distribute.Strategy settings, and delegates the actual training logic to Model.train_step.

    This function is cached the first time Model.fit or Model.train_on_batch is called. The cache is cleared whenever Model.compile is called. You can skip the cache and generate again the function with force=True.

    Args
    force Whether to regenerate the train function and skip the cached function if available.

    Returns
    Function. The function created by this method should accept a tf.data.Iterator, and return a dict containing values that will be passed to tf.keras.Callbacks.on_train_batch_end, such as {'loss': 0.2, 'accuracy': 0.7}.

    predefined_hyperparameters

    View source

    Returns a better than default set of hyper-parameters.

    They can be used directly with the hyperparameter_template argument of the model constructor.

    These hyper-parameters outperform the default hyper-parameters (either generally or in specific scenarios). Like default hyper-parameters, existing pre-defined hyper-parameters cannot change.

    predict

    Generates output predictions for the input samples.

    Computation is done in batches. This method is designed for batch processing of large numbers of inputs. It is not intended for use inside of loops that iterate over your data and process small numbers of inputs at a time.

    For small numbers of inputs that fit in one batch, directly use __call__() for faster execution, e.g., model(x), or model(x, training=False) if you have layers such as tf.keras.layers.BatchNormalization that behave differently during inference. You may pair the individual model call with a tf.function for additional performance inside your inner loop. If you need access to numpy array values instead of tensors after your model call, you can use tensor.numpy() to get the numpy array value of an eager tensor.

    Also, note the fact that test loss is not affected by regularization layers like noise and dropout.

    Args
    x Input samples. It could be:

    • A Numpy array (or array-like), or a list of arrays (in case the model has multiple inputs).
    • A TensorFlow tensor, or a list of tensors (in case the model has multiple inputs).
    • A tf.data dataset.
    • A generator or keras.utils.Sequence instance. A more detailed description of unpacking behavior for iterator types (Dataset, generator, Sequence) is given in the Unpacking behavior for iterator-like inputs section of Model.fit.
    batch_size Integer or None. Number of samples per batch. If unspecified, batch_size will default to 32. Do not specify the batch_size if your data is in the form of dataset, generators, or keras.utils.Sequence instances (since they generate batches).
    verbose "auto", 0, 1, or 2. Verbosity mode. 0 = silent, 1 = progress bar, 2 = single line. "auto" becomes 1 for most cases, and to 2 when used with ParameterServerStrategy. Note that the progress bar is not particularly useful when logged to a file, so verbose=2 is recommended when not running interactively (e.g. in a production environment). Defaults to 'auto'.
    steps Total number of steps (batches of samples) before declaring the prediction round finished. Ignored with the default value of None. If x is a tf.data dataset and steps is None, predict() will run until the input dataset is exhausted.
    callbacks List of keras.callbacks.Callback instances. List of callbacks to apply during prediction. See callbacks.
    max_queue_size Integer. Used for generator or keras.utils.Sequence input only. Maximum size for the generator queue. If unspecified, max_queue_size will default to 10.
    workers Integer. Used for generator or keras.utils.Sequence input only. Maximum number of processes to spin up when using process-based threading. If unspecified, workers will default to 1.
    use_multiprocessing Boolean. Used for generator or keras.utils.Sequence input only. If True, use process-based threading. If unspecified, use_multiprocessing will default to False. Note that because this implementation relies on multiprocessing, you should not pass non-pickleable arguments to the generator as they can't be passed easily to children processes.

    See the discussion of Unpacking behavior for iterator-like inputs for Model.fit. Note that Model.predict uses the same interpretation rules as Model.fit and Model.evaluate, so inputs must be unambiguous for all three methods.

    Returns
    Numpy array(s) of predictions.

    Raises
    RuntimeError If model.predict is wrapped in a tf.function.
    ValueError In case of mismatch between the provided input data and the model's expectations, or in case a stateful model receives a number of samples that is not a multiple of the batch size.

    predict_get_leaves

    View source

    Gets the index of the active leaf of each tree.

    The active leaf is the leave that that receive the example during inference.

    The returned value "leaves[i,j]" is the index of the active leave for the i-th example and the j-th tree. Leaves are indexed by depth first exploration with the negative child visited before the positive one (similarly as "iterate_on_nodes()" iteration). Leaf indices are also available with LeafNode.leaf_idx.

    Args
    x Input samples as a tf.data.Dataset.

    Returns
    Index of the active leaf for each tree in the model.

    predict_on_batch

    Returns predictions for a single batch of samples.

    Args
    x Input data. It could be:

    • A Numpy array (or array-like), or a list of arrays (in case the model has multiple inputs).
    • A TensorFlow tensor, or a list of tensors (in case the model has multiple inputs).

    Returns
    Numpy array(s) of predictions.

    Raises
    RuntimeError If model.predict_on_batch is wrapped in a tf.function.

    predict_step

    The logic for one inference step.

    This method can be overridden to support custom inference logic. This method is called by Model.make_predict_function.

    This method should contain the mathematical logic for one step of inference. This typically includes the forward pass.

    Configuration details for how this logic is run (e.g. tf.function and tf.distribute.Strategy settings), should be left to Model.make_predict_function, which can also be overridden.

    Args
    data A nested structure of Tensors.

    Returns
    The result of one inference step, typically the output of calling the Model on data.

    ranking_group

    View source

    reset_metrics

    Resets the state of all the metrics in the model.

    Examples:

    inputs = tf.keras.layers.Input(shape=(3,))
    outputs = tf.keras.layers.Dense(2)(inputs)
    model = tf.keras.models.Model(inputs=inputs, outputs=outputs)
    model.compile(optimizer="Adam", loss="mse", metrics=["mae"])
    x = np.random.random((2, 3))
    y = np.random.randint(0, 2, (2, 2))
    _ = model.fit(x, y, verbose=0)
    assert all(float(m.result()) for m in model.metrics)
    model.reset_metrics()
    assert all(float(m.result()) == 0 for m in model.metrics)

    reset_states

    save

    View source

    Saves the model as a TensorFlow SavedModel.

    The exported SavedModel contains a standalone Yggdrasil Decision Forests model in the "assets" sub-directory. The Yggdrasil model can be used directly using the Yggdrasil API. However, this model does not contain the "preprocessing" layer (if any).

    Args
    filepath Path to the output model.
    overwrite If true, override an already existing model. If false, raise an error if a model already exist.
    **kwargs Arguments passed to the core keras model's save.

    save_own_variables

    Saves the state of the layer.

    You can override this method to take full control of how the state of the layer is saved upon calling model.save().

    Args
    store Dict where the state of the model will be saved.

    save_spec

    Returns the tf.TensorSpec of call args as a tuple (args, kwargs).

    This value is automatically defined after calling the model for the first time. Afterwards, you can use it when exporting the model for serving:

    model = tf.keras.Model(...)
    
    @tf.function
    def serve(*args, **kwargs):
      outputs = model(*args, **kwargs)
      # Apply postprocessing steps, or add additional outputs.
      ...
      return outputs
    
    # arg_specs is `[tf.TensorSpec(...), ...]`. kwarg_specs, in this
    # example, is an empty dict since functional models do not use keyword
    # arguments.
    arg_specs, kwarg_specs = model.save_spec()
    
    model.save(path, signatures={
      'serving_default': serve.get_concrete_function(*arg_specs,
                                                     **kwarg_specs)
    })
    

    Args
    dynamic_batch Whether to set the batch sizes of all the returned tf.TensorSpec to None. (Note that when defining functional or Sequential models with tf.keras.Input([...], batch_size=X), the batch size will always be preserved). Defaults to True.

    Returns
    If the model inputs are defined, returns a tuple (args, kwargs). All elements in args and kwargs are tf.TensorSpec. If the model inputs are not defined, returns None. The model inputs are automatically set when calling the model, model.fit, model.evaluate or model.predict.

    save_weights

    Saves all layer weights.

    Either saves in HDF5 or in TensorFlow format based on the save_format argument.

    When saving in HDF5 format, the weight file has:

    • layer_names (attribute), a list of strings (ordered names of model layers).
    • For every layer, a group named layer.name
      • For every such layer group, a group attribute weight_names, a list of strings (ordered names of weights tensor of the layer).
      • For every weight in the layer, a dataset storing the weight value, named after the weight tensor.

    When saving in TensorFlow format, all objects referenced by the network are saved in the same format as tf.train.Checkpoint, including any Layer instances or Optimizer instances assigned to object attributes. For networks constructed from inputs and outputs using tf.keras.Model(inputs, outputs), Layer instances used by the network are tracked/saved automatically. For user-defined classes which inherit from tf.keras.Model, Layer instances must be assigned to object attributes, typically in the constructor. See the documentation of tf.train.Checkpoint and tf.keras.Model for details.

    While the formats are the same, do not mix save_weights and tf.train.Checkpoint. Checkpoints saved by Model.save_weights should be loaded using Model.load_weights. Checkpoints saved using tf.train.Checkpoint.save should be restored using the corresponding tf.train.Checkpoint.restore. Prefer tf.train.Checkpoint over save_weights for training checkpoints.

    The TensorFlow format matches objects and variables by starting at a root object, self for save_weights, and greedily matching attribute names. For Model.save this is the Model, and for Checkpoint.save this is the Checkpoint even if the Checkpoint has a model attached. This means saving a tf.keras.Model using save_weights and loading into a tf.train.Checkpoint with a Model attached (or vice versa) will not match the Model's variables. See the guide to training checkpoints for details on the TensorFlow format.

    Args
    filepath String or PathLike, path to the file to save the weights to. When saving in TensorFlow format, this is the prefix used for checkpoint files (multiple files are generated). Note that the '.h5' suffix causes weights to be saved in HDF5 format.
    overwrite Whether to silently overwrite any existing file at the target location, or provide the user with a manual prompt.
    save_format Either 'tf' or 'h5'. A filepath ending in '.h5' or '.keras' will default to HDF5 if save_format is None. Otherwise, None becomes 'tf'. Defaults to None.
    options Optional tf.train.CheckpointOptions object that specifies options for saving weights.

    Raises
    ImportError If h5py is not available when attempting to save in HDF5 format.

    set_weights

    Sets the weights of the layer, from NumPy arrays.

    The weights of a layer represent the state of the layer. This function sets the weight values from numpy arrays. The weight values should be passed in the order they are created by the layer. Note that the layer's weights must be instantiated before calling this function, by calling the layer.

    For example, a Dense layer returns a list of two values: the kernel matrix and the bias vector. These can be used to set the weights of another Dense layer:

    layer_a = tf.keras.layers.Dense(1,
      kernel_initializer=tf.constant_initializer(1.))
    a_out = layer_a(tf.convert_to_tensor([[1., 2., 3.]]))
    layer_a.get_weights()
    [array([[1.],
           [1.],
           [1.]], dtype=float32), array([0.], dtype=float32)]
    layer_b = tf.keras.layers.Dense(1,
      kernel_initializer=tf.constant_initializer(2.))
    b_out = layer_b(tf.convert_to_tensor([[10., 20., 30.]]))
    layer_b.get_weights()
    [array([[2.],
           [2.],
           [2.]], dtype=float32), array([0.], dtype=float32)]
    layer_b.set_weights(layer_a.get_weights())
    layer_b.get_weights()
    [array([[1.],
           [1.],
           [1.]], dtype=float32), array([0.], dtype=float32)]

    Args
    weights a list of NumPy arrays. The number of arrays and their shape must match number of the dimensions of the weights of the layer (i.e. it should match the output of get_weights).

    Raises
    ValueError If the provided weights list does not match the layer's specifications.

    summary

    View source

    Shows information about the model.

    support_distributed_training

    View source

    test_on_batch

    Test the model on a single batch of samples.

    Args
    x Input data. It could be:

    • A Numpy array (or array-like), or a list of arrays (in case the model has multiple inputs).
    • A TensorFlow tensor, or a list of tensors (in case the model has multiple inputs).
    • A dict mapping input names to the corresponding array/tensors, if the model has named inputs.
    y Target data. Like the input data x, it could be either Numpy array(s) or TensorFlow tensor(s). It should be consistent with x (you cannot have Numpy inputs and tensor targets, or inversely).
    sample_weight Optional array of the same length as x, containing weights to apply to the model's loss for each sample. In the case of temporal data, you can pass a 2D array with shape (samples, sequence_length), to apply a different weight to every timestep of every sample.
    reset_metrics If True, the metrics returned will be only for this batch. If False, the metrics will be statefully accumulated across batches.
    return_dict If True, loss and metric results are returned as a dict, with each key being the name of the metric. If False, they are returned as a list.

    Returns
    Scalar test loss (if the model has a single output and no metrics) or list of scalars (if the model has multiple outputs and/or metrics). The attribute model.metrics_names will give you the display labels for the scalar outputs.

    Raises
    RuntimeError If model.test_on_batch is wrapped in a tf.function.

    test_step

    The logic for one evaluation step.

    This method can be overridden to support custom evaluation logic. This method is called by Model.make_test_function.

    This function should contain the mathematical logic for one step of evaluation. This typically includes the forward pass, loss calculation, and metrics updates.

    Configuration details for how this logic is run (e.g. tf.function and tf.distribute.Strategy settings), should be left to Model.make_test_function, which can also be overridden.

    Args
    data A nested structure of Tensors.

    Returns
    A dict containing values that will be passed to tf.keras.callbacks.CallbackList.on_train_batch_end. Typically, the values of the Model's metrics are returned.

    to_json

    Returns a JSON string containing the network configuration.

    To load a network from a JSON save file, use keras.models.model_from_json(json_string, custom_objects={}).

    Args
    **kwargs Additional keyword arguments to be passed to *json.dumps().

    Returns
    A JSON string.

    to_yaml

    Returns a yaml string containing the network configuration.

    To load a network from a yaml save file, use keras.models.model_from_yaml(yaml_string, custom_objects={}).

    custom_objects should be a dictionary mapping the names of custom losses / layers / etc to the corresponding functions / classes.

    Args
    **kwargs Additional keyword arguments to be passed to yaml.dump().

    Returns
    A YAML string.

    Raises
    RuntimeError announces that the method poses a security risk

    train_on_batch

    View source

    No supported for Tensorflow Decision Forests models.

    Decision forests are not trained in batches the same way neural networks are. To avoid confusion, train_on_batch is disabled.

    Args
    *args Ignored
    **kwargs Ignored.

    train_step

    View source

    Collects training examples.

    uplift_treatment

    View source

    valid_step

    View source

    Collects validation examples.

    with_name_scope

    Decorator to automatically enter the module name scope.

    class MyModule(tf.Module):
      @tf.Module.with_name_scope
      def __call__(self, x):
        if not hasattr(self, &#x27;w'):
          self.w = tf.Variable(tf.random.normal([x.shape[1], 3]))
        return tf.matmul(x, self.w)

    Using the above module would produce tf.Variables and tf.Tensors whose names included the module name:

    mod = MyModule()
    mod(tf.ones([1, 2]))
    <tf.Tensor: shape=(1, 3), dtype=float32, numpy=..., dtype=float32)>
    mod.w
    <tf.Variable &#x27;my_module/Variable:0' shape=(2, 3) dtype=float32,
    numpy=..., dtype=float32)>

    Args
    method The method to wrap.

    Returns
    The original method wrapped such that it enters the module's name scope.

    yggdrasil_model_path_tensor

    View source

    Gets the path to yggdrasil model, if available.

    The effective path can be obtained with:

    yggdrasil_model_path_tensor().numpy().decode("utf-8")
    

    Args
    multitask_model_index Index of the sub-model. Only used for multitask models.

    Returns
    Path to the Yggdrasil model.

    yggdrasil_model_prefix

    View source

    Gets the prefix of the internal yggdrasil model.

    __call__