Migration examples: Canned Estimators

View on TensorFlow.org Run in Google Colab View source on GitHub Download notebook

Canned (or Premade) Estimators have traditionally been used in TensorFlow 1 as quick and easy ways to train models for a variety of typical use cases. TensorFlow 2 provides straightforward approximate substitutes for a number of them by way of Keras models. For those canned estimators that do not have built-in TensorFlow 2 substitutes, you can still build your own replacement fairly easily.

This guide will walk you through a few examples of direct equivalents and custom substitutions to demonstrate how TensorFlow 1's tf.estimator-derived models can be migrated to TensorFlow 2 with Keras.

Namely, this guide includes examples for migrating:

A common precursor to the training of a model is feature preprocessing, which is done for TensorFlow 1 Estimator models with tf.feature_column. For more information on feature preprocessing in TensorFlow 2, see this guide on migrating from feature columns to the Keras preprocessing layers API.

Setup

Start with a couple of necessary TensorFlow imports,

pip install tensorflow_decision_forests
import keras
import pandas as pd
import tensorflow as tf
import tensorflow.compat.v1 as tf1
import tensorflow_decision_forests as tfdf
2022-12-14 03:47:10.052663: W tensorflow/compiler/xla/stream_executor/platform/default/dso_loader.cc:64] Could not load dynamic library 'libnvinfer.so.7'; dlerror: libnvinfer.so.7: cannot open shared object file: No such file or directory
2022-12-14 03:47:10.052759: W tensorflow/compiler/xla/stream_executor/platform/default/dso_loader.cc:64] Could not load dynamic library 'libnvinfer_plugin.so.7'; dlerror: libnvinfer_plugin.so.7: cannot open shared object file: No such file or directory
2022-12-14 03:47:10.052768: W tensorflow/compiler/tf2tensorrt/utils/py_utils.cc:38] TF-TRT Warning: Cannot dlopen some TensorRT libraries. If you would like to use Nvidia GPU with TensorRT, please make sure the missing libraries mentioned above are installed properly.

prepare some simple data for demonstration from the standard Titanic dataset,

x_train = pd.read_csv('https://storage.googleapis.com/tf-datasets/titanic/train.csv')
x_eval = pd.read_csv('https://storage.googleapis.com/tf-datasets/titanic/eval.csv')
x_train['sex'].replace(('male', 'female'), (0, 1), inplace=True)
x_eval['sex'].replace(('male', 'female'), (0, 1), inplace=True)

x_train['alone'].replace(('n', 'y'), (0, 1), inplace=True)
x_eval['alone'].replace(('n', 'y'), (0, 1), inplace=True)

x_train['class'].replace(('First', 'Second', 'Third'), (1, 2, 3), inplace=True)
x_eval['class'].replace(('First', 'Second', 'Third'), (1, 2, 3), inplace=True)

x_train.drop(['embark_town', 'deck'], axis=1, inplace=True)
x_eval.drop(['embark_town', 'deck'], axis=1, inplace=True)

y_train = x_train.pop('survived')
y_eval = x_eval.pop('survived')
# Data setup for TensorFlow 1 with `tf.estimator`
def _input_fn():
  return tf1.data.Dataset.from_tensor_slices((dict(x_train), y_train)).batch(32)


def _eval_input_fn():
  return tf1.data.Dataset.from_tensor_slices((dict(x_eval), y_eval)).batch(32)


FEATURE_NAMES = [
    'age', 'fare', 'sex', 'n_siblings_spouses', 'parch', 'class', 'alone'
]

feature_columns = []
for fn in FEATURE_NAMES:
  feat_col = tf1.feature_column.numeric_column(fn, dtype=tf.float32)
  feature_columns.append(feat_col)

and create a method to instantiate a simplistic sample optimizer to use with various TensorFlow 1 Estimator and TensorFlow 2 Keras models.

def create_sample_optimizer(tf_version):
  if tf_version == 'tf1':
    optimizer = lambda: tf.keras.optimizers.legacy.Ftrl(
        l1_regularization_strength=0.001,
        learning_rate=tf1.train.exponential_decay(
            learning_rate=0.1,
            global_step=tf1.train.get_global_step(),
            decay_steps=10000,
            decay_rate=0.9))
  elif tf_version == 'tf2':
    optimizer = tf.keras.optimizers.legacy.Ftrl(
        l1_regularization_strength=0.001,
        learning_rate=tf.keras.optimizers.schedules.ExponentialDecay(
            initial_learning_rate=0.1, decay_steps=10000, decay_rate=0.9))
  return optimizer

Example 1: Migrating from LinearEstimator

TensorFlow 1: Using LinearEstimator

In TensorFlow 1, you can use tf.estimator.LinearEstimator to create a baseline linear model for regression and classification problems.

linear_estimator = tf.estimator.LinearEstimator(
    head=tf.estimator.BinaryClassHead(),
    feature_columns=feature_columns,
    optimizer=create_sample_optimizer('tf1'))
INFO:tensorflow:Using default config.
WARNING:tensorflow:Using temporary folder as model directory: /tmpfs/tmp/tmpzcvqvtaa
INFO:tensorflow:Using config: {'_model_dir': '/tmpfs/tmp/tmpzcvqvtaa', '_tf_random_seed': None, '_save_summary_steps': 100, '_save_checkpoints_steps': None, '_save_checkpoints_secs': 600, '_session_config': allow_soft_placement: true
graph_options {
  rewrite_options {
    meta_optimizer_iterations: ONE
  }
}
, '_keep_checkpoint_max': 5, '_keep_checkpoint_every_n_hours': 10000, '_log_step_count_steps': 100, '_train_distribute': None, '_device_fn': None, '_protocol': None, '_eval_distribute': None, '_experimental_distribute': None, '_experimental_max_worker_delay_secs': None, '_session_creation_timeout_secs': 7200, '_checkpoint_save_graph_def': True, '_service': None, '_cluster_spec': ClusterSpec({}), '_task_type': 'worker', '_task_id': 0, '_global_id_in_cluster': 0, '_master': '', '_evaluation_master': '', '_is_chief': True, '_num_ps_replicas': 0, '_num_worker_replicas': 1}
linear_estimator.train(input_fn=_input_fn, steps=100)
linear_estimator.evaluate(input_fn=_eval_input_fn, steps=10)
WARNING:tensorflow:From /tmpfs/src/tf_docs_env/lib/python3.9/site-packages/tensorflow/python/training/training_util.py:396: Variable.initialized_value (from tensorflow.python.ops.variables) is deprecated and will be removed in a future version.
Instructions for updating:
Use Variable.read_value. Variables in 2.X are initialized automatically both in eager and graph (inside tf.defun) contexts.
INFO:tensorflow:Calling model_fn.
WARNING:tensorflow:From /tmpfs/src/tf_docs_env/lib/python3.9/site-packages/keras/optimizers/optimizer_v2/ftrl.py:173: calling Constant.__init__ (from tensorflow.python.ops.init_ops) with dtype is deprecated and will be removed in a future version.
Instructions for updating:
Call initializer instance with the dtype argument instead of passing it to the constructor
INFO:tensorflow:Done calling model_fn.
INFO:tensorflow:Create CheckpointSaverHook.
INFO:tensorflow:Graph was finalized.
INFO:tensorflow:Running local_init_op.
INFO:tensorflow:Done running local_init_op.
INFO:tensorflow:Calling checkpoint listeners before saving checkpoint 0...
INFO:tensorflow:Saving checkpoints for 0 into /tmpfs/tmp/tmpzcvqvtaa/model.ckpt.
INFO:tensorflow:Calling checkpoint listeners after saving checkpoint 0...
INFO:tensorflow:loss = 0.6931472, step = 0
INFO:tensorflow:Calling checkpoint listeners before saving checkpoint 20...
INFO:tensorflow:Saving checkpoints for 20 into /tmpfs/tmp/tmpzcvqvtaa/model.ckpt.
INFO:tensorflow:Calling checkpoint listeners after saving checkpoint 20...
INFO:tensorflow:Loss for final step: 0.552688.
INFO:tensorflow:Calling model_fn.
INFO:tensorflow:Done calling model_fn.
INFO:tensorflow:Starting evaluation at 2022-12-14T03:47:17
INFO:tensorflow:Graph was finalized.
INFO:tensorflow:Restoring parameters from /tmpfs/tmp/tmpzcvqvtaa/model.ckpt-20
INFO:tensorflow:Running local_init_op.
INFO:tensorflow:Done running local_init_op.
INFO:tensorflow:Evaluation [1/10]
INFO:tensorflow:Evaluation [2/10]
INFO:tensorflow:Evaluation [3/10]
INFO:tensorflow:Evaluation [4/10]
INFO:tensorflow:Evaluation [5/10]
INFO:tensorflow:Evaluation [6/10]
INFO:tensorflow:Evaluation [7/10]
INFO:tensorflow:Evaluation [8/10]
INFO:tensorflow:Evaluation [9/10]
INFO:tensorflow:Inference Time : 0.51355s
INFO:tensorflow:Finished evaluation at 2022-12-14-03:47:18
INFO:tensorflow:Saving dict for global step 20: accuracy = 0.70075756, accuracy_baseline = 0.625, auc = 0.75472915, auc_precision_recall = 0.65362054, average_loss = 0.5759378, global_step = 20, label/mean = 0.375, loss = 0.5704811, precision = 0.6388889, prediction/mean = 0.41331065, recall = 0.46464646
INFO:tensorflow:Saving 'checkpoint_path' summary for global step 20: /tmpfs/tmp/tmpzcvqvtaa/model.ckpt-20
{'accuracy': 0.70075756,
 'accuracy_baseline': 0.625,
 'auc': 0.75472915,
 'auc_precision_recall': 0.65362054,
 'average_loss': 0.5759378,
 'label/mean': 0.375,
 'loss': 0.5704811,
 'precision': 0.6388889,
 'prediction/mean': 0.41331065,
 'recall': 0.46464646,
 'global_step': 20}

TensorFlow 2: Using Keras LinearModel

In TensorFlow 2, you can create an instance of the Keras tf.compat.v1.keras.models.LinearModel which is the substitute to the tf.estimator.LinearEstimator. The tf.compat.v1.keras path is used to signify that the pre-made model exists for compatibility.

linear_model = tf.compat.v1.keras.experimental.LinearModel()
linear_model.compile(loss='mse', optimizer=create_sample_optimizer('tf2'), metrics=['accuracy'])
linear_model.fit(x_train, y_train, epochs=10)
linear_model.evaluate(x_eval, y_eval, return_dict=True)
Epoch 1/10
20/20 [==============================] - 0s 2ms/step - loss: 3.1981 - accuracy: 0.6380
Epoch 2/10
20/20 [==============================] - 0s 2ms/step - loss: 0.3036 - accuracy: 0.6635
Epoch 3/10
20/20 [==============================] - 0s 2ms/step - loss: 0.2018 - accuracy: 0.6858
Epoch 4/10
20/20 [==============================] - 0s 2ms/step - loss: 0.1875 - accuracy: 0.7113
Epoch 5/10
20/20 [==============================] - 0s 2ms/step - loss: 0.2024 - accuracy: 0.7273
Epoch 6/10
20/20 [==============================] - 0s 2ms/step - loss: 0.2423 - accuracy: 0.7337
Epoch 7/10
20/20 [==============================] - 0s 2ms/step - loss: 0.1737 - accuracy: 0.7719
Epoch 8/10
20/20 [==============================] - 0s 2ms/step - loss: 0.1727 - accuracy: 0.7879
Epoch 9/10
20/20 [==============================] - 0s 2ms/step - loss: 0.1900 - accuracy: 0.7671
Epoch 10/10
20/20 [==============================] - 0s 2ms/step - loss: 0.1667 - accuracy: 0.7974
9/9 [==============================] - 0s 2ms/step - loss: 0.1807 - accuracy: 0.7386
{'loss': 0.18073849380016327, 'accuracy': 0.7386363744735718}

Example 2: Migrating from DNNEstimator

TensorFlow 1: Using DNNEstimator

In TensorFlow 1, you can use tf.estimator.DNNEstimator to create a baseline deep neural network (DNN) model for regression and classification problems.

dnn_estimator = tf.estimator.DNNEstimator(
    head=tf.estimator.BinaryClassHead(),
    feature_columns=feature_columns,
    hidden_units=[128],
    activation_fn=tf.nn.relu,
    optimizer=create_sample_optimizer('tf1'))
INFO:tensorflow:Using default config.
WARNING:tensorflow:Using temporary folder as model directory: /tmpfs/tmp/tmpmdq9lsrm
INFO:tensorflow:Using config: {'_model_dir': '/tmpfs/tmp/tmpmdq9lsrm', '_tf_random_seed': None, '_save_summary_steps': 100, '_save_checkpoints_steps': None, '_save_checkpoints_secs': 600, '_session_config': allow_soft_placement: true
graph_options {
  rewrite_options {
    meta_optimizer_iterations: ONE
  }
}
, '_keep_checkpoint_max': 5, '_keep_checkpoint_every_n_hours': 10000, '_log_step_count_steps': 100, '_train_distribute': None, '_device_fn': None, '_protocol': None, '_eval_distribute': None, '_experimental_distribute': None, '_experimental_max_worker_delay_secs': None, '_session_creation_timeout_secs': 7200, '_checkpoint_save_graph_def': True, '_service': None, '_cluster_spec': ClusterSpec({}), '_task_type': 'worker', '_task_id': 0, '_global_id_in_cluster': 0, '_master': '', '_evaluation_master': '', '_is_chief': True, '_num_ps_replicas': 0, '_num_worker_replicas': 1}
dnn_estimator.train(input_fn=_input_fn, steps=100)
dnn_estimator.evaluate(input_fn=_eval_input_fn, steps=10)
INFO:tensorflow:Calling model_fn.
INFO:tensorflow:Done calling model_fn.
INFO:tensorflow:Create CheckpointSaverHook.
INFO:tensorflow:Graph was finalized.
INFO:tensorflow:Running local_init_op.
INFO:tensorflow:Done running local_init_op.
2022-12-14 03:47:19.999250: W tensorflow/core/common_runtime/type_inference.cc:339] Type inference failed. This indicates an invalid graph that escaped type checking. Error message: INVALID_ARGUMENT: expected compatible input types, but input 1:
type_id: TFT_OPTIONAL
args {
  type_id: TFT_PRODUCT
  args {
    type_id: TFT_TENSOR
    args {
      type_id: TFT_INT64
    }
  }
}
 is neither a subtype nor a supertype of the combined inputs preceding it:
type_id: TFT_OPTIONAL
args {
  type_id: TFT_PRODUCT
  args {
    type_id: TFT_TENSOR
    args {
      type_id: TFT_INT32
    }
  }
}

    while inferring type of node 'dnn/zero_fraction/cond/output/_18'
INFO:tensorflow:Calling checkpoint listeners before saving checkpoint 0...
INFO:tensorflow:Saving checkpoints for 0 into /tmpfs/tmp/tmpmdq9lsrm/model.ckpt.
INFO:tensorflow:Calling checkpoint listeners after saving checkpoint 0...
INFO:tensorflow:loss = 3.088563, step = 0
INFO:tensorflow:Calling checkpoint listeners before saving checkpoint 20...
INFO:tensorflow:Saving checkpoints for 20 into /tmpfs/tmp/tmpmdq9lsrm/model.ckpt.
INFO:tensorflow:Calling checkpoint listeners after saving checkpoint 20...
INFO:tensorflow:Loss for final step: 0.58986354.
INFO:tensorflow:Calling model_fn.
INFO:tensorflow:Done calling model_fn.
INFO:tensorflow:Starting evaluation at 2022-12-14T03:47:21
INFO:tensorflow:Graph was finalized.
INFO:tensorflow:Restoring parameters from /tmpfs/tmp/tmpmdq9lsrm/model.ckpt-20
INFO:tensorflow:Running local_init_op.
INFO:tensorflow:Done running local_init_op.
INFO:tensorflow:Evaluation [1/10]
INFO:tensorflow:Evaluation [2/10]
INFO:tensorflow:Evaluation [3/10]
INFO:tensorflow:Evaluation [4/10]
INFO:tensorflow:Evaluation [5/10]
INFO:tensorflow:Evaluation [6/10]
INFO:tensorflow:Evaluation [7/10]
INFO:tensorflow:Evaluation [8/10]
INFO:tensorflow:Evaluation [9/10]
INFO:tensorflow:Inference Time : 0.46118s
INFO:tensorflow:Finished evaluation at 2022-12-14-03:47:21
INFO:tensorflow:Saving dict for global step 20: accuracy = 0.70075756, accuracy_baseline = 0.625, auc = 0.7111723, auc_precision_recall = 0.6200676, average_loss = 0.6077014, global_step = 20, label/mean = 0.375, loss = 0.6034344, precision = 0.6388889, prediction/mean = 0.4131366, recall = 0.46464646
INFO:tensorflow:Saving 'checkpoint_path' summary for global step 20: /tmpfs/tmp/tmpmdq9lsrm/model.ckpt-20
{'accuracy': 0.70075756,
 'accuracy_baseline': 0.625,
 'auc': 0.7111723,
 'auc_precision_recall': 0.6200676,
 'average_loss': 0.6077014,
 'label/mean': 0.375,
 'loss': 0.6034344,
 'precision': 0.6388889,
 'prediction/mean': 0.4131366,
 'recall': 0.46464646,
 'global_step': 20}

TensorFlow 2: Using Keras to create a custom DNN model

In TensorFlow 2, you can create a custom DNN model to substitute for one generated by tf.estimator.DNNEstimator, with similar levels of user-specified customization (for instance, as in the previous example, the ability to customize a chosen model optimizer).

A similar workflow can be used to replace tf.estimator.experimental.RNNEstimator with a Keras recurrent neural network (RNN) model. Keras provides a number of built-in, customizable choices by way of tf.keras.layers.RNN, tf.keras.layers.LSTM, and tf.keras.layers.GRU. To learn more, check out the Built-in RNN layers: a simple example section of RNN with Keras guide.

dnn_model = tf.keras.models.Sequential(
    [tf.keras.layers.Dense(128, activation='relu'),
     tf.keras.layers.Dense(1)])

dnn_model.compile(loss='mse', optimizer=create_sample_optimizer('tf2'), metrics=['accuracy'])
dnn_model.fit(x_train, y_train, epochs=10)
dnn_model.evaluate(x_eval, y_eval, return_dict=True)
Epoch 1/10
20/20 [==============================] - 0s 2ms/step - loss: 86.4269 - accuracy: 0.5805
Epoch 2/10
20/20 [==============================] - 0s 2ms/step - loss: 0.3415 - accuracy: 0.6746
Epoch 3/10
20/20 [==============================] - 0s 2ms/step - loss: 0.2722 - accuracy: 0.6715
Epoch 4/10
20/20 [==============================] - 0s 2ms/step - loss: 0.2191 - accuracy: 0.6826
Epoch 5/10
20/20 [==============================] - 0s 2ms/step - loss: 0.2050 - accuracy: 0.6906
Epoch 6/10
20/20 [==============================] - 0s 2ms/step - loss: 0.2013 - accuracy: 0.6922
Epoch 7/10
20/20 [==============================] - 0s 2ms/step - loss: 0.1900 - accuracy: 0.6906
Epoch 8/10
20/20 [==============================] - 0s 2ms/step - loss: 0.1857 - accuracy: 0.7193
Epoch 9/10
20/20 [==============================] - 0s 2ms/step - loss: 0.1781 - accuracy: 0.7193
Epoch 10/10
20/20 [==============================] - 0s 2ms/step - loss: 0.1815 - accuracy: 0.7209
9/9 [==============================] - 0s 2ms/step - loss: 0.1917 - accuracy: 0.7424
{'loss': 0.19173277914524078, 'accuracy': 0.7424242496490479}

Example 3: Migrating from DNNLinearCombinedEstimator

TensorFlow 1: Using DNNLinearCombinedEstimator

In TensorFlow 1, you can use tf.estimator.DNNLinearCombinedEstimator to create a baseline combined model for regression and classification problems with customization capacity for both its linear and DNN components.

optimizer = create_sample_optimizer('tf1')

combined_estimator = tf.estimator.DNNLinearCombinedEstimator(
    head=tf.estimator.BinaryClassHead(),
    # Wide settings
    linear_feature_columns=feature_columns,
    linear_optimizer=optimizer,
    # Deep settings
    dnn_feature_columns=feature_columns,
    dnn_hidden_units=[128],
    dnn_optimizer=optimizer)
INFO:tensorflow:Using default config.
WARNING:tensorflow:Using temporary folder as model directory: /tmpfs/tmp/tmppyljhe1d
INFO:tensorflow:Using config: {'_model_dir': '/tmpfs/tmp/tmppyljhe1d', '_tf_random_seed': None, '_save_summary_steps': 100, '_save_checkpoints_steps': None, '_save_checkpoints_secs': 600, '_session_config': allow_soft_placement: true
graph_options {
  rewrite_options {
    meta_optimizer_iterations: ONE
  }
}
, '_keep_checkpoint_max': 5, '_keep_checkpoint_every_n_hours': 10000, '_log_step_count_steps': 100, '_train_distribute': None, '_device_fn': None, '_protocol': None, '_eval_distribute': None, '_experimental_distribute': None, '_experimental_max_worker_delay_secs': None, '_session_creation_timeout_secs': 7200, '_checkpoint_save_graph_def': True, '_service': None, '_cluster_spec': ClusterSpec({}), '_task_type': 'worker', '_task_id': 0, '_global_id_in_cluster': 0, '_master': '', '_evaluation_master': '', '_is_chief': True, '_num_ps_replicas': 0, '_num_worker_replicas': 1}
combined_estimator.train(input_fn=_input_fn, steps=100)
combined_estimator.evaluate(input_fn=_eval_input_fn, steps=10)
INFO:tensorflow:Calling model_fn.
INFO:tensorflow:Done calling model_fn.
INFO:tensorflow:Create CheckpointSaverHook.
INFO:tensorflow:Graph was finalized.
INFO:tensorflow:Running local_init_op.
INFO:tensorflow:Done running local_init_op.
INFO:tensorflow:Calling checkpoint listeners before saving checkpoint 0...
INFO:tensorflow:Saving checkpoints for 0 into /tmpfs/tmp/tmppyljhe1d/model.ckpt.
INFO:tensorflow:Calling checkpoint listeners after saving checkpoint 0...
INFO:tensorflow:loss = 4.2737427, step = 0
INFO:tensorflow:Calling checkpoint listeners before saving checkpoint 20...
INFO:tensorflow:Saving checkpoints for 20 into /tmpfs/tmp/tmppyljhe1d/model.ckpt.
INFO:tensorflow:Calling checkpoint listeners after saving checkpoint 20...
INFO:tensorflow:Loss for final step: 0.54818135.
INFO:tensorflow:Calling model_fn.
INFO:tensorflow:Done calling model_fn.
INFO:tensorflow:Starting evaluation at 2022-12-14T03:47:26
INFO:tensorflow:Graph was finalized.
INFO:tensorflow:Restoring parameters from /tmpfs/tmp/tmppyljhe1d/model.ckpt-20
INFO:tensorflow:Running local_init_op.
INFO:tensorflow:Done running local_init_op.
INFO:tensorflow:Evaluation [1/10]
INFO:tensorflow:Evaluation [2/10]
INFO:tensorflow:Evaluation [3/10]
INFO:tensorflow:Evaluation [4/10]
INFO:tensorflow:Evaluation [5/10]
INFO:tensorflow:Evaluation [6/10]
INFO:tensorflow:Evaluation [7/10]
INFO:tensorflow:Evaluation [8/10]
INFO:tensorflow:Evaluation [9/10]
INFO:tensorflow:Inference Time : 0.53725s
INFO:tensorflow:Finished evaluation at 2022-12-14-03:47:26
INFO:tensorflow:Saving dict for global step 20: accuracy = 0.7121212, accuracy_baseline = 0.625, auc = 0.7566269, auc_precision_recall = 0.65287393, average_loss = 0.5770411, global_step = 20, label/mean = 0.375, loss = 0.5672319, precision = 0.6619718, prediction/mean = 0.4006542, recall = 0.47474748
INFO:tensorflow:Saving 'checkpoint_path' summary for global step 20: /tmpfs/tmp/tmppyljhe1d/model.ckpt-20
{'accuracy': 0.7121212,
 'accuracy_baseline': 0.625,
 'auc': 0.7566269,
 'auc_precision_recall': 0.65287393,
 'average_loss': 0.5770411,
 'label/mean': 0.375,
 'loss': 0.5672319,
 'precision': 0.6619718,
 'prediction/mean': 0.4006542,
 'recall': 0.47474748,
 'global_step': 20}

TensorFlow 2: Using Keras WideDeepModel

In TensorFlow 2, you can create an instance of the Keras tf.compat.v1.keras.models.WideDeepModel to substitute for one generated by tf.estimator.DNNLinearCombinedEstimator, with similar levels of user-specified customization (for instance, as in the previous example, the ability to customize a chosen model optimizer).

This WideDeepModel is constructed on the basis of a constituent LinearModel and a custom DNN Model, both of which are discussed in the preceding two examples. A custom linear model can also be used in place of the built-in Keras LinearModel if desired.

If you would like to build your own model instead of using a canned estimator, check out the Keras Sequential model guide. For more information on custom training and optimizers, check out the Custom training: walkthrough guide.

# Create LinearModel and DNN Model as in Examples 1 and 2
optimizer = create_sample_optimizer('tf2')

linear_model = tf.compat.v1.keras.experimental.LinearModel()
linear_model.compile(loss='mse', optimizer=optimizer, metrics=['accuracy'])
linear_model.fit(x_train, y_train, epochs=10, verbose=0)

dnn_model = tf.keras.models.Sequential(
    [tf.keras.layers.Dense(128, activation='relu'),
     tf.keras.layers.Dense(1)])
dnn_model.compile(loss='mse', optimizer=optimizer, metrics=['accuracy'])
combined_model = tf.compat.v1.keras.experimental.WideDeepModel(linear_model,
                                                               dnn_model)
combined_model.compile(
    optimizer=[optimizer, optimizer], loss='mse', metrics=['accuracy'])
combined_model.fit([x_train, x_train], y_train, epochs=10)
combined_model.evaluate(x_eval, y_eval, return_dict=True)
Epoch 1/10
20/20 [==============================] - 0s 3ms/step - loss: 597.0383 - accuracy: 0.4992
Epoch 2/10
20/20 [==============================] - 0s 3ms/step - loss: 0.2878 - accuracy: 0.6874
Epoch 3/10
20/20 [==============================] - 0s 2ms/step - loss: 0.1988 - accuracy: 0.7560
Epoch 4/10
20/20 [==============================] - 0s 3ms/step - loss: 0.1739 - accuracy: 0.7831
Epoch 5/10
20/20 [==============================] - 0s 3ms/step - loss: 0.1632 - accuracy: 0.7927
Epoch 6/10
20/20 [==============================] - 0s 3ms/step - loss: 0.1603 - accuracy: 0.7943
Epoch 7/10
20/20 [==============================] - 0s 3ms/step - loss: 0.1569 - accuracy: 0.8054
Epoch 8/10
20/20 [==============================] - 0s 2ms/step - loss: 0.1525 - accuracy: 0.8022
Epoch 9/10
20/20 [==============================] - 0s 2ms/step - loss: 0.1659 - accuracy: 0.7974
Epoch 10/10
20/20 [==============================] - 0s 2ms/step - loss: 0.1547 - accuracy: 0.7959
9/9 [==============================] - 0s 2ms/step - loss: 0.1731 - accuracy: 0.7500
{'loss': 0.1730780303478241, 'accuracy': 0.75}

Example 4: Migrating from BoostedTreesEstimator

TensorFlow 1: Using BoostedTreesEstimator

In TensorFlow 1, you could use tf.estimator.BoostedTreesEstimator to create a baseline to create a baseline Gradient Boosting model using an ensemble of decision trees for regression and classification problems. This functionality is no longer included in TensorFlow 2.

bt_estimator = tf1.estimator.BoostedTreesEstimator(
    head=tf.estimator.BinaryClassHead(),
    n_batches_per_layer=1,
    max_depth=10,
    n_trees=1000,
    feature_columns=feature_columns)
bt_estimator.train(input_fn=_input_fn, steps=1000)
bt_estimator.evaluate(input_fn=_eval_input_fn, steps=100)

TensorFlow 2: Using TensorFlow Decision Forests

In TensorFlow 2, tf.estimator.BoostedTreesEstimator is replaced by tfdf.keras.GradientBoostedTreesModel from the TensorFlow Decision Forests package.

TensorFlow Decision Forests provides various advantages over the tf.estimator.BoostedTreesEstimator, notably regarding quality, speed, ease of use and flexibility. To learn about TensorFlow Decision Forests, start with the beginner colab.

The following example shows how to train a Gradient Boosted Trees model using TensorFlow 2:

Install TensorFlow Decision Forests.

pip install tensorflow_decision_forests

Create a TensorFlow dataset. Note that Decision Forests natively support many types of features and do not need pre-processing.

train_dataframe = pd.read_csv('https://storage.googleapis.com/tf-datasets/titanic/train.csv')
eval_dataframe = pd.read_csv('https://storage.googleapis.com/tf-datasets/titanic/eval.csv')

# Convert the Pandas Dataframes into TensorFlow datasets.
train_dataset = tfdf.keras.pd_dataframe_to_tf_dataset(train_dataframe, label="survived")
eval_dataset = tfdf.keras.pd_dataframe_to_tf_dataset(eval_dataframe, label="survived")

Train the model on the train_dataset dataset.

# Use the default hyper-parameters of the model.
gbt_model = tfdf.keras.GradientBoostedTreesModel()
gbt_model.fit(train_dataset)
Warning: The `num_threads` constructor argument is not set and the number of CPU is os.cpu_count()=32 > 32. Setting num_threads to 32. Set num_threads manually to use more than 32 cpus.
WARNING:absl:The `num_threads` constructor argument is not set and the number of CPU is os.cpu_count()=32 > 32. Setting num_threads to 32. Set num_threads manually to use more than 32 cpus.
Use /tmpfs/tmp/tmpwmxg0bbf as temporary training directory
Reading training dataset...
WARNING:tensorflow:From /tmpfs/src/tf_docs_env/lib/python3.9/site-packages/tensorflow/python/autograph/pyct/static_analysis/liveness.py:83: Analyzer.lamba_check (from tensorflow.python.autograph.pyct.static_analysis.liveness) is deprecated and will be removed after 2023-09-23.
Instructions for updating:
Lambda fuctions will be no more assumed to be used in the statement where they are used, or at least in the same block. https://github.com/tensorflow/tensorflow/issues/56089
WARNING:tensorflow:From /tmpfs/src/tf_docs_env/lib/python3.9/site-packages/tensorflow/python/autograph/pyct/static_analysis/liveness.py:83: Analyzer.lamba_check (from tensorflow.python.autograph.pyct.static_analysis.liveness) is deprecated and will be removed after 2023-09-23.
Instructions for updating:
Lambda fuctions will be no more assumed to be used in the statement where they are used, or at least in the same block. https://github.com/tensorflow/tensorflow/issues/56089
Training dataset read in 0:00:03.171120. Found 627 examples.
Training model...
2022-12-14 03:47:33.727569: W external/ydf/yggdrasil_decision_forests/learner/gradient_boosted_trees/gradient_boosted_trees.cc:1765] Subsample hyperparameter given but sampling method does not match.
2022-12-14 03:47:33.727613: W external/ydf/yggdrasil_decision_forests/learner/gradient_boosted_trees/gradient_boosted_trees.cc:1778] GOSS alpha hyperparameter given but GOSS is disabled.
2022-12-14 03:47:33.727620: W external/ydf/yggdrasil_decision_forests/learner/gradient_boosted_trees/gradient_boosted_trees.cc:1787] GOSS beta hyperparameter given but GOSS is disabled.
2022-12-14 03:47:33.727633: W external/ydf/yggdrasil_decision_forests/learner/gradient_boosted_trees/gradient_boosted_trees.cc:1799] SelGB ratio hyperparameter given but SelGB is disabled.
Model trained in 0:00:00.224369
Compiling model...
[INFO 2022-12-14T03:47:33.942071121+00:00 kernel.cc:1175] Loading model from path /tmpfs/tmp/tmpwmxg0bbf/model/ with prefix 25097dbecf6546dd
[INFO 2022-12-14T03:47:33.945814229+00:00 abstract_model.cc:1306] Engine "GradientBoostedTreesQuickScorerExtended" built
[INFO 2022-12-14T03:47:33.945845819+00:00 kernel.cc:1021] Use fast generic engine
WARNING:tensorflow:AutoGraph could not transform <function simple_ml_inference_op_with_handle at 0x7f6ac875ad30> and will run it as-is.
Please report this to the TensorFlow team. When filing the bug, set the verbosity to 10 (on Linux, `export AUTOGRAPH_VERBOSITY=10`) and attach the full output.
Cause: could not get source code
To silence this warning, decorate the function with @tf.autograph.experimental.do_not_convert
WARNING:tensorflow:AutoGraph could not transform <function simple_ml_inference_op_with_handle at 0x7f6ac875ad30> and will run it as-is.
Please report this to the TensorFlow team. When filing the bug, set the verbosity to 10 (on Linux, `export AUTOGRAPH_VERBOSITY=10`) and attach the full output.
Cause: could not get source code
To silence this warning, decorate the function with @tf.autograph.experimental.do_not_convert
WARNING: AutoGraph could not transform <function simple_ml_inference_op_with_handle at 0x7f6ac875ad30> and will run it as-is.
Please report this to the TensorFlow team. When filing the bug, set the verbosity to 10 (on Linux, `export AUTOGRAPH_VERBOSITY=10`) and attach the full output.
Cause: could not get source code
To silence this warning, decorate the function with @tf.autograph.experimental.do_not_convert
Model compiled.
<keras.callbacks.History at 0x7f69f44494c0>

Evaluate the quality of the model on the eval_dataset dataset.

gbt_model.compile(metrics=['accuracy'])
gbt_evaluation = gbt_model.evaluate(eval_dataset, return_dict=True)
print(gbt_evaluation)
1/1 [==============================] - 0s 279ms/step - loss: 0.0000e+00 - accuracy: 0.8295
{'loss': 0.0, 'accuracy': 0.8295454382896423}

Gradient Boosted Trees is just one of the many decision forest algorithms available in TensorFlow Decision Forests. For example, Random Forests (available as tfdf.keras.GradientBoostedTreesModel is very resistant to overfitting) while CART (available as tfdf.keras.CartModel) is great for model interpretation.

In the next example, train and plot a Random Forest model.

# Train a Random Forest model
rf_model = tfdf.keras.RandomForestModel()
rf_model.fit(train_dataset)

# Evaluate the Random Forest model
rf_model.compile(metrics=['accuracy'])
rf_evaluation = rf_model.evaluate(eval_dataset, return_dict=True)
print(rf_evaluation)
Warning: The `num_threads` constructor argument is not set and the number of CPU is os.cpu_count()=32 > 32. Setting num_threads to 32. Set num_threads manually to use more than 32 cpus.
WARNING:absl:The `num_threads` constructor argument is not set and the number of CPU is os.cpu_count()=32 > 32. Setting num_threads to 32. Set num_threads manually to use more than 32 cpus.
Use /tmpfs/tmp/tmp5544m0nu as temporary training directory
Reading training dataset...
Training dataset read in 0:00:00.175397. Found 627 examples.
Training model...
Model trained in 0:00:00.191206
Compiling model...
[INFO 2022-12-14T03:47:35.854534193+00:00 kernel.cc:1175] Loading model from path /tmpfs/tmp/tmp5544m0nu/model/ with prefix 838d0a6768af4637
[INFO 2022-12-14T03:47:35.953742261+00:00 kernel.cc:1021] Use fast generic engine
Model compiled.
1/1 [==============================] - 0s 128ms/step - loss: 0.0000e+00 - accuracy: 0.8333
{'loss': 0.0, 'accuracy': 0.8333333134651184}

In the final example, train and evaluate a CART model.

# Train a CART model
cart_model = tfdf.keras.CartModel()
cart_model.fit(train_dataset)

# Plot the CART model
tfdf.model_plotter.plot_model_in_colab(cart_model, max_depth=2)
Warning: The `num_threads` constructor argument is not set and the number of CPU is os.cpu_count()=32 > 32. Setting num_threads to 32. Set num_threads manually to use more than 32 cpus.
WARNING:absl:The `num_threads` constructor argument is not set and the number of CPU is os.cpu_count()=32 > 32. Setting num_threads to 32. Set num_threads manually to use more than 32 cpus.
Use /tmpfs/tmp/tmpjm_0ja_l as temporary training directory
Reading training dataset...
Training dataset read in 0:00:00.174151. Found 627 examples.
Training model...
Model trained in 0:00:00.016527
Compiling model...
2022-12-14 03:47:36.447234: W external/ydf/yggdrasil_decision_forests/model/random_forest/random_forest.cc:607] ValidationEvaluation requires OOB evaluation enabled.Random Forest models should be trained with compute_oob_performances:true. CART models do not support OOB evaluation.
[INFO 2022-12-14T03:47:36.458485956+00:00 kernel.cc:1175] Loading model from path /tmpfs/tmp/tmpjm_0ja_l/model/ with prefix 833cb2cdd91d4dbb
[INFO 2022-12-14T03:47:36.45885296+00:00 kernel.cc:1021] Use fast generic engine
Model compiled.