def test_train_sequential_with_distribution_strategy(
self, distribution, cloning):
keras_model = simple_sequential_model()
keras_model.compile(
loss='categorical_crossentropy',
metrics=[tf.keras.metrics.CategoricalAccuracy()],
optimizer=tf.keras.optimizers.RMSprop(learning_rate=0.01),
cloning=cloning)
config = run_config_lib.RunConfig(tf_random_seed=_RANDOM_SEED,
model_dir=self._base_dir,
train_distribute=distribution)
with self.cached_session():
est_keras = keras_lib.model_to_estimator(keras_model=keras_model,
config=config)
before_eval_results = est_keras.evaluate(
input_fn=get_ds_test_input_fn, steps=1)
est_keras.train(input_fn=get_ds_train_input_fn,
steps=_TRAIN_SIZE / 16)
after_eval_results = est_keras.evaluate(
input_fn=get_ds_test_input_fn, steps=1)
self.assertLess(after_eval_results['loss'],
before_eval_results['loss'])
tf.compat.v1.summary.FileWriterCache.clear()
tf.compat.v1.gfile.DeleteRecursively(self._config.model_dir)
def do_test_multi_inputs_multi_outputs_with_input_fn(
self, distribution, train_input_fn, eval_input_fn):
config = run_config_lib.RunConfig(tf_random_seed=_RANDOM_SEED,
model_dir=self._base_dir,
train_distribute=distribution)
with self.cached_session():
model = multi_inputs_multi_outputs_model()
est_keras = keras_lib.model_to_estimator(keras_model=model,
config=config)
baseline_eval_results = est_keras.evaluate(input_fn=eval_input_fn,
steps=1)
est_keras.train(input_fn=train_input_fn, steps=_TRAIN_SIZE / 16)
eval_results = est_keras.evaluate(input_fn=eval_input_fn, steps=1)
self.assertLess(eval_results['loss'],
baseline_eval_results['loss'])
def test_train_premade_linear_model(self):
(x_train, y_train
), _, train_inp_fn, eval_inp_fn = get_resource_for_simple_model()
linear_model = tf.keras.experimental.LinearModel(units=1)
opt = tf.keras.optimizers.SGD(0.1)
linear_model.compile(opt, 'mse', ['mse'])
linear_model.fit(x_train, y_train, epochs=10)
est = keras_lib.model_to_estimator(keras_model=linear_model,
config=self._config,
checkpoint_format='saver')
before_eval_results = est.evaluate(input_fn=eval_inp_fn, steps=1)
est.train(input_fn=train_inp_fn, steps=500)
after_eval_results = est.evaluate(input_fn=eval_inp_fn, steps=1)
self.assertLess(after_eval_results['loss'],
before_eval_results['loss'])
self.assertLess(after_eval_results['loss'], 0.1)
def test_train_premade_linear_model_with_dense_features(self):
vocab_list = ['alpha', 'beta', 'gamma']
vocab_val = [0.4, 0.6, 0.9]
data = np.random.choice(vocab_list, size=256)
y = np.zeros_like(data, dtype=np.float32)
for vocab, val in zip(vocab_list, vocab_val):
indices = np.where(data == vocab)
y[indices] = val + np.random.uniform(
low=-0.01, high=0.01, size=indices[0].shape)
cat_column = tf.feature_column.categorical_column_with_vocabulary_list(
key='symbol', vocabulary_list=vocab_list)
ind_column = tf.feature_column.indicator_column(cat_column)
keras_input = tf.keras.layers.Input(name='symbol',
shape=3,
dtype=tf.dtypes.string)
feature_layer = tf.compat.v1.keras.layers.DenseFeatures([ind_column])
h = feature_layer({'symbol': keras_input})
linear_model = tf.keras.experimental.LinearModel(units=1)
h = linear_model(h)
model = tf.keras.models.Model(inputs=keras_input, outputs=h)
opt = tf.keras.optimizers.SGD(0.1)
model.compile(opt, 'mse', ['mse'])
train_input_fn = numpy_io.numpy_input_fn(x={'symbol': data},
y=y,
num_epochs=20,
shuffle=False)
eval_input_fn = numpy_io.numpy_input_fn(x={'symbol': data},
y=y,
num_epochs=20,
shuffle=False)
est = keras_lib.model_to_estimator(keras_model=model,
config=self._config,
checkpoint_format='saver')
before_eval_results = est.evaluate(input_fn=eval_input_fn, steps=1)
est.train(input_fn=train_input_fn, steps=30)
after_eval_results = est.evaluate(input_fn=eval_input_fn, steps=1)
self.assertLess(after_eval_results['loss'],
before_eval_results['loss'])
self.assertLess(after_eval_results['loss'], 0.05)
def model_to_estimator(keras_model=None,
keras_model_path=None,
custom_objects=None,
model_dir=None,
config=None,
checkpoint_format='saver',
metric_names_map=None,
export_outputs=None):
"""Constructs an `Estimator` instance from given keras model.
If you use infrastructure or other tooling that relies on Estimators, you can
still build a Keras model and use model_to_estimator to convert the Keras
model to an Estimator for use with downstream systems.
For usage example, please see:
[Creating estimators from Keras Models](
https://www.tensorflow.org/guide/estimators#creating_estimators_from_keras_models).
Sample Weights:
Estimators returned by `model_to_estimator` are configured so that they can
handle sample weights (similar to `keras_model.fit(x, y, sample_weights)`).
To pass sample weights when training or evaluating the Estimator, the first
item returned by the input function should be a dictionary with keys
`features` and `sample_weights`. Example below:
```python
keras_model = tf.keras.Model(...)
keras_model.compile(...)
estimator = tf.keras.estimator.model_to_estimator(keras_model)
def input_fn():
return dataset_ops.Dataset.from_tensors(
({'features': features, 'sample_weights': sample_weights},
targets))
estimator.train(input_fn, steps=1)
```
Example with customized export signature:
```python
inputs = {'a': tf.keras.Input(..., name='a'),
'b': tf.keras.Input(..., name='b')}
outputs = {'c': tf.keras.layers.Dense(..., name='c')(inputs['a']),
'd': tf.keras.layers.Dense(..., name='d')(inputs['b'])}
keras_model = tf.keras.Model(inputs, outputs)
keras_model.compile(...)
export_outputs = {'c': tf.estimator.export.RegressionOutput,
'd': tf.estimator.export.ClassificationOutput}
estimator = tf.keras.estimator.model_to_estimator(
keras_model, export_outputs=export_outputs)
def input_fn():
return dataset_ops.Dataset.from_tensors(
({'features': features, 'sample_weights': sample_weights},
targets))
estimator.train(input_fn, steps=1)
```
Args:
keras_model: A compiled Keras model object. This argument is mutually
exclusive with `keras_model_path`. Estimator's `model_fn` uses the
structure of the model to clone the model. Defaults to `None`.
keras_model_path: Path to a compiled Keras model saved on disk, in HDF5
format, which can be generated with the `save()` method of a Keras model.
This argument is mutually exclusive with `keras_model`.
Defaults to `None`.
custom_objects: Dictionary for cloning customized objects. This is
used with classes that is not part of this pip package. For example, if
user maintains a `relu6` class that inherits from `tf.keras.layers.Layer`,
then pass `custom_objects={'relu6': relu6}`. Defaults to `None`.
model_dir: Directory to save `Estimator` model parameters, graph, summary
files for TensorBoard, etc. If unset a directory will be created with
`tempfile.mkdtemp`
config: `RunConfig` to config `Estimator`. Allows setting up things in
`model_fn` based on configuration such as `num_ps_replicas`, or
`model_dir`. Defaults to `None`. If both `config.model_dir` and the
`model_dir` argument (above) are specified the `model_dir` **argument**
takes precedence.
checkpoint_format: Sets the format of the checkpoint saved by the estimator
when training. May be `saver` or `checkpoint`, depending on whether to
save checkpoints from `tf.train.Saver` or `tf.train.Checkpoint`. This
argument currently defaults to `saver`. When 2.0 is released, the default
will be `checkpoint`. Estimators use name-based `tf.train.Saver`
checkpoints, while Keras models use object-based checkpoints from
`tf.train.Checkpoint`. Currently, saving object-based checkpoints from
`model_to_estimator` is only supported by Functional and Sequential
models. Defaults to 'saver'.
metric_names_map: Optional dictionary mapping Keras model output metric
names to custom names. This can be used to override the default Keras
model output metrics names in a multi IO model use case and provide custom
names for the `eval_metric_ops` in Estimator.
The Keras model metric names can be obtained using `model.metrics_names`
excluding any loss metrics such as total loss and output losses.
For example, if your Keras model has two outputs `out_1` and `out_2`,
with `mse` loss and `acc` metric, then `model.metrics_names` will be
`['loss', 'out_1_loss', 'out_2_loss', 'out_1_acc', 'out_2_acc']`.
The model metric names excluding the loss metrics will be
`['out_1_acc', 'out_2_acc']`.
export_outputs: Optional dictionary. This can be used to override the
default Keras model output exports in a multi IO model use case and
provide custom names for the `export_outputs` in
`tf.estimator.EstimatorSpec`. Default is None, which is equivalent to
{'serving_default': `tf.estimator.export.PredictOutput`}. If not None,
the keys must match the keys of `model.output_names`.
A dict `{name: output}` where:
* name: An arbitrary name for this output.
* output: an `ExportOutput` class such as `ClassificationOutput`,
`RegressionOutput`, or `PredictOutput`. Single-headed models only need
to specify one entry in this dictionary. Multi-headed models should
specify one entry for each head, one of which must be named using
`tf.saved_model.signature_constants.DEFAULT_SERVING_SIGNATURE_DEF_KEY`
If no entry is provided, a default `PredictOutput` mapping to
`predictions` will be created.
Returns:
An Estimator from given keras model.
Raises:
ValueError: If neither keras_model nor keras_model_path was given.
ValueError: If both keras_model and keras_model_path was given.
ValueError: If the keras_model_path is a GCS URI.
ValueError: If keras_model has not been compiled.
ValueError: If an invalid checkpoint_format was given.
"""
try:
from tensorflow_estimator.python.estimator import keras_lib # pylint: disable=g-import-not-at-top
except ImportError:
raise NotImplementedError(
'tf.keras.estimator.model_to_estimator function not available in your '
'installation.')
_model_to_estimator_usage_gauge.get_cell('v1').set(True)
return keras_lib.model_to_estimator( # pylint:disable=unexpected-keyword-arg
keras_model=keras_model,
keras_model_path=keras_model_path,
custom_objects=custom_objects,
model_dir=model_dir,
config=config,
checkpoint_format=checkpoint_format,
use_v2_estimator=False,
metric_names_map=metric_names_map,
export_outputs=export_outputs)
def test_train_premade_widedeep_model_with_feature_layers(self):
vocab_list = ['alpha', 'beta', 'gamma']
vocab_val = [0.4, 0.6, 0.9]
data = np.random.choice(vocab_list, size=256)
y = np.zeros_like(data, dtype=np.float32)
for vocab, val in zip(vocab_list, vocab_val):
indices = np.where(data == vocab)
y[indices] = val + np.random.uniform(
low=-0.01, high=0.01, size=indices[0].shape)
cat_column = tf.feature_column.categorical_column_with_vocabulary_list(
key='symbol', vocabulary_list=vocab_list)
ind_column = tf.feature_column.indicator_column(cat_column)
# TODO(tanzheny): use emb column for dense part once b/139667019 is fixed.
# emb_column = feature_column.embedding_column(cat_column, dimension=5)
keras_input = tf.keras.layers.Input(name='symbol',
shape=3,
dtype=tf.dtypes.string)
# build linear part with feature layer.
linear_feature_layer = tf.compat.v1.keras.layers.DenseFeatures(
[ind_column])
linear_model = tf.keras.experimental.LinearModel(
units=1, name='Linear', kernel_initializer='zeros')
combined_linear = tf.keras.models.Sequential(
[linear_feature_layer, linear_model])
# build dnn part with feature layer.
dnn_feature_layer = tf.compat.v1.keras.layers.DenseFeatures(
[ind_column])
dense_layer = tf.keras.layers.Dense(units=1,
name='DNNDense',
kernel_initializer='zeros')
combined_dnn = tf.keras.models.Sequential(
[dnn_feature_layer, dense_layer])
# build and compile wide deep.
wide_deep_model = tf.keras.experimental.WideDeepModel(
combined_linear, combined_dnn)
wide_deep_model._set_inputs({'symbol': keras_input})
sgd_opt = tf.keras.optimizers.SGD(0.1)
adam_opt = tf.keras.optimizers.Adam(0.1)
wide_deep_model.compile([sgd_opt, adam_opt], 'mse', ['mse'])
# build estimator.
train_input_fn = numpy_io.numpy_input_fn(x={'symbol': data},
y=y,
num_epochs=20,
shuffle=False)
eval_input_fn = numpy_io.numpy_input_fn(x={'symbol': data},
y=y,
num_epochs=20,
shuffle=False)
est = keras_lib.model_to_estimator(keras_model=wide_deep_model,
config=self._config,
checkpoint_format='saver')
before_eval_results = est.evaluate(input_fn=eval_input_fn, steps=1)
est.train(input_fn=train_input_fn, steps=20)
after_eval_results = est.evaluate(input_fn=eval_input_fn, steps=1)
self.assertLess(after_eval_results['loss'],
before_eval_results['loss'])
self.assertLess(after_eval_results['loss'], 0.1)
