def create_exp_decaying_learning_rate(initial_learning_rate = 0.0001,
decay_steps = 10000,
decay_rate = 0.9,
staircase = True):
"""Create a learning rate that decays exponentially with global_steps.
Args:
initial_learning_rate: A scalar float32 or float64 Tensor or a Python
number. The initial learning rate.
decay_steps: A scalar int32 or int64 Tensor or a Python number. Must be
positive. See the decay computation in `tf.exponential_decay`.
decay_rate: A scalar float32 or float64 Tensor or a Python number. The decay
rate.
staircase: Boolean. If True decay the learning rate at discrete intervals.
Returns:
learning_rate: Scaler tf.Tensor with the learning rate depending on the
globat_step.
"""
learning_rate = tf.exponential_decay(
learning_rate=initial_learning_rate,
global_step=tf.get_or_create_global_step(),
decay_steps=decay_steps,
decay_rate=decay_rate,
staircase=staircase)
return learning_rate
def exponential_decay(initial_value = 0.0001,
decay_steps = 10000,
decay_rate = 0.9,
staircase = True):
"""Create a value that decays exponentially with global_step.
Args:
initial_value: A scalar float32 or float64 Tensor or a Python
number. The initial value returned for global_step == 0.
decay_steps: A scalar int32 or int64 Tensor or a Python number. Must be
positive. See the decay computation in `tf.exponential_decay`.
decay_rate: A scalar float32 or float64 Tensor or a Python number. The decay
rate.
staircase: Boolean. If True, decay the value at discrete intervals.
Returns:
value: Scalar tf.Tensor with the value decaying based on the globat_step.
"""
global_step = tf.train.get_or_create_global_step()
value = tf.exponential_decay(
learning_rate=initial_value,
global_step=global_step,
decay_steps=decay_steps,
decay_rate=decay_rate,
staircase=staircase)
return value
def __init__(self, learning_rate, no_inputs):
# create feature columns
feature_cols = []
for i in range(no_inputs):
feature_cols.append(
tf.feature_column.categorical_column_with_vocabulary_list(
str(i + 1), vocabulary_list=['S', 'O', 'E']))
hidden_layer_units = [27, 27, 27]
# instantiate estimator
self.estimator = tf.estimator.DNNRegressor(
feature_columns=feature_cols,
model_dir='train',
hidden_units=hidden_layer_units,
optimizer=lambda: tf.AdamOptimizer(
learning_rate=tf.exponential_decay(learning_rate=0.1,
global_step=tf.
get_global_step(),
decay_steps=10000,
decay_rate=0.96)))
estimator = DNNRegressor(
feature_columns=[categorical_feature_a_emb, categorical_feature_b_emb],
hidden_units=[1024, 512, 256],
optimizer=tf.train.ProximalAdagradOptimizer(
learning_rate=0.1,
l1_regularization_strength=0.001
))
# Or estimator using an optimizer with a learning rate decay.
estimator = DNNRegressor(
feature_columns=[categorical_feature_a_emb, categorical_feature_b_emb],
hidden_units=[1024, 512, 256],
optimizer=lambda: tf.AdamOptimizer(
learning_rate=tf.exponential_decay(
learning_rate=0.1,
global_step=tf.get_global_step(),
decay_steps=10000,
decay_rate=0.96))
# Or estimator with warm-starting from a previous checkpoint.
estimator = DNNRegressor(
feature_columns=[categorical_feature_a_emb, categorical_feature_b_emb],
hidden_units=[1024, 512, 256],
warm_start_from="/path/to/checkpoint/dir")
# Input builders
def input_fn_train: # returns x, y
pass
estimator.train(input_fn=input_fn_train, steps=100)
def input_fn_eval: # returns x, y
learning_rate=tf.compat.v1.train.exponential_decay(
learning_rate=0.1,
global_step=tf.compat.v1.train.get_global_step(),
decay_steps=10000,
decay_rate=0.96)),
# warm-start settings
model_dir="/Users/songfeng/workspace/github/tensorflowDemo/model/widedeep2",
# warm_start_from="/Users/songfeng/workspace/github/tensorflowDemo/model/widedeep2"
)
# To apply L1 and L2 regularization, you can set dnn_optimizer to:
# To apply learning rate decay, you can set dnn_optimizer to a callable:
lambda: tf.AdamOptimizer(learning_rate=tf.exponential_decay(learning_rate=0.1,
global_step=tf.
get_global_step(),
decay_steps=10000,
decay_rate=0.96))
def input_fn_train():
return train_ds
def input_fn_eval():
return train_ds.take(1)
def input_fn_predict():
return train_ds.take(1)
