def build_estimator(model_dir, model_type):
"""Build an estimator using official tf.estimator API.
Args:
model_dir: model save base directory
model_type: one of {`wide_deep`, `deep`, `wide_deep`}
Returns:
model instance of tf.estimator.Estimator class
"""
wide_columns, deep_columns = _build_model_columns()
_build_distribution()
# Create a tf.estimator.RunConfig to ensure the model is run on CPU, which
# trains faster than GPU for this model.
run_config = tf.estimator.RunConfig(**CONF.runconfig).replace(
session_config=tf.ConfigProto(device_count={'GPU': 0}))
if model_type == 'wide_deep':
return tf.estimator.LinearClassifier(
model_dir=model_dir,
feature_columns=wide_columns,
weight_column=weight_column,
optimizer=tf.train.FtrlOptimizer( # can not read from conf
learning_rate=0.1,
l1_regularization_strength=0.5,
l2_regularization_strength=1),
partitioner=None,
config=run_config)
elif model_type == 'deep':
return tf.estimator.DNNClassifier(
model_dir=model_dir,
feature_columns=deep_columns,
hidden_units=CONF.model["dnn_hidden_units"],
optimizer=tf.train.ProximalAdagradOptimizer(
learning_rate=0.1,
l1_regularization_strength=0.1,
l2_regularization_strength=0.1
), # {'Adagrad', 'Adam', 'Ftrl', 'RMSProp', 'SGD'}
activation_fn=activation_fn(CONF.model["dnn_activation_function"]
), # tf.nn.relu vs 'tf.nn.relu'
dropout=CONF.model["dnn_dropout"],
weight_column=weight_column,
input_layer_partitioner=None,
config=run_config)
else:
return tf.estimator.DNNLinearCombinedClassifier(
model_dir=
model_dir, # self._model_dir = model_dir or self._config.model_dir
linear_feature_columns=wide_columns,
linear_optimizer=tf.train.FtrlOptimizer(
learning_rate=0.1,
l1_regularization_strength=0.5,
l2_regularization_strength=1),
dnn_feature_columns=deep_columns,
dnn_optimizer=tf.train.ProximalAdagradOptimizer(
learning_rate=0.1,
l1_regularization_strength=0.1,
l2_regularization_strength=0.1),
dnn_hidden_units=CONF.model["dnn_hidden_units"],
dnn_activation_fn=activation_fn(
CONF.model["dnn_activation_function"]),
dnn_dropout=CONF.model["dnn_dropout"],
n_classes=2,
weight_column=weight_column,
label_vocabulary=None,
input_layer_partitioner=None,
config=run_config)
from __future__ import print_function
import six
import tensorflow as tf
from tensorflow.python.estimator.canned import head as head_lib
import os
import sys
PACKAGE_DIR = os.path.dirname(os.path.dirname(os.path.abspath(__file__)))
sys.path.insert(0, PACKAGE_DIR)
from lib.read_conf import Config
from lib.utils.model_util import add_layer_summary, get_optimizer_instance, activation_fn
CONF = Config().model
ACTIVATION_FN = activation_fn(CONF['dnn_activation_function'])
DROPOUT = CONF['dnn_dropout']
BATCH_NORM = CONF['dnn_batch_normalization']
DNN_L1 = CONF['dnn_l1']
DNN_L2 = CONF['dnn_l2']
regularizer_list = []
# WENQI
if DNN_L1 and DNN_L2:
REG = tf.keras.regularizers.L1L2(l1=DNN_L1, l2=DNN_L2)
elif DNN_L1:
REG = tf.keras.regularizers.l1(DNN_L1)
elif DNN_L2:
REG = tf.keras.regularizers.l2(0.5 * (DNN_L2))
else:
REG = None