def build_model(opts, as_components=False):
"""Builds a Keras model for Criteo data."""
layers_tup = uq_utils.get_layer_builders(opts.method, opts.dropout_rate,
data_lib.NUM_TRAIN_EXAMPLES)
_, dense_layer, dense_last, dropout_fn, dropout_fn_last = layers_tup
fcs_int, fcs_cat = make_feature_columns(opts)
input_layer = make_input_layers()
features = input_layer
dense_int = keras.layers.DenseFeatures(fcs_int)(features)
dense_cat = keras.layers.DenseFeatures(fcs_cat)(features)
net = tf.concat([dense_int, dense_cat], axis=-1)
logging.info('Dense layer shape: %s', net.shape)
# TODO(yovadia): Consider explicit normalization according to data stats.
net = keras.layers.BatchNormalization()(net)
for size in opts.layer_sizes:
net = dropout_fn(net)
net = dense_layer(size, activation='relu')(net)
prelogits = dropout_fn_last(net)
# Sigmoid output necessary to get useful AUC metric outputs.
lastlayer = dense_last(1, activation='sigmoid')
probs = lastlayer(prelogits)
if as_components:
trunc = keras.Model(inputs=input_layer, outputs=prelogits)
embedings_in = keras.layers.Input(shape=prelogits.shape[1:])
head = keras.Model(inputs=embedings_in,
outputs=lastlayer(embedings_in))
return trunc, head
return keras.Model(inputs=input_layer, outputs=probs)
def _build_mlp(opts):
"""Builds a multi-layer perceptron Keras model."""
layer_builders = uq_utils.get_layer_builders(opts.method, opts.dropout_rate,
opts.num_train_examples)
_, dense_layer, dense_last, dropout_fn, dropout_fn_last = layer_builders
inputs = keras.layers.Input(_MNIST_SHAPE)
net = keras.layers.Flatten(input_shape=_MNIST_SHAPE)(inputs)
for size in opts.mlp_layer_sizes:
net = dropout_fn(net)
net = dense_layer(size, activation='relu')(net)
net = dropout_fn_last(net)
logits = dense_last(_NUM_CLASSES)(net)
return keras.Model(inputs=inputs, outputs=logits)
def _build_lenet(opts):
"""Builds a LeNet Keras model."""
layer_builders = uq_utils.get_layer_builders(opts.method, opts.dropout_rate,
opts.num_train_examples)
conv2d, dense_layer, dense_last, dropout_fn, dropout_fn_last = layer_builders
inputs = keras.layers.Input(_MNIST_SHAPE)
net = inputs
net = conv2d(32, kernel_size=(3, 3),
activation='relu',
input_shape=_MNIST_SHAPE)(net)
net = conv2d(64, (3, 3), activation='relu')(net)
net = keras.layers.MaxPooling2D(pool_size=(2, 2))(net)
net = dropout_fn(net)
net = keras.layers.Flatten()(net)
net = dense_layer(128, activation='relu')(net)
net = dropout_fn_last(net)
logits = dense_last(_NUM_CLASSES, activation='relu')(net)
return keras.Model(inputs=inputs, outputs=logits)
