def build_network(self, inputs, phase_train=True, nclass=1001):
try:
from official.recommendation import neumf_model # pylint: disable=g-import-not-at-top
except ImportError as e:
if 'neumf_model' not in e.message:
raise
raise ImportError('To use the experimental NCF model, you must clone the '
'repo https://github.com/tensorflow/models and add '
'tensorflow/models to the PYTHONPATH.')
del nclass
users, items, _ = inputs
params = {
'num_users': _NUM_USERS_20M,
'num_items': _NUM_ITEMS_20M,
'model_layers': (256, 256, 128, 64),
'mf_dim': 64,
'mf_regularization': 0,
'mlp_reg_layers': (0, 0, 0, 0),
'use_tpu': False
}
if self.data_type == tf.float32:
keras_model = neumf_model.construct_model(users, items, params)
logits = keras_model.output
else:
assert self.data_type == tf.float16
tf.keras.backend.set_floatx('float16')
# We cannot rely on the variable_scope's fp16 custom getter here, because
# the NCF model uses keras layers, which ignore variable scopes. So we use
# a variable_creator_scope instead.
with tf.variable_creator_scope(_fp16_variable_creator):
keras_model = neumf_model.construct_model(users, items, params)
logits = tf.cast(keras_model.output, tf.float32)
return model.BuildNetworkResult(logits=logits, extra_info=None)
def build_network(self, inputs, phase_train=True, nclass=1001,
data_type=tf.float32):
try:
from official.recommendation import neumf_model # pylint: disable=g-import-not-at-top
except ImportError as e:
if 'neumf_model' not in e.message:
raise
raise ImportError('To use the experimental NCF model, you must clone the '
'repo https://github.com/tensorflow/models and add '
'tensorflow/models to the PYTHONPATH.')
del nclass
if data_type != tf.float32:
raise ValueError('NCF model only supports float32 for now.')
users, items = inputs
params = {
'num_users': _NUM_USERS_20M,
'num_items': _NUM_ITEMS_20M,
'model_layers': (256, 256, 128, 64),
'mf_dim': 64,
'mf_regularization': 0,
'mlp_reg_layers': (0, 0, 0, 0),
}
logits = neumf_model.construct_model(users, items, params)
return model.BuildNetworkResult(logits=logits, extra_info=None)
def build_network(self, inputs, phase_train=True, nclass=29):
"""Builds the forward pass of the deepspeech2 model.
Args:
inputs: The input list of the model.
phase_train: True during training. False during evaluation.
nclass: Number of classes that the input spectrogram can belong to.
Returns:
A BuildNetworkResult which contains the logits and model-specific extra
information.
"""
inputs = inputs[0] # Get the spectrogram feature.
# Two cnn layers.
inputs = self._conv_bn_layer(
inputs,
padding=(20, 5),
filters=DeepSpeech2Model.CONV_FILTERS,
kernel_size=(41, 11),
strides=(2, 2),
layer_id=1,
training=phase_train)
inputs = self._conv_bn_layer(
inputs,
padding=(10, 5),
filters=DeepSpeech2Model.CONV_FILTERS,
kernel_size=(21, 11),
strides=(2, 1),
layer_id=2,
training=phase_train)
# output of conv_layer2 with the shape of
# [batch_size (N), times (T), features (F), channels (C)].
# Convert the conv output to rnn input.
# batch_size = tf.shape(inputs)[0]
feat_size = inputs.get_shape().as_list()[2]
inputs = tf.reshape(
inputs,
[self.batch_size, -1, feat_size * DeepSpeech2Model.CONV_FILTERS])
# RNN layers.
rnn_cell = DeepSpeech2Model.SUPPORTED_RNNS[self.rnn_type]
for layer_counter in xrange(self.num_rnn_layers):
# No batch normalization on the first layer.
use_batch_norm = (layer_counter != 0)
inputs = self._rnn_layer(inputs, rnn_cell, self.rnn_hidden_size,
layer_counter + 1, use_batch_norm,
self.is_bidirectional, phase_train)
# FC layer with batch norm.
inputs = self._batch_norm(inputs, phase_train)
logits = tf.layers.dense(inputs, nclass, use_bias=self.use_bias)
return model_lib.BuildNetworkResult(logits=logits, extra_info=None)
def build_network(self,
images,
phase_train=True,
nclass=1001,
data_type=tf.float32):
try:
from official.recommendation import neumf_model # pylint: disable=g-import-not-at-top
except ImportError:
raise ImportError(
'To use the experimental NCF model, you must clone the '
'repo https://github.com/tensorflow/models and add '
'tensorflow/models to the PYTHONPATH.')
del nclass
if data_type != tf.float32:
raise ValueError('NCF model only supports float32 for now.')
batch_size = int(images.shape[0])
# Create synthetic users and items. tf_cnn_benchmarks only passes images to
# this function, which we cannot use in the NCF model. We use functions as
# initializers for XLA compatibility.
def users_init_val():
return tf.random_uniform((batch_size, ),
minval=0,
maxval=_NUM_USERS_20M,
dtype=tf.int32)
users = tf.Variable(users_init_val,
dtype=tf.int32,
trainable=False,
collections=[tf.GraphKeys.LOCAL_VARIABLES],
name='synthetic_users')
def items_init_val():
return tf.random_uniform((batch_size, ),
minval=0,
maxval=_NUM_ITEMS_20M,
dtype=tf.int32)
items = tf.Variable(items_init_val,
dtype=tf.int32,
trainable=False,
collections=[tf.GraphKeys.LOCAL_VARIABLES],
name='synthetic_items')
params = {
'num_users': _NUM_USERS_20M,
'num_items': _NUM_ITEMS_20M,
'model_layers': (256, 256, 128, 64),
'mf_dim': 64,
'mf_regularization': 0,
'mlp_reg_layers': (0, 0, 0, 0),
}
logits = neumf_model.construct_model(users, items, params)
return model.BuildNetworkResult(logits=logits, extra_info=None)
def manually_compute_losses(self, inputs, num_workers, params):
with tf.Graph().as_default(), tf.device('/cpu:0'):
a = tf.Variable(self.VAR_A_INITIAL_VALUE, name='A')
b = tf.Variable(self.VAR_B_INITIAL_VALUE, name='B')
inputs_placeholder = tf.placeholder(tf.float32, (None, 1, 1, 1),
name='inputs_placeholder')
inputs_reshaped = tf.reshape(inputs_placeholder, (-1, 1))
loss = self.loss_function(
None,
model.BuildNetworkResult(logits=inputs_reshaped * a * b,
extra_info=None))
return manually_compute_losses(inputs, inputs_placeholder, loss,
num_workers, params)
def build_network(self, images, phase_train=True, nclass=1001,
data_type=tf.float32):
# pylint: disable=g-import-not-at-top
try:
from official.resnet.r1.imagenet_main import ImagenetModel
except ImportError:
tf.logging.fatal('Please include tensorflow/models to the PYTHONPATH.')
raise
images = tf.cast(images, data_type)
model_class = ImagenetModel(resnet_size=self.resnet_size,
resnet_version=self.version,
# The official model dtype seems to be ignored,
# as the dtype it uses is the dtype of the input
# images. Doesn't hurt to set it though.
dtype=data_type)
logits = model_class(images, phase_train)
logits = tf.cast(logits, tf.float32)
return model_lib.BuildNetworkResult(logits=logits, extra_info=None)
