def train_item_graph(subgraph):
_, subgraph['p_item_vec'] = LatentFactor(
l2_reg=l2_reg,
init='normal',
id_=subgraph['p_item_id'],
shape=[total_items, dim_embed],
subgraph=subgraph,
scope='item')
_, subgraph['p_item_bias'] = LatentFactor(l2_reg=l2_reg,
init='zero',
id_=subgraph['p_item_id'],
shape=[total_items, 1],
subgraph=subgraph,
scope='item_bias')
_, subgraph['n_item_vec'] = LatentFactor(
l2_reg=l2_reg,
init='normal',
id_=subgraph['n_item_id'],
shape=[total_items, dim_embed],
subgraph=subgraph,
scope='item')
_, subgraph['n_item_bias'] = LatentFactor(l2_reg=l2_reg,
init='zero',
id_=subgraph['n_item_id'],
shape=[total_items, 1],
subgraph=subgraph,
scope='item_bias')
def _build_item_extractions(self, train=True):
if train:
self._item_vec = TemporalLatentFactor(
l2_reg=self._l2_reg,
init=self._pretrained_item_embeddings,
ids=self._item_id_input,
mlp_dims=self._item_dims,
shape=[self._max_item, self._dim_embed],
train=True,
scope='item',
reuse=False)
self._item_bias = LatentFactor(l2_reg=self._l2_reg,
init='zero',
ids=self._item_id_input,
shape=[self._max_item, 1],
scope='item_bias',
reuse=False)
self._loss_nodes += [self._item_vec, self._item_bias]
else:
self._item_vec_serving = TemporalLatentFactor(
l2_reg=self._l2_reg,
init=self._pretrained_item_embeddings,
ids=self._item_id_serving,
mlp_dims=self._item_dims,
shape=[self._max_item, self._dim_embed],
train=False,
scope='item',
reuse=True)
self._item_bias_serving = LatentFactor(l2_reg=self._l2_reg,
init='zero',
ids=self._item_id_serving,
shape=[self._max_item, 1],
scope='item_bias',
reuse=True)
def train_item_graph(subgraph):
_, subgraph["p_item_vec"] = LatentFactor(
l2_reg=l2_reg,
init="normal",
id_=subgraph["p_item_id"],
shape=[total_items, dim_item_embed],
subgraph=subgraph,
scope="item",
)
_, subgraph["p_item_bias"] = LatentFactor(
l2_reg=l2_reg,
init="zero",
id_=subgraph["p_item_id"],
shape=[total_items, 1],
subgraph=subgraph,
scope="item_bias",
)
_, subgraph["n_item_vec"] = LatentFactor(
l2_reg=l2_reg,
init="normal",
id_=subgraph["n_item_id"],
shape=[total_items, dim_item_embed],
subgraph=subgraph,
scope="item",
)
_, subgraph["n_item_bias"] = LatentFactor(
l2_reg=l2_reg,
init="zero",
id_=subgraph["n_item_id"],
shape=[total_items, 1],
subgraph=subgraph,
scope="item_bias",
)
def _build_item_extractions(self, train=True):
self._add_module('item_vec',
LatentFactor(l2_reg=self._l2_reg, init='normal', ids=self._get_input('item_id', train=train),
shape=[self._max_item, self._dim_embed], scope='item', reuse=not train),
train=train)
self._add_module('item_bias',
LatentFactor(l2_reg=self._l2_reg, init='zero', ids=self._get_input('item_id', train=train),
shape=[self._max_item, 1], scope='item_bias', reuse=not train),
train=train)
def _build_item_extractions(self, train=True):
if train:
self._item_vec = LatentFactor(
l2_reg=self._l2_reg,
init='normal',
ids=self._item_id_input,
shape=[self._max_item, self._dim_embed],
scope='item',
reuse=False)
self._item_bias = LatentFactor(l2_reg=self._l2_reg,
init='zero',
ids=self._item_id_input,
shape=[self._max_item, 1],
scope='item_bias',
reuse=False)
self._loss_nodes += [self._item_vec, self._item_bias]
else:
self._item_vec_serving = LatentFactor(
l2_reg=self._l2_reg,
init='normal',
ids=self._item_id_serving,
shape=[self._max_item, self._dim_embed],
scope='item',
reuse=True)
self._item_bias_serving = LatentFactor(l2_reg=self._l2_reg,
init='zero',
ids=self._item_id_serving,
shape=[self._max_item, 1],
scope='item_bias',
reuse=True)
def _add_itemgraph(subgraph):
item_embedding, _ = LatentFactor(id_=None,
shape=[total_items, dim_embed],
subgraph=subgraph,
scope='item')
item_bias_embedding, _ = LatentFactor(id_=None,
shape=[total_items, 1],
subgraph=subgraph,
scope='item_bias')
subgraph.register_global_operation(
tf.assign(item_embedding, subgraph['item_lf_cache']), 'cache')
subgraph.register_global_operation(
tf.assign(item_bias_embedding, subgraph['item_bias_cache']),
'cache')
def user_graph(subgraph):
_, subgraph['user_vec'] = LatentFactor(l2_reg=l2_reg,
init='normal',
id_=subgraph['user_id'],
shape=[total_users, dim_user_embed],
subgraph=subgraph,
scope='user')
def _add_usergraph(subgraph):
user_embedding, _ = LatentFactor(id_=None,
shape=[total_users, dim_embed],
subgraph=subgraph,
scope='user')
subgraph.register_global_operation(
tf.assign(user_embedding, subgraph['user_lf_cache']), 'cache')
def user_graph(subgraph):
_, user_gender = LatentFactor(
l2_reg=l2_reg_embed,
shape=[user_dict['gender'], dim_user_embed['gender']],
id_=subgraph['user_gender'],
subgraph=subgraph,
init='normal',
scope='user_gender')
_, user_geo = LatentFactor(
l2_reg=l2_reg_embed,
shape=[user_dict['geo'], dim_user_embed['geo']],
id_=subgraph['user_geo'],
subgraph=subgraph,
init='normal',
scope='user_geo')
subgraph['user_vec'] = tf.concat([user_gender, user_geo], axis=1)
def _build_item_extractions(self, train=True):
super(ConcatVisualBPR, self)._build_item_extractions(train)
if train:
self._loss_nodes.remove(self._p_item_vec)
self._loss_nodes.remove(self._n_item_vec)
self._p_item_lf = LatentFactor(
init='normal',
l2_reg=self._l2_reg,
ids=self._p_item_id_input,
shape=[self._max_item, self._dim_embed - self._dim_ve],
scope='item',
reuse=False)
self._p_item_vf = MultiLayerFC(
in_tensor=self._p_item_vfeature_input,
dims=[self._dim_ve],
scope='item_MLP',
reuse=False)
self._n_item_lf = LatentFactor(
init='normal',
l2_reg=self._l2_reg,
ids=self._n_item_id_input,
shape=[self._max_item, self._dim_embed - self._dim_ve],
scope='item',
reuse=True)
self._n_item_vf = MultiLayerFC(
in_tensor=self._n_item_vfeature_input,
dims=[self._dim_ve],
scope='item_MLP',
reuse=True)
else:
self._item_lf_serving = LatentFactor(
init='normal',
l2_reg=self._l2_reg,
ids=self._item_id_serving,
shape=[self._max_item, self._dim_embed - self._dim_ve],
scope='item',
reuse=True)
self._item_vf_serving = MultiLayerFC(
in_tensor=self._item_vfeature_serving,
dims=[self._dim_ve],
scope='item_MLP',
reuse=True)
def censor_user_vec(subgraph):
user_embedding, _ = LatentFactor(l2_reg=None,
init='normal',
id_=None,
shape=[total_users, dim_user_embed],
scope='user')
user_censor_ops = censor_vec(user_embedding, subgraph['user_id'])
subgraph.register_global_operation(user_censor_ops, 'censor_embedding')
def item_graph(subgraph):
_, subgraph['seq_item_vec'] = LatentFactor(
l2_reg=l2_reg_embed,
init='normal',
id_=subgraph['seq_item_id'],
shape=[total_items, dim_item_embed],
subgraph=subgraph,
scope='item')
def item_graph(subgraph):
_, subgraph["seq_item_vec"] = LatentFactor(
l2_reg=l2_reg_embed,
init="normal",
id_=subgraph["seq_item_id"],
shape=[total_items, dim_item_embed],
subgraph=subgraph,
scope="item",
)
def user_graph(subgraph):
_, subgraph["user_vec"] = LatentFactor(
l2_reg=l2_reg,
init="normal",
id_=subgraph["user_id"],
shape=[total_users, dim_user_embed],
subgraph=subgraph,
scope="user",
)
def user_graph(subgraph):
_, user_gender = LatentFactor(
l2_reg=l2_reg_embed,
shape=[user_dict["gender"], dim_user_embed["gender"]],
id_=subgraph["user_gender"],
subgraph=subgraph,
init="normal",
scope="user_gender",
)
_, user_geo = LatentFactor(
l2_reg=l2_reg_embed,
shape=[user_dict["geo"], dim_user_embed["geo"]],
id_=subgraph["user_geo"],
subgraph=subgraph,
init="normal",
scope="user_geo",
)
subgraph["user_vec"] = tf.concat([user_gender, user_geo], axis=1)
def _build_user_extractions(self, train=True):
if train:
self._user_vec = LatentFactor(
l2_reg=self._l2_reg,
init='normal',
ids=self._user_id_input,
shape=[self._max_user, self._dim_embed],
scope='user',
reuse=False)
self._loss_nodes += [self._user_vec]
else:
self._user_vec_serving = LatentFactor(
l2_reg=self._l2_reg,
init='normal',
ids=self._user_id_serving,
shape=[self._max_user, self._dim_embed],
scope='user',
reuse=True)
def censor_user_vec(subgraph):
user_embedding, _ = LatentFactor(
l2_reg=None,
init="normal",
id_=None,
shape=[total_users, dim_user_embed],
scope="user",
)
user_censor_ops = censor_vec(user_embedding, subgraph["user_id"])
subgraph.register_global_operation(user_censor_ops, "censor_embedding")
def _build_user_extractions(self, train=True):
self._add_module('user_vec',
LatentFactor(l2_reg=self._l2_reg,
init='normal',
ids=self._get_input('user_id',
train=train),
shape=[self._max_user, self._dim_embed],
scope='user',
reuse=(not train)),
train=train)
def censor_item_vec(subgraph):
item_embedding, _ = LatentFactor(l2_reg=None,
init='normal',
id_=None,
shape=[total_items, dim_item_embed],
subgraph=subgraph,
scope='item')
item_censor_ops = censor_vec(
item_embedding,
tf.concat([subgraph['p_item_id'], subgraph['n_item_id']], axis=0))
subgraph.register_global_operation(item_censor_ops, 'censor_embedding')
def censor_item_vec(subgraph):
item_embedding, _ = LatentFactor(
l2_reg=None,
init="normal",
id_=None,
shape=[total_items, dim_item_embed],
subgraph=subgraph,
scope="item",
)
item_censor_ops = censor_vec(
item_embedding,
tf.concat([subgraph["p_item_id"], subgraph["n_item_id"]], axis=0),
)
subgraph.register_global_operation(item_censor_ops, "censor_embedding")
def _build_extra_extractions(self, train=True):
if train:
self._add_module('p_extra_vec',
LatentFactor(
l2_reg=self._l2_reg,
init='normal',
ids=self._get_input('p_item_timestamp',
train=train),
shape=[self._max_item, self._dim_embed],
scope='item',
reuse=not train),
train=train)
self._add_module('n_extra_vec',
LatentFactor(
l2_reg=self._l2_reg,
init='normal',
ids=self._get_input('n_item_timestamp',
train=train),
shape=[self._max_item, self._dim_embed],
scope='item',
reuse=not train),
train=train)
def _build_item_extractions(self, train=True):
self._add_module('item_vec',
TemporalLatentFactor(
l2_reg=self._l2_reg,
init=self._pretrained_item_embeddings,
ids=self._get_input('item_id', train=train),
mlp_dims=self._item_dims,
shape=[self._max_item, self._dim_embed],
train=train,
scope='item',
reuse=not train),
train=train)
self._add_module('item_bias',
LatentFactor(l2_reg=self._l2_reg,
init='zero',
ids=self._get_input('item_id',
train=train),
shape=[self._max_item, 1],
scope='item_bias',
reuse=not train),
train=train)
class ItrMLP(Recommender):
def __init__(self,
batch_size,
dim_embed,
max_user,
max_item,
pretrained_user_embeddings,
pretrained_item_embeddings,
user_dims,
item_dims,
test_batch_size=None,
l2_reg=None,
opt='SGD',
sess_config=None):
self._dim_embed = dim_embed
self._pretrained_user_embeddings = pretrained_user_embeddings
self._pretrained_item_embeddings = pretrained_item_embeddings
self._user_dims = user_dims
self._item_dims = item_dims
super(ItrMLP, self).__init__(batch_size=batch_size,
test_batch_size=test_batch_size,
max_user=max_user,
max_item=max_item,
l2_reg=l2_reg,
opt=opt,
sess_config=sess_config)
def _initialize(self, init_dict):
super(ItrMLP, self)._initialize(init_dict=init_dict)
print(colored('[Pretrain user MLP into identity]', 'blue'))
self._user_vec.pretrain_mlp_as_identity(self._sess)
print(colored('[Pretrain item MLP into identity]', 'blue'))
self._item_vec.pretrain_mlp_as_identity(self._sess)
def update_embeddings(self):
self._user_vec.forward_update_embeddings(self._sess)
self._item_vec.forward_update_embeddings(self._sess)
def _input_mappings(self, batch_data, train):
if train:
return {
self._user_id_input: batch_data['user_id_input'],
self._item_id_input: batch_data['item_id_input'],
self._labels: batch_data['labels']
}
else:
return {
self._user_id_serving: batch_data['user_id_input'],
self._item_id_serving: batch_data['item_id_input']
}
def _build_user_inputs(self, train=True):
if train:
self._user_id_input = self._input(dtype='int32',
shape=[self._batch_size],
name='user_id_input')
else:
self._user_id_serving = self._input(dtype='int32',
shape=[self._test_batch_size],
name='user_id_serving')
def _build_item_inputs(self, train=True):
if train:
self._item_id_input = self._input(dtype='int32',
shape=[self._batch_size],
name='item_id_input')
else:
self._item_id_serving = self._input(dtype='int32',
shape=[self._test_batch_size],
name='item_id_serving')
def _build_extra_inputs(self, train=True):
if train:
self._labels = self._input(dtype='float32',
shape=[self._batch_size],
name='labels')
def _build_user_extractions(self, train=True):
if train:
self._user_vec = TemporalLatentFactor(
l2_reg=self._l2_reg,
init=self._pretrained_user_embeddings,
ids=self._user_id_input,
mlp_dims=self._user_dims,
shape=[self._max_user, self._dim_embed],
train=True,
scope='user',
reuse=False)
self._loss_nodes += [self._user_vec]
else:
self._user_vec_serving = TemporalLatentFactor(
l2_reg=self._l2_reg,
init=self._pretrained_user_embeddings,
ids=self._user_id_serving,
mlp_dims=self._user_dims,
shape=[self._max_user, self._dim_embed],
train=False,
scope='user',
reuse=True)
def _build_item_extractions(self, train=True):
if train:
self._item_vec = TemporalLatentFactor(
l2_reg=self._l2_reg,
init=self._pretrained_item_embeddings,
ids=self._item_id_input,
mlp_dims=self._item_dims,
shape=[self._max_item, self._dim_embed],
train=True,
scope='item',
reuse=False)
self._item_bias = LatentFactor(l2_reg=self._l2_reg,
init='zero',
ids=self._item_id_input,
shape=[self._max_item, 1],
scope='item_bias',
reuse=False)
self._loss_nodes += [self._item_vec, self._item_bias]
else:
self._item_vec_serving = TemporalLatentFactor(
l2_reg=self._l2_reg,
init=self._pretrained_item_embeddings,
ids=self._item_id_serving,
mlp_dims=self._item_dims,
shape=[self._max_item, self._dim_embed],
train=False,
scope='item',
reuse=True)
self._item_bias_serving = LatentFactor(l2_reg=self._l2_reg,
init='zero',
ids=self._item_id_serving,
shape=[self._max_item, 1],
scope='item_bias',
reuse=True)
def _build_default_interactions(self, train=True):
if train:
self._interaction_train = PointwiseMSE(
user=self._user_vec.get_outputs()[0],
item=self._item_vec.get_outputs()[0],
item_bias=self._item_bias.get_outputs()[0],
labels=self._labels,
a=1.0,
b=1.0,
sigmoid=True,
train=True,
scope='PointwiseMSE',
reuse=False)
self._loss_nodes.append(self._interaction_train)
else:
self._interaction_serve = PointwiseMSE(
user=self._user_vec_serving.get_outputs()[0],
item=self._item_vec_serving.get_outputs()[0],
item_bias=self._item_bias_serving.get_outputs()[0],
sigmoid=True,
train=False,
batch_serving=False,
scope='PointwiseMSE',
reuse=True)
def _build_serving_graph(self):
super(ItrMLP, self)._build_serving_graph()
self._scores = self._interaction_serve.get_outputs()[0]
def _build_extra_extractions(self, train=True):
if train:
self._add_module('labels_vec',
LatentFactor(
l2_reg=self._l2_reg,
init='normal',
ids=self._get_input('labels', train=train),
shape=[self._max_item, self._dim_embed],
scope='item',
reuse=tf.AUTO_REUSE),
train=True)
self._add_module('source_system_tab_vec',
LatentFactor(
l2_reg=self._l2_reg,
init='normal',
ids=self._get_input('source_system_tab',
train=train),
shape=[self._max_item, self._dim_embed],
scope='item',
reuse=tf.AUTO_REUSE),
train=True)
self._add_module('source_screen_name_vec',
LatentFactor(
l2_reg=self._l2_reg,
init='normal',
ids=self._get_input('source_screen_name',
train=train),
shape=[self._max_item, self._dim_embed],
scope='item',
reuse=tf.AUTO_REUSE),
train=True)
self._add_module('source_type_vec',
LatentFactor(
l2_reg=self._l2_reg,
init='normal',
ids=self._get_input('source_type',
train=train),
shape=[self._max_item, self._dim_embed],
scope='item',
reuse=tf.AUTO_REUSE),
train=True)
else:
self._add_module('labels_vec',
LatentFactor(
l2_reg=self._l2_reg,
init='normal',
ids=self._get_input('labels', train=train),
shape=[self._max_item, self._dim_embed],
scope='item',
reuse=True),
train=False)
self._add_module('source_system_tab_vec',
LatentFactor(
l2_reg=self._l2_reg,
init='normal',
ids=self._get_input('source_system_tab',
train=train),
shape=[self._max_item, self._dim_embed],
scope='item',
reuse=True),
train=False)
self._add_module('source_screen_name_vec',
LatentFactor(
l2_reg=self._l2_reg,
init='normal',
ids=self._get_input('source_screen_name',
train=train),
shape=[self._max_item, self._dim_embed],
scope='item',
reuse=True),
train=False)
self._add_module('source_type_vec',
LatentFactor(
l2_reg=self._l2_reg,
init='normal',
ids=self._get_input('source_type',
train=train),
shape=[self._max_item, self._dim_embed],
scope='item',
reuse=True),
train=False)
def _build_item_extractions(self, train=True):
if train:
self._add_module(
'p_item_lf',
LatentFactor(
init='normal',
l2_reg=self._l2_reg,
ids=self._get_input('p_item_id'),
shape=[self._max_item, self._dim_embed - self._dim_ve],
scope='item',
reuse=False))
self._add_module(
'p_item_vf',
MultiLayerFC(in_tensor=self._get_input('p_item_vfeature'),
dims=[self._dim_ve],
scope='item_MLP',
reuse=False))
self._add_module(
'p_item_bias',
LatentFactor(l2_reg=self._l2_reg,
init='zero',
ids=self._get_input('p_item_id'),
shape=[self._max_item, 1],
scope='item_bias',
reuse=False))
self._add_module(
'n_item_lf',
LatentFactor(
init='normal',
l2_reg=self._l2_reg,
ids=self._get_input('n_item_id'),
shape=[self._max_item, self._dim_embed - self._dim_ve],
scope='item',
reuse=True))
self._add_module(
'n_item_vf',
MultiLayerFC(in_tensor=self._get_input('n_item_vfeature'),
dims=[self._dim_ve],
scope='item_MLP',
reuse=True))
self._add_module(
'n_item_bias',
LatentFactor(l2_reg=self._l2_reg,
init='zero',
ids=self._get_input('n_item_id'),
shape=[self._max_item, 1],
scope='item_bias',
reuse=True))
else:
self._add_module('item_lf',
LatentFactor(init='normal',
l2_reg=self._l2_reg,
ids=self._get_input('item_id',
train=train),
shape=[
self._max_item,
self._dim_embed - self._dim_ve
],
scope='item',
reuse=True),
train=False)
self._add_module('item_vf',
MultiLayerFC(in_tensor=self._get_input(
'item_vfeature', train=train),
dims=[self._dim_ve],
scope='item_MLP',
reuse=True),
train=False)
self._add_module('item_bias',
LatentFactor(l2_reg=self._l2_reg,
init='zero',
ids=self._get_input('item_id',
train=train),
shape=[self._max_item, 1],
scope='item_bias',
reuse=True),
train=False)
def _build_user_extractions(self, train=True):
if train:
self._add_module('song_id_vec',
LatentFactor(
l2_reg=self._l2_reg,
init='normal',
ids=self._get_input('song_id', train=train),
shape=[self._max_item, self._dim_embed],
scope='user',
reuse=tf.AUTO_REUSE),
train=True)
self._add_module('artist_vec',
LatentFactor(
l2_reg=self._l2_reg,
init='normal',
ids=self._get_input('artist', train=train),
shape=[self._max_item, self._dim_embed],
scope='user',
reuse=tf.AUTO_REUSE),
train=True)
self._add_module('genre_vec',
LatentFactor(
l2_reg=self._l2_reg,
init='normal',
ids=self._get_input('genre', train=train),
shape=[self._max_item, self._dim_embed],
scope='user',
reuse=tf.AUTO_REUSE),
train=True)
self._add_module('language_vec',
LatentFactor(
l2_reg=self._l2_reg,
init='normal',
ids=self._get_input('language', train=train),
shape=[self._max_item, self._dim_embed],
scope='user',
reuse=tf.AUTO_REUSE),
train=True)
self._add_module('lyricist_vec',
LatentFactor(
l2_reg=self._l2_reg,
init='normal',
ids=self._get_input('lyricist', train=train),
shape=[self._max_item, self._dim_embed],
scope='user',
reuse=tf.AUTO_REUSE),
train=True)
self._add_module('composer_vec',
LatentFactor(
l2_reg=self._l2_reg,
init='normal',
ids=self._get_input('composer', train=train),
shape=[self._max_item, self._dim_embed],
scope='user',
reuse=tf.AUTO_REUSE),
train=True)
else:
self._add_module('song_id_vec',
LatentFactor(
l2_reg=self._l2_reg,
init='normal',
ids=self._get_input('song_id', train=train),
shape=[self._max_item, self._dim_embed],
scope='user',
reuse=True),
train=False)
self._add_module('artist_vec',
LatentFactor(
l2_reg=self._l2_reg,
init='normal',
ids=self._get_input('artist', train=train),
shape=[self._max_item, self._dim_embed],
scope='user',
reuse=True),
train=False)
self._add_module('genre_vec',
LatentFactor(
l2_reg=self._l2_reg,
init='normal',
ids=self._get_input('genre', train=train),
shape=[self._max_item, self._dim_embed],
scope='item',
reuse=True),
train=False)
self._add_module('language_vec',
LatentFactor(
l2_reg=self._l2_reg,
init='normal',
ids=self._get_input('language', train=train),
shape=[self._max_item, self._dim_embed],
scope='user',
reuse=True),
train=False)
self._add_module('lyricist_vec',
LatentFactor(
l2_reg=self._l2_reg,
init='normal',
ids=self._get_input('lyricist', train=train),
shape=[self._max_item, self._dim_embed],
scope='user',
reuse=True),
train=False)
self._add_module('composer_vec',
LatentFactor(
l2_reg=self._l2_reg,
init='normal',
ids=self._get_input('composer', train=train),
shape=[self._max_item, self._dim_embed],
scope='user',
reuse=True),
train=False)
class PMF(Recommender):
def __init__(self,
batch_size,
dim_embed,
max_user,
max_item,
test_batch_size=None,
l2_reg=None,
opt='SGD',
sess_config=None):
self._dim_embed = dim_embed
super(PMF, self).__init__(batch_size=batch_size,
test_batch_size=test_batch_size,
max_user=max_user,
max_item=max_item,
l2_reg=l2_reg,
opt=opt,
sess_config=sess_config)
def _input_mappings(self, batch_data, train):
if train:
return {
self._user_id_input: batch_data['user_id_input'],
self._item_id_input: batch_data['item_id_input'],
self._labels: batch_data['labels']
}
else:
return {self._user_id_serving: batch_data['user_id_input']}
def _build_user_inputs(self, train=True):
if train:
self._user_id_input = self._input(dtype='int32',
shape=[self._batch_size],
name='user_id_input')
else:
self._user_id_serving = self._input(dtype='int32',
shape=[None],
name='user_id_serving')
def _build_item_inputs(self, train=True):
if train:
self._item_id_input = self._input(dtype='int32',
shape=[self._batch_size],
name='item_id_input')
else:
self._item_id_serving = None
def _build_extra_inputs(self, train=True):
if train:
self._labels = self._input(dtype='float32',
shape=[self._batch_size],
name='labels')
def _build_user_extractions(self, train=True):
if train:
self._user_vec = LatentFactor(
l2_reg=self._l2_reg,
init='normal',
ids=self._user_id_input,
shape=[self._max_user, self._dim_embed],
scope='user',
reuse=False)
self._loss_nodes += [self._user_vec]
else:
self._user_vec_serving = LatentFactor(
l2_reg=self._l2_reg,
init='normal',
ids=self._user_id_serving,
shape=[self._max_user, self._dim_embed],
scope='user',
reuse=True)
def _build_item_extractions(self, train=True):
if train:
self._item_vec = LatentFactor(
l2_reg=self._l2_reg,
init='normal',
ids=self._item_id_input,
shape=[self._max_item, self._dim_embed],
scope='item',
reuse=False)
self._item_bias = LatentFactor(l2_reg=self._l2_reg,
init='zero',
ids=self._item_id_input,
shape=[self._max_item, 1],
scope='item_bias',
reuse=False)
self._loss_nodes += [self._item_vec, self._item_bias]
else:
self._item_vec_serving = LatentFactor(
l2_reg=self._l2_reg,
init='normal',
ids=self._item_id_serving,
shape=[self._max_item, self._dim_embed],
scope='item',
reuse=True)
self._item_bias_serving = LatentFactor(l2_reg=self._l2_reg,
init='zero',
ids=self._item_id_serving,
shape=[self._max_item, 1],
scope='item_bias',
reuse=True)
def _build_default_interactions(self, train=True):
if train:
self._interaction_train = PointwiseMSE(
user=self._user_vec.get_outputs()[0],
item=self._item_vec.get_outputs()[0],
item_bias=self._item_bias.get_outputs()[0],
labels=self._labels,
a=1.0,
b=1.0,
train=True,
scope='PointwiseMSE',
reuse=False)
self._loss_nodes.append(self._interaction_train)
else:
self._interaction_serve = PointwiseMSE(
user=self._user_vec_serving.get_outputs()[0],
item=self._item_vec_serving.get_outputs()[0],
item_bias=self._item_bias_serving.get_outputs()[0],
train=False,
scope='PointwiseMSE',
reuse=True)
def _build_serving_graph(self):
super(PMF, self)._build_serving_graph()
self._scores = self._interaction_serve.get_outputs()[0]