"""
Module description:
"""
__version__ = '0.3.1'
__author__ = 'Vito Walter Anelli, Claudio Pomo, Daniele Malitesta'
__email__ = 'vitowalter.anelli@poliba.it, claudio.pomo@poliba.it, daniele.malitesta@poliba.it'
import os
import numpy as np
import tensorflow as tf
from tensorflow import keras
os.environ['TF_CPP_MIN_LOG_LEVEL'] = '3'
[docs]class BPRMF_batch_model(keras.Model):
def __init__(self,
factors=200,
learning_rate=0.001,
l_w=0, l_b=0,
num_users=100,
num_items=100,
random_seed=42,
name="NNBPRMF",
**kwargs):
super().__init__(name=name, **kwargs)
tf.random.set_seed(random_seed)
self._factors = factors
self._learning_rate = learning_rate
self._l_w = l_w
self._l_b = l_b
self._num_items = num_items
self._num_users = num_users
self.initializer = tf.initializers.GlorotUniform()
self.Bi = tf.Variable(tf.zeros(self._num_items), name='Bi', dtype=tf.float32)
self.Gu = tf.Variable(self.initializer(shape=[self._num_users, self._factors]), name='Gu', dtype=tf.float32)
self.Gi = tf.Variable(self.initializer(shape=[self._num_items, self._factors]), name='Gi', dtype=tf.float32)
self.optimizer = tf.optimizers.Adam(self._learning_rate)
[docs] @tf.function
def call(self, inputs, training=None):
user, item = inputs
beta_i = tf.squeeze(tf.nn.embedding_lookup(self.Bi, item))
gamma_u = tf.squeeze(tf.nn.embedding_lookup(self.Gu, user))
gamma_i = tf.squeeze(tf.nn.embedding_lookup(self.Gi, item))
xui = beta_i + tf.reduce_sum(gamma_u * gamma_i, 1)
return xui, beta_i, gamma_u, gamma_i
[docs] @tf.function
def train_step(self, batch):
user, pos, neg = batch
with tf.GradientTape() as tape:
# Clean Inference
xu_pos, beta_pos, gamma_u, gamma_pos = self(inputs=(user, pos), training=True)
xu_neg, beta_neg, gamma_u, gamma_neg = self(inputs=(user, neg), training=True)
difference = tf.clip_by_value(xu_pos - xu_neg, -80.0, 1e8)
loss = tf.reduce_sum(tf.nn.softplus(-difference))
# Regularization Component
reg_loss = self._l_w * tf.reduce_sum([tf.nn.l2_loss(gamma_u),
tf.nn.l2_loss(gamma_pos),
tf.nn.l2_loss(gamma_neg)]) \
+ self._l_b * tf.nn.l2_loss(beta_pos) \
+ self._l_b * tf.nn.l2_loss(beta_neg) / 10
# Loss to be optimized
loss += reg_loss
grads = tape.gradient(loss, [self.Bi, self.Gu, self.Gi])
self.optimizer.apply_gradients(zip(grads, [self.Bi, self.Gu, self.Gi]))
return loss
[docs] @tf.function
def predict(self, start, stop, **kwargs):
return self.Bi + tf.matmul(self.Gu[start:stop], self.Gi, transpose_b=True)
[docs] @tf.function
def get_top_k(self, predictions, train_mask, k=100):
return tf.nn.top_k(tf.where(train_mask, predictions, -np.inf), k=k, sorted=True)
[docs] @tf.function
def get_positions(self, predictions, train_mask, items, inner_test_user_true_mask):
predictions = tf.gather(predictions, inner_test_user_true_mask)
train_mask = tf.gather(train_mask, inner_test_user_true_mask)
equal = tf.reshape(items, [len(items), 1])
i = tf.argsort(tf.where(train_mask, predictions, -np.inf), axis=-1,
direction='DESCENDING', stable=False, name=None)
positions = tf.where(tf.equal(equal, i))[:, 1]
return 1 - (positions / tf.reduce_sum(tf.cast(train_mask, tf.int64), axis=1))
[docs] def get_config(self):
raise NotImplementedError