CFDNN_MLP_02.py

import tensorflow as tf
import numpy as np
import pandas as pd
import math
import heapq
from tqdm import tqdm

# Load and prepare our data.
uids, iids, df_train, df_test, df_neg, users, items, item_lookup = load_dataset()


#-------------
# HYPERPARAMS
#-------------

num_neg = 4
epochs = 20
batch_size = 256
learning_rate = 0.001


#-------------------------
# TENSORFLOW GRAPH
#-------------------------

# Set up our Tensorflow graph
graph = tf.Graph()

with graph.as_default():

    # Define input placeholders for user, item and label.
    user = tf.placeholder(tf.int32, shape=(None, 1))
    item = tf.placeholder(tf.int32, shape=(None, 1))
    label = tf.placeholder(tf.int32, shape=(None, 1))

    # User feature embedding
    u_var = tf.Variable(tf.random_normal([len(users), 32], stddev=0.05), name='user_embedding')
    user_embedding = tf.nn.embedding_lookup(u_var, user)

    # Item feature embedding
    i_var = tf.Variable(tf.random_normal([len(items), 32], stddev=0.05), name='item_embedding')
    item_embedding = tf.nn.embedding_lookup(i_var, item)

    # Flatten our user and item embeddings.
    user_embedding = tf.keras.layers.Flatten()(user_embedding)
    item_embedding = tf.keras.layers.Flatten()(item_embedding)

    # Concatenate our two embedding vectors together
    concatenated = tf.keras.layers.concatenate([user_embedding, item_embedding])

    # Add a first dropout layer.
    dropout = tf.keras.layers.Dropout(0.2)(concatenated)

    # Below we add our four hidden layers along with batch
    # normalization and dropouts. We use relu as the activation function.
    layer_1 = tf.keras.layers.Dense(64, activation='relu', name='layer1')(dropout)
    batch_norm1 = tf.keras.layers.BatchNormalization(name='batch_norm1')(layer_1)
    dropout1 = tf.keras.layers.Dropout(0.2, name='dropout1')(batch_norm1)

    layer_2 = tf.keras.layers.Dense(32, activation='relu', name='layer2')(layer_1)
    batch_norm2 = tf.keras.layers.BatchNormalization(name='batch_norm1')(layer_2)
    dropout2 = tf.keras.layers.Dropout(0.2, name='dropout1')(batch_norm2)

    layer_3 = tf.keras.layers.Dense(16, activation='relu', name='layer3')(layer_2)
    layer_4 = tf.keras.layers.Dense(8, activation='relu', name='layer4')(layer_3)

    # Our final single neuron output layer.
    output_layer = tf.keras.layers.Dense(1,
            kernel_initializer="lecun_uniform",
            name='output_layer')(layer_4)

    # Define our loss function as binary cross entropy.
    labels = tf.cast(label, tf.float32)
    logits = output_layer
    loss = tf.reduce_mean(tf.nn.sigmoid_cross_entropy_with_logits(
                labels=labels,
                logits=logits))

    # Train using the Adam optimizer to minimize our loss.
    opt = tf.train.AdamOptimizer(learning_rate = learning_rate)
    step = opt.minimize(loss)

    # Initialize all tensorflow variables.
    init = tf.global_variables_initializer()

session = tf.Session(config=None, graph=graph)
session.run(init)