tensorflow-graph.py

#Before we dive in we need to declare some variables
vocabulary_size = len(artist_lookup)
#How big we want our final vectors to be
embedding_size = 64
#The number of training samples passed per epoch
batch_size = 64
#Number of negative samples to use in NCE [see below]
num_sampled = 16

#BAND2VEC - Tensorflow Time!
import tensorflow as tf
import math 

graph = tf.Graph()

with graph.as_default():
    #Defining variables and functions in a scope is good practice - adds a prefix to the operations
    with tf.name_scope('inputs'):
        #Tensorflow Placeholders are the mouths of the neural network - they will constantly be fed new information
        training_inputs = tf.placeholder(tf.int32, shape=[batch_size])
        training_labels = tf.placeholder(tf.int32, shape=[batch_size, 1])

    #Using the CPU
    with tf.device('/cpu:0'):
        with tf.name_scope('embeddings'):
            #The embeddings - variables are maintained across runs and need to be initialised with a shape and type
            #Each row is a band represented by a vector of length 'embedding_size'
            embeddings = tf.Variable(tf.random_uniform([vocabulary_size, embedding_size], -1.0, 1.0))
            #Like passing muliple indices to a numpy array we get the vectors quickly with this function
            embed = tf.nn.embedding_lookup(embeddings, training_inputs)
        
        with tf.name_scope('weights'):
            #Like embeddings we initialise our weights and also...
            nce_weights = tf.Variable(
                tf.truncated_normal(
                    [vocabulary_size, embedding_size],
                    stddev=1.0 / math.sqrt(embedding_size)))
            
        with tf.name_scope('biases'):
            #...our biases
            nce_biases = tf.Variable(tf.zeros([vocabulary_size]))
            
    with tf.name_scope('loss'):
        #Finally our loss function - see below for an explanation of the Noise Contrastive Estimation Approach
        loss = tf.reduce_mean(
            tf.nn.nce_loss(
                weights=nce_weights,
                biases=nce_biases,
                labels=training_labels,
                inputs=embed,
                num_sampled=num_sampled,
                num_classes=vocabulary_size))
        
    with tf.name_scope('optimizer'):
        optimizer = tf.train.GradientDescentOptimizer(1.0).minimize(loss)