TensorFlow Recommenders Addons is a library for building recommender system models using TensorFlow.
It makes TensorFlow support dynamic sparse weights training for large scale recommendation systems (Recommenders), which are one of most common and impactful use cases in the industry. We hope to encourage sharing of best practices in the industry, get consensus and product feedback to help evolve TensorFlow better, and facilitate the contributions of RFCs and PRs in this domain.
Make sure you have TensorFlow 2.x installed, and install from pip:
pip install tensorflow-recommenders-addons
Have a look at our tutorials and API reference.
Building a factorization model for the Movielens 100K dataset is very simple (Colab):
import tensorflow as tf
import tensorflow_recommenders_addons as tfra
# graph defination
x, labels = dataset
w = tfra.get_variable(name="dynamic_embeddings",
devices=[
"/job:ps/replica:0/task:0/CPU:0",
"/job:ps/replica:0/task:1/CPU:0"
],
initializer=tf.random_normal_initializer(0, 0.005),
dim=1)
z = tfra.embedding_lookup(params=w, ids=x, name="wide-sparse-weights")
# graph defination
opt = tfra.DynamicEmbeddingOptimizer(tf.train.AdamOptimizer(0.001))
update = opt.minimize(loss)
saver = tf.train.Saver()
# training loop
with tf.Session() as sess:
for _ in range(100):
sess.run(update)
saver.save(sess, './ckpt/ckpt')