Tensorflow Recommenders

TensorFlow User Group Summit SSA
Une Session sur TensorFlow Recommenders- une bibliothèque de composants que vous pouvez utiliser pour créer, vous l'avez deviné, des systèmes de recommandation.

Transcript

1. 1 Yannick Serge Obam ML GDE Tensorflow recommenders 1 @OBAMSerge Construire les systemes de recommandations avec TensorFlow
2. 2 C’est quoi un Système de recommandation 2
3. • Filtrage collaboraif • Filtrage basé sur le contenu • Hybride 03 philosophies de modélisation
4. Filtrage collaboratif 4 Source :https://interstices.info/ • User-based • Item-based • Factorisation de matrice
5. Filtrage basé sur le contenu 5 Source :https://interstices.info/
6. Approche hybride 6 Source :https://interstices.info/
7. 7 Tensorflow recommenders - TFRS 7
8. TensorFlow recommenders • Bibliothèque permettant de créer des modèles système pour les outils de recommandation. • S’appuie sur tensorflow 2.0 et Keras 8
9. Conduit par des besoins pratiques • Apprentissage multitache • Modelisation des interactions des features • Entrainement sur TPU 9
10. Facilité d’utilisation • Construction du Modele par composition des couches Keras • Model fitting via keras 10
11. Facilité d’utilisation • Construction du Modele par composition des couches Keras • Model fitting via keras 11
12. Two-Tower Recommender Example 12
13. import tensorflow_datasets as tfds import tensorflow_recommenders as tfrs # Data on ratings. ratings = tfds.load("movielens/100k-ratings", split="train") # Movie features. movies = tfds.load("movielens/100k-movies", split="train")
14. import tensorflow_datasets as tfds import tensorflow_recommenders as tfrs # Data on ratings. ratings = tfds.load("movielens/100k-ratings", split="train") # Movie features. movies = tfds.load("movielens/100k-movies", split="train") # The user and movie models can be arbitrary Keras models. user_model = tf.keras.Sequential([ tf.keras.layers.Embedding(1000, 32), tf.keras.layers.Dense(64, activation=”relu”) ]) movie_model = tf.keras.layers.Embedding(1700, 64)
15. import tensorflow_datasets as tfds import tensorflow_recommenders as tfrs # Data on ratings. ratings = tfds.load("movielens/100k-ratings", split="train") # Movie features. movies = tfds.load("movielens/100k-movies", split="train") # The user and movie models can be arbitrary Keras models. user_model = tf.keras.Sequential([ tf.keras.layers.Embedding(1000, 32), tf.keras.layers.Dense(64, activation=”relu”) ]) movie_model = tf.keras.layers.Embedding(1700, 64) for x in ratings.take(1).as_numpy_iterator(): pprint.pprint(x)
19. 19 model = MovielensModel( user_model=user_model, movie_model=movie_model, # The retrieval task will optimize for retrieving the best movies. task=tfrs.tasks.Retrieval( # Model retrieval accuracy measured across all recommendable movies. metrics=tfrs.metrics.FactorizedTopK( candidates=movies.batch(128).map(movie_model) )) )
21. 21 # Train the model. model.compile(optimizer=tf.keras.optimizers.Adagrad(0.5)) model.fit(ratings.batch(4096), epochs=3) # Evaluate retrieval performance. model.evaluate(ratings.batch(4096))
24. 24 Resources: TensorFlow Recommenders homepage → https://goo.gle/2IJAkrK TensorFlow Recommenders repository on GitHub → https://goo.gle/3ef3tYy Building a retrieval model using TensorFlow Recommenders → https://goo.gle/3yUAqkK https://www.tensorflow.org/guide/ https://www.tensorflow.org/tutorials Question & Bugs: https://github.com/tensorflow/tensorflow/issues Merci