I am new to the Subclassing API (and fairly new to ML in general) and I am learning to use the TensorFlow Recommenders library and have been working through the documentation. I would like to know how to run and review the output of the individual sub models that feed into the final model. I have recreated an excerpt of the first sub model (I have compiled and fit the full model and it runs okay). I would like to run this UserFeaturesModel in isolation, to be able to inspect the concatenated data that it outputs so that I can determine if it is doing what it should, what its shape is, and to understand more about how it feeds into the next sub model. But I am not sure how to compile and fit this model, as it requires a loss function when I try to compile, but is this model not just simply running embeddings and therefore has no loss function? Apologies if this is a stupid question but I have spent days trying to figure this out and have tried looking through previous questions on Stack and watched many tutorials - please could someone show me to run this model and generate an output that I can review? # imports import tensorflow as tf import tensorflow_recommenders as tfrs import tensorflow_datasets as tfds import numpy as np # load datasets from tf ratings = tfds.load("movielens/100k-ratings", split="train") movies = tfds.load("movielens/100k-movies", split="train") # keep required feature columns for each dataset ratings = ratings.map(lambda x: { "movie_title": x["movie_title"], "user_id": x["user_id"], "user_rating": x["user_rating"], "timestamp": x["timestamp"], "raw_user_age": x["raw_user_age"], }) movies = movies.map(lambda x: x["movie_title"]) # create datasets of user_ids and movie_titles user_ids = ratings.batch(1000).map(lambda x: x['user_id']) movie_titles = movies.batch(1000) # find unique vocabs from above datasets unique_user_ids = np.unique(np.concatenate(list(user_ids))) unique_movie_titles = np.unique(np.concatenate(list(movie_titles))) # create timestamp array for bucketization and normalization timestamps = np.concatenate(list(ratings.map(lambda x: x["timestamp"]).batch(100))) max_timestamp = timestamps.max() min_timestamp = timestamps.min() timestamp_buckets = np.linspace(min_timestamp, max_timestamp, num = 2000) # create user_age array for normalization user_age = np.concatenate(list(ratings.map(lambda x: x["raw_user_age"]).batch(100))) # vocab sizes for embedding user_vocab_size = len(unique_user_ids) + 1 movie_vocab_size = len(unique_movie_titles) + 1 timestamp_buckets_size = len(timestamp_buckets) + 1 # set random seed tf.random.set_seed(42) # shuffle and split data 80:20 shuffled_data = ratings.shuffle(100000, seed=42, reshuffle_each_iteration=False) train = shuffled_data.take(80000) test = shuffled_data.skip(80000).take(20000) # batch and cache the train and test datasets cached_train = train.shuffle(80000).batch(512).cache() cached_test = test.batch(512).cache() # define model class UserFeaturesModel(tf.keras.Model): # model construction and attributes def __init__(self): super().__init__() embedding_dim = 32 # user_id feature embeddings self.user_id_embeddings = tf.keras.Sequential([ tf.keras.layers.StringLookup(vocabulary=unique_user_ids, mask_token=None), tf.keras.layers.Embedding(user_vocab_size, embedding_dim) ]) # timestamp bucket featrure embeddings self.timestamp_bucket_embeddings = tf.keras.Sequential([ tf.keras.layers.Discretization(timestamp_buckets.tolist()), tf.keras.layers.Embedding(timestamp_buckets_size, embedding_dim) ]) # timestamp normalization self.normalized_timestamp = tf.keras.layers.Normalization(axis=None) self.normalized_timestamp.adapt(timestamps) # age normalization self.normalized_user_age = tf.keras.layers.Normalization(axis=None) self.normalized_user_age.adapt(user_age) # model call and forward pass def call(self, inputs): return tf.concat([ self.user_id_embeddings(inputs['user_id']), self.timestamp_bucket_embeddings(inputs['timestamp']), tf.reshape(self.normalized_timestamp(inputs['timestamp']),(-1,1)), tf.reshape(self.normalized_user_age(inputs['raw_user_age']),(-1,1)) ], axis=1) # instantiate model model = UserFeaturesModel() # run model - not sure how to compile, fit and run model and inspect output? I would really appreciate any help as I am trying to get a better understanding of the Subclassing API. Continue reading...
