python - How to compile and run a model in Keras subclassing API? - Stack Overflow

admin管理员组

文章数量:1345072

I am learning to use the TensorFlow Recommenders library and the subclassing API 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?

run model - not sure how to compile, fit and run model and inspect output?

I am learning to use the TensorFlow Recommenders library and the subclassing API 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?

run model - not sure how to compile, fit and run model and inspect output?

• python
• tensorflow
• machine-learning
• keras

Share Improve this question Follow edited yesterday Vy Do 52.9k6969 gold badges255255 silver badges387387 bronze badges asked yesterday AlexEastAlexEast 2366 bronze badges 7

• 1 I believe your confusion stems from the fact that there is nothing to compile or to fit. Generate a valid input and use the call method. – Frede Commented yesterday
• Thank you @Frede for your help. I am now trying to do this but I can't get the data into the correct format, please could you advise me as to how to do this? – AlexEast Commented yesterday
• 1 Have you tried transforming each value in your dict independently into a tensor with tf.constant()? You don't have to (and shouldn't) transform the dict as a whole into a tensor. – Frede Commented yesterday
• 1 I scaled the model back to just include the 'user_id_embedding' section and then passed in a tf.constant() version of the 'user_id' value and it worked fine and I got back the expected embeddings. I just can't figure out how to get all of the different datatypes into one 'input' format. The fuller version of the model uses the cached dataset and works fine, and I have tried to inspect this an understand how it works, but I've not found the answer yet. Thank you for your help. – AlexEast Commented yesterday
• 1 Take a look into the documentation. You have to specify the type when using tf.constant(). Btw. ChatGPT is a nice tool, try it. – Frede Commented yesterday

 |  Show 2 more comments

1 Answer 1

Sorted by: Reset to default 0

I have solved the issue and include the answer here to help others with the same issue.

I have created a dictionary and used tf.constant() to create a Tensor for an example value, and specified the data type and I now get the expected output.

inputs = {
    "user_id": tf.constant(["138"], dtype=tf.string),
    "timestamp": tf.constant([879024327], dtype=tf.int64),
    "raw_user_age": tf.constant([46.0], dtype=tf.float32)
}


model = UserFeaturesModel()

output = model(inputs)
print(output)

本文标签： pythonHow to compile and run a model in Keras subclassing APIStack Overflow