import tensorflow as tf
COLUMNS = ("user_id,item_id,rating,timestamp,datetime,year,month,day,week,dayofweek,"
"age,gender,occupation,zipcode,zipcode1,zipcode2,zipcode3,"
"title,release,video_release,imdb,unknown,action,adventure,animation,children,"
"comedy,crime,documentary,drama,fantasy,filmnoir,horror,musical,mystery,romance,"
"scifi,thriller,war,western,release_date,release_year").split(",")
GENRE = ("unknown,action,adventure,animation,children,comedy,crime,documentary,drama,fantasy,"
"filmnoir,horror,musical,mystery,romance,scifi,thriller,war,western").split(",")
LABEL_COL = "rating"
DEFAULTS = [[0], [0], [0], [0], ["null"], [0], [0], [0], [0], [0],
[0], ["null"], ["null"], ["null"], ["null"], ["null"], ["null"],
["null"], ["null"], ["null"], ["null"], [0], [0], [0], [0], [0],
[0], [0], [0], [0], [0], [0], [0], [0], [0], [0],
[0], [0], [0], [0], ["null"], [0]]
def get_feature_columns(embedding_size=4):
user_fc = tf.feature_column.categorical_column_with_hash_bucket("user_id", 1000, tf.int32)
item_fc = tf.feature_column.categorical_column_with_hash_bucket("item_id", 2000, tf.int32)
# user features
age_fc = tf.feature_column.numeric_column("age")
age_buckets = tf.feature_column.bucketized_column(age_fc, list(range(15, 66, 10)))
gender_fc = tf.feature_column.categorical_column_with_vocabulary_list(
"gender", ["F", "M"],
num_oov_buckets=1
)
occupation_fc = tf.feature_column.categorical_column_with_hash_bucket("occupation", 50)
zipcode_fc = tf.feature_column.categorical_column_with_hash_bucket("zipcode", 1000)
# item features
release_year_fc = tf.feature_column.numeric_column("release_year")
release_year_buckets = tf.feature_column.bucketized_column(release_year_fc, list(range(1930, 1991, 10)))
genre_fc = [tf.feature_column.categorical_column_with_identity(col, 2) for col in GENRE]
linear_columns = [user_fc, item_fc, age_buckets, gender_fc, occupation_fc, zipcode_fc, release_year_buckets] + genre_fc
deep_columns = [tf.feature_column.embedding_column(fc, embedding_size) for fc in linear_columns]
return {"linear": linear_columns, "deep": deep_columns}
def get_input_fn(csv_path, mode=tf.estimator.ModeKeys.TRAIN, batch_size=32, cutoff=5):
def input_fn():
def parse_csv(value):
columns = tf.decode_csv(value, DEFAULTS)
features = dict(zip(COLUMNS, columns))
label = features.pop(LABEL_COL)
label = tf.math.greater_equal(label, cutoff)
return features, label
# read, parse, shuffle and batch dataset
dataset = tf.data.TextLineDataset(csv_path).skip(1) # skip header
if mode == tf.estimator.ModeKeys.TRAIN:
# shuffle and repeat
dataset = dataset.shuffle(16 * batch_size).repeat()
dataset = dataset.map(parse_csv, num_parallel_calls=8)
dataset = dataset.batch(batch_size)
return dataset
return input_fn
def serving_input_fn():
feature_placeholders = {
"user_id": tf.placeholder(tf.int32, [None]),
"item_id": tf.placeholder(tf.int32, [None]),
"age": tf.placeholder(tf.int32, [None]),
"gender": tf.placeholder(tf.string, [None]),
"occupation": tf.placeholder(tf.string, [None]),
"zipcode": tf.placeholder(tf.string, [None]),
"release_year": tf.placeholder(tf.int32, [None]),
}
feature_placeholders.update({
col: tf.placeholder_with_default(tf.constant([0]), [None]) for col in GENRE
})
features = {
key: tf.expand_dims(tensor, -1)
for key, tensor in feature_placeholders.items()
}
return tf.estimator.export.ServingInputReceiver(
features=features,
receiver_tensors=feature_placeholders
)
