How can I preprocess my Mapdataset to fit my model input? - tensorflow

Question

I use a MapDataset compose of label in text and a vector of float in string.
Here is the way I read the content of my tfrecord:
def extract_data(tfrecord_ds):
feature_description = {
'classes_text': tf.io.FixedLenFeature((), tf.string),
'data': tf.io.FixedLenFeature([], tf.string)
}
def _parse_data_function(example_proto):
return tf.compat.v1.parse_single_example(example_proto, feature_description)
parsed_dataset = tfrecord_ds.map(_parse_data_function)
dataset = parsed_dataset.cache().shuffle(1000).batch(32).prefetch(tf.data.AUTOTUNE)
return dataset
I want to convert the label_text to int according to label.txt file and the data string to vector of float.
I want to use this data to train a custom model like this:
my_model = tf.keras.Sequential([
tf.keras.layers.Input(shape=(1024), dtype=tf.float32,
name='input_embedding'),
tf.keras.layers.Dense(512, activation='relu'),
tf.keras.layers.Dense(num_classes)
], name='audio_detector')
How can I process my MapDataset from (string,string) to (int, float_array) to be able to train my model?
Edit:
Here is the way I encode my data:
features = {}
features['classes_text'] = tf.train.Feature(
bytes_list=tf.train.BytesList(value=[audio_data_generator.label.encode()]))
bytes = embedding.numpy().tobytes()
features['data'] = tf.train.Feature(bytes_list=tf.train.BytesList(value=[bytes]))
tf_example = tf.train.Example(features=tf.train.Features(feature=features))
writer.write(tf_example.SerializeToString())

Answer 1

It is easier to encode the embedding using tf.train.FloatList.
When writing to tfrecords use:
features = {
'classes_text': tf.train.Feature(bytes_list=tf.train.BytesList(value=[label.encode()])),
'data': tf.train.Feature(float_list=tf.train.FloatList(value=embedding))
}
tf_example = tf.train.Example(features=tf.train.Features(feature=features))
writer.write(tf_example.SerializeToString())
And when reading give the embedding size to tf.io.FixedLenFeature, for example:
embedding_size = 10
feature_description = {
'classes_text': tf.io.FixedLenFeature((), tf.string),
'data': tf.io.FixedLenFeature([embedding_size], tf.float32)
}
To convert label_text to int you can use tf.lookup.StaticVocabularyTable.
# Assuming lable.txt contains a single label per line.
with open('label.txt', 'r') as fin:
categories = [line.strip() for line in fin.readlines()]
init = tf.lookup.KeyValueTensorInitializer(
keys=tf.constant(categories),
values=tf.constant(list(range(len(categories))), dtype=tf.int64))
label_table = tf.lookup.StaticVocabularyTable(
init,
num_oov_buckets=1)
feature_description = {
'classes_text': tf.io.FixedLenFeature((), tf.string),
'data': tf.io.FixedLenFeature([embedding_size], tf.float32)
}
def _parse_data_function(example_proto):
example = tf.compat.v1.parse_single_example(example_proto, feature_description)
# Apply the label lookup.
example['classes_text'] = label_table.lookup(example['classes_text'])
return example
parsed_dataset = tfrecord_ds.map(_parse_data_function)
dataset = parsed_dataset.cache().shuffle(1000).batch(32).prefetch(tf.data.AUTOTUNE)
Edit
If you wish to keep the way you save data you can use np.frombuffer to convert the numpy vectors to from binary stings. You will have to wrap this code in a tf.function and tf.py_function though.
def decode_embedding(embedding_bytes):
return np.frombuffer(embedding_bytes.numpy())
#tf.function()
def tf_decode_embedding(embedding_bytes):
return tf.py_function(decode_embedding, inp=[embedding_bytes], Tout=tf.float32)
feature_description = {
'classes_text': tf.io.FixedLenFeature((), tf.string),
'data': tf.io.FixedLenFeature([], tf.string)
}
def _parse_data_function(example_proto):
example = tf.compat.v1.parse_single_example(example_proto, feature_description)
example['classes_text'] = label_table.lookup(example['classes_text'])
example['data'] = tf_decode_embedding(example['data'])
return example
parsed_dataset = tfrecord_ds.map(_parse_data_function)
dataset = parsed_dataset.cache().shuffle(1000).batch(32).prefetch(tf.data.AUTOTUNE)