Mapping strings to ints during preprocessing stage

TensorFlow General Discussion
, ,
1

Hi everyone!

I’m looking to train several models - image classification - on a large dataset which I want to represent as TFRecord files. Today I know how to do it if I save the image (tf.string) and class number (int64). I’d like to be able to set the class numbers during preprocess from a saved string.

Motivation - the dataset I’m working on contains images which can be classified as a more general category (i.e. dogs, cats) and more specifically (labrador, bulldog, persian cat). The labels are mostly more specific, so an image can go though a cascade of a more general category (model of cat/dog/negative), and then a more nuanced model (labrador/bulldog/neg).

I’m relatively new to the framework, and I was able to express what I want in eager execution mode using .numpy() method (code below).

My question is how can this mapping be done efficiently? what methods should I look for?

Thanks :slight_smile:

import tensorflow as tf
import functools

# example: this model classifies dogs and cats
example_label_map = {
    "bulldog": "dog",
    "labrador": "dog",
    "persian cat": "cat",
    "cow": "neg",
}

# map the N classes to an int of 0..N-1
labels_to_class_int = {
    "neg": 0,   # negative class
    "dog": 1,
    "cat": 2,
}

def parse_tfrec_function(example, labels_map):
    """
    labels is a map of labels to class numbers. e.g. "bulldog" -> 1, "labrador" -> 1, "persian cat" -> 2, ...
    """
    image_feature_description = {
        "image": tf.io.FixedLenFeature([], tf.string),
        "label": tf.io.FixedLenFeature([], tf.string)
    }

    features = tf.io.parse_single_example(example, image_feature_description)

    image = tf.io.parse_tensor(features["image"], tf.uint8)
    label = features["label"].numpy().decode('utf-8')
    class_num = labels_map[label]

    return image, class_num


tfrec_file = ['/srv/mickey/build-cls/test/data/dataset.tfrec']
dataset = tf.data.TFRecordDataset(tfrec_file)

labels_map = {}
for k, v in example_label_map.items():
    labels_map[k] = labels_to_class_int[v]

print(labels_map)

parser_fn = functools.partial(parse_tfrec_function, labels_map=labels_map)
parset_dataset = dataset.map(parser_fn) # this breaks due to AttributeError for Tensor object
3

You can use function tf.keras.utils.image_dataset_from_directory() to read images from directory and map the labels as integer value.

main_directory/
...cat/
......cat_image_1.jpg
......cat_image_2.jpg
...dog/
......dog_image_1.jpg
......dog_image_2.jpg
...neg/
......neg_image_1.jpg
......neg_image_2.jpg

Calling image_dataset_from_directory(main_directory, labels='inferred') will return a tf.data.Dataset that yields batches of images from the subdirectories cat and dog and neg , together with labels 0 ,1 and 2 (0 corresponding to cat and 1 corresponding to dog, 2 corresponding to neg ).

Specify the label_mode as int to get int value.

label_mode

- 'int': means that the labels are encoded as integers(e.g. for `sparse_categorical_crossentropy` loss).

For more information on the library please refer here