将标签编码器实现为 TensorFlow 预处理层

Question

From my understanding from SKLearn's documentation , LabelEncoder in SKLearn encodes values between 0 and the number of classes subtracted by 1 (ie n_classes - 1).

I wanted to use something similar as a part of a Tensorflow preprocessing operation to avoid using SKLearn for a package. For example, I understand the preprocessing layer provides APIs for OneHot encoding and Categorical Encoding easily as follows:

tf.keras.layers.CategoryEncoding(
    num_tokens=None, output_mode='multi_hot', sparse=False, **kwargs
)

Is there any way to use LabelEncoder by certain arguments in the CategoryEncoding API, or do I have to define a brand new pre-processing layer using the abstract base class template provided in the Tensorflow documentations?

If so, is there any possible reference on how I can write my own class for using LabelEncoder as a Tensorflow layer?

Answer 1

IIUC, you just need sparse integer labels. So, maybe try something simple and naive first:

classes = ['fish1', 'fish2', 'fish3']

data = ['fish1', 'fish2', 'fish3', 'fish2', 'fish3', 'fish1']

class_indices = dict(zip(classes, range(len(classes))))
labels = list(map(class_indices.get, data))

print(labels) 

[0, 1, 2, 1, 2, 0]

Or with Tensorflow , you can use StaticHashTable :

import tensorflow as tf

classes = ['fish1', 'fish2', 'fish3']
data = tf.constant(['fish1', 'fish2', 'fish3', 'fish2', 'fish3', 'fish1'])

table = tf.lookup.StaticHashTable(
    tf.lookup.KeyValueTensorInitializer(tf.constant(classes), tf.range(len(classes))),
    default_value=-1)

label_encoder = tf.keras.layers.Lambda(lambda x: table.lookup(x))

print(label_encoder(data))

tf.Tensor([0 1 2 1 2 0], shape=(6,), dtype=int32)