Feature Extraction

TensorFlow Hub provides the option of downloading a model without the final layer. In this case, you will be using a TensorFlow module as a feature extractor; you can design your custom final layers on top of it. In TensorFlow Hub, a module used for feature extraction is known as a feature vector:

import tensorflow_hub as hub
MODULE_HANDLE = 'https://tfhub.dev/google/efficientnet/b0'\
                '/feature-vector/1'
module = hub.load(MODULE_HANDLE)

Note

To find all the available feature vectors on TensorFlow Hub, you can use its search engine: https://tfhub.dev/s?module-type=image-feature-vector&tf-version=tf2.

Once loaded, you can add your own final layer to the feature vector with the Sequential API:

model = tf.keras.Sequential([
    hub.KerasLayer(MODULE_HANDLE, input_shape=(224, 224, 3)),
    tf.keras...