Different between Feature vector and classification model on TF-Hub

[teddcp2]

hi all,

I found on TFHUB that there are 2 model exists from same arch(Efficient B0) i.e Feature vector and Classification. What's the difference between them? @mrdbourke

My thought
Classification model is the original model that is designed to predict 1k classes where model based on feature vector only includes weights/patterns to accustom our need. Even so this seems to be correct, why are we setting include_top=false for base model of feature vector models? Is not it contradicting the my first assumption? If so, What's the difference between them?

[teddcp2]

Ok. I found the answer Which is dumb actually. There 2 ways we can use extrernal models. One through the TF-Hub and another is keras.applications. We need to set the include_top=False on Keras application layer, not in models based on tf-hub.

Differences in how to use it with Tf-hub and Keras

# Feature vector based model
m = tf.keras.Sequential([
    hub.KerasLayer("https://tfhub.dev/tensorflow/efficientnet/b0/feature-vector/1",
                   trainable=False),  # Can be True, see below.
    tf.keras.layers.Dense(num_classes, activation='softmax')
])

# Classification based model
m = tf.keras.Sequential([
    hub.KerasLayer("https://tfhub.dev/tensorflow/efficientnet/b0/classification/1")
])

# Keras application layer
tf.keras.applications.EfficientNetB0(
    include_top=True,
    weights="imagenet",
   ......
)

1- link for more on TF-hub

2- keras applications

thanks for reading...

[mrdbourke]

Thank you for providing the answer!

And you're right, tf.keras.applications vs TF-Hub can be a little confusing to begin with... as well as feature vector vs classification models.

[teddcp2]

@mrdbourke , What if I want to customize the number of unfrozen layers for fine-tuning. With keras, we can easily do number of trainable layers , but with TF Hub, entire layers will be trainable.. right? Is there any way to do so with TFHUB ?

hub.KerasLayer("https://tfhub.dev/google/imagenet/efficientnet_v2_imagenet1k_s/feature_vector/2", trainable=True)

Any way to set the number of layers with TFHub?

[mrdbourke]

I'm not aware of a specific way to do so with TFHUB models, however, this Stack Overflow thread might have something: https://stackoverflow.com/questions/61550788/is-there-a-way-to-freeze-specific-layers-in-a-keraslayer

Though looks like it might be only targeting the whole model 🤔

[teddcp2]

Yep.. same thought.. :)