Train with transfer learning EfficientNetV2 TFHub model

General Discussion
#1

I’m looking for a working Python sample that given an image dataset with a class for each folder it trains a custom model using transfer learning on EfficientNetV2XL downloaded form TfHub and loaded with

model = tf.saved_model.load(path)

Thanks.

#3

here you go: Retraining an Image Classifier  |  TensorFlow Hub

I think this can give you a very good start

#4

Thanks for the answer but the problem is that model obtained from tfhub downloaded archive with

model = tf.saved_model.load(extraction_path)

is a different object than model obtained with

model = tf.keras.Sequential([
        tf.keras.layers.InputLayer(input_shape=IMAGE_SIZE + (3,)),
        hub.KerasLayer(model_handle, trainable=do_fine_tuning),
        tf.keras.layers.Dropout(rate=0.2),
        tf.keras.layers.Dense(len(class_names), kernel_regularizer=tf.keras.regularizers.l2(0.0001))
    ])
    model.build((None,)+IMAGE_SIZE+(3,))

as in the example of the link (and other examples I have seen)

and I cannot call e.g. methods as model.fit or model.predict on model = tf.saved_model.load(extraction_path)

In what way I should proceed if I want to use the saved models?

#6

Sorry, it’s not clear what you want to do.

A model from tfhub, ideally shoud be used in one of two way:

  1. hub.load(model_path)
  2. hub.KerasLayer(model_path)

for 1, you will get a model that can do inference directly (hub.load  |  TensorFlow Hub). This is equivalent to tf.saved_model.load and the result is a TensorFlow 2 low-level module. This is not a Keras object so it doesn’t have the fit and predict methods

For 2, you will get a layer that can be used to compose a model. You can still use it for inference, it will work but it’s not optmal. (hub.KerasLayer  |  TensorFlow Hub). The returned object is wrapped such that it can be used as a Keras layer.

so 1 and 2 are the same model but presented in a different way

For what you want to do, you can follow the colab I shared, fine tune your model, save it and then load it anyway you want and just run inference on it
does it makes sense?

in both cases, what you have is a