This snippet was used to recover a proper keras model from a saved model which contained a submodel (ie., one of its layers was actually another model), in order to apply model optimization a posteriori (quantization, pruning). It can be extended to handle more type of layers.

trained_to_sequential.py

from tensorflow.keras.models import Sequential
import tensorflow.keras.layers as keras_layers


clone = Sequential()
# Here, layer[1] is a model. 
layers = [model.layers[0]] + model.layers[1].layers + model.layers[2:]

for layer in layers:
    if str(type(layer)).endswith("InputLayer'>"):
        clone.add(keras_layers.Input(shape=(192,192,3)))
    elif str(type(layer)).endswith("Conv2D'>"):
        l = layer
        clone.add(keras_layers.Conv2D(**layer.get_config()))
        clone.layers[-1].set_weights(layer.get_weights())
    elif str(type(layer)).endswith("MaxPooling2D'>"):
        l = layer
        clone.add(keras_layers.MaxPooling2D(**layer.get_config()))
        clone.layers[-1].set_weights(layer.get_weights())
    elif str(type(layer)).endswith("Dense'>"):
        l = layer
        clone.add(keras_layers.Dense(**layer.get_config()))
        clone.layers[-1].set_weights(layer.get_weights())
    elif str(type(layer)).endswith("Flatten'>"):
        l = layer
        clone.add(keras_layers.Flatten(**layer.get_config()))
        clone.layers[-1].set_weights(layer.get_weights())
    else:
        print(str(type(layer)))