Project - Training from Scratch vs Transfer Learning

7 / 8

Creating new model based on existing model A

  • Let us first see how many trainable parameters are there for model_B we trained previously.

  • Then we shall create a new model model_B_on_A which has the pre-trained parameters of model_A but customized final dense layer with only 1 neuron.

  • Finally, we shall compare the performance of both the models - model_B and model_B_on_A.

  • See the model_B summary using summary() on model_B.

    model_B.<<your code comes here >>

    We see that there are 275,801 trainable parameters for model_B.

  • Now, before creating model_B_on_A(a model based on pre-trained layers of model_A), we shall clone the model_A and set its trained weights so that when you train model_B_on_A, it will not affect model_A.

    We could copy the model_A architechture using keras.models.clone_model.

    • Create model_A_clone which is the copy of model_A.

      model_A_clone = keras.models.clone_model(model_A)
    • Get the weights of model_A using get_weights(), and set the model parameters for model_A_clone using set_weights().

      model_A_clone.<< your code comes here >>(model_A.get_weights())
  • Now, create a new model model_B_on_A, based on existing layers of model_A.

    << your code comes here >> = keras.models.Sequential(model_A.layers[:-1])
  • Add the final dense layer with 1 neuron to the model_B_on_A. Set the activation to "sigmoid", as this is a binary classification problem.

    model_B_on_A.add(keras.layers.Dense(1, activation=<< your code comes here >>))
  • Set all the layers, except the last layer, of model_B_on_A to be non-trainable.

    for layer in model_B_on_A.layers[:-1]:
        layer.trainable = False
  • Now check the number of trainable parameters of model_B_on_A.


    We observe there are only 51 parameters to train in model_B_on_A, while there are as many as 275,801 trainable parameters for model_B.

  • Compile the model model_B_on_A by using model.compile.

    • Set loss="binary_crossentropy".

    • Set optimizer=keras.optimizers.SGD(lr=1e-3)

      model_B_on_A.compile(loss=<< your code comes here >>,
               optimizer=<< your code comes here >>,
  • Now train the model_B_on_A uaing

    history =, y_train_B, epochs=5,
                       validation_data=(X_valid_B, y_valid_B))
Get Hint See Answer

Note - Having trouble with the assessment engine? Follow the steps listed here

Loading comments...