Hello all, I would really appreciate some assistance on how to modify the retrain.py to retrain from deeper within the model than just the final layer. As far as I understand, as it stands, the retrain.py script scraps the final layer and retrains a new one in its place and then adds a softmax to the end with the prerequisite number of classes to fit your purpose. Now my question is - how can one go deeper than that and retrain from, say, the middle of the architecture in order to be able to recognize more features. Say I want to add some dense layers like so:
x=base_model.outputx=GlobalAveragePooling2D()(x)x=Dense(1024,activation='relu')(x) #we add dense layers so that the model can learn more complex functions and classify for better results.x=Dense(1024,activation='relu')(x) #dense layer 2x=Dense(512,activation='relu')(x) #dense layer 3preds=Dense(17,activation='softmax')(x) #final layer with softmax activation
and freeze all the top layers like so:
# Freeze pretrained weightsfor layer in model.layers:layer.trainable=False
I've managed to narrow down that the function that needs to be augmented to account for this is add_final_retrain_ops , but how exactly one goes about incorporating the changes, I have no idea. Does someone have experience with this?
Hello all, I would really appreciate some assistance on how to modify the
retrain.pyto retrain from deeper within the model than just the final layer. As far as I understand, as it stands, theretrain.pyscript scraps the final layer and retrains a new one in its place and then adds a softmax to the end with the prerequisite number of classes to fit your purpose. Now my question is - how can one go deeper than that and retrain from, say, the middle of the architecture in order to be able to recognize more features. Say I want to add some dense layers like so:x=base_model.outputx=GlobalAveragePooling2D()(x)x=Dense(1024,activation='relu')(x) #we add dense layers so that the model can learn more complex functions and classify for better results.x=Dense(1024,activation='relu')(x) #dense layer 2x=Dense(512,activation='relu')(x) #dense layer 3preds=Dense(17,activation='softmax')(x) #final layer with softmax activationand freeze all the top layers like so:
# Freeze pretrained weightsfor layer in model.layers:layer.trainable=FalseI've managed to narrow down that the function that needs to be augmented to account for this is
add_final_retrain_ops, but how exactly one goes about incorporating the changes, I have no idea. Does someone have experience with this?