I'm trying to apply transfer learning using VGG16 but I'm struggling to find an example with multiple outputs.
I've written my own architecture as below which I used to manually train my model - though I don't understand how I can compile the same metrics using a pre-trained model.
How do I do this?
from keras.applications.vgg16 import VGG16
vgg16_model = VGG16(weights='imagenet',
include_top=False,
input_shape=(224,224,3))
vgg16_model.summary()
def modelG():
input=tf.keras.layers.Input(shape=(224,224,3))
x=input
x=layers.Conv2D(8,(3,3),activation='relu')(x)
x=layers.Conv2D(8,(3,3),activation='relu')(x)
x=layers.MaxPooling2D(2)(x)
x=layers.Dropout(0.1)(x)
x=layers.Conv2D(16,(3,3),activation='relu')(x)
x=layers.Conv2D(16,(3,3),activation='relu')(x)
x=layers.MaxPooling2D(2)(x)
x=layers.Dropout(0.1)(x)
x=layers.Conv2D(32,(3,3),activation='relu')(x)
x=layers.Conv2D(32,(3,3),activation='relu')(x)
x=layers.MaxPooling2D(2)(x)
x=layers.Dropout(0.1)(x)
x=layers.Conv2D(84,(3,3),activation='relu')(x)
x=layers.Dropout(0.1)(x)
x=layers.Flatten()(x)
out_col=layers.Dense(512,activation='relu')(x)
out_ren=layers.Dense(512,activation='relu')(x)
out_col= layers.Dense(1,activation='sigmoid',name='col_out')(out_col)
out_ren=layers.Dense(1,activation='relu',name='ren_out')(out_ren)
multiOutputModel=tf.keras.models.Model(inputs=input, outputs=[out_col, out_ren])
# Compile
multiOutputModel.compile(
optimizer='adam',
loss={
'ren_out': 'mean_squared_error',
'col_out': 'binary_crossentropy'},
loss_weights={
'ren_out': 4.0,
'col_out': 0.1},
metrics={
'ren_out': 'mean_absolute_error',
'col_out': 'accuracy'})
tf.keras.utils.plot_model(modelB, 'modelB.png',show_shapes=True)
return multiOutputModel
multiOutputModel.summary()
CodePudding user response:
you would do it as you would for the single output case
base_model=tf.keras.applications.VGG19(include_top=False, weights="imagenet",input_shape=img_shape, pooling='max')
x=base_model.output
out_col=layers.Dense(512,activation='relu')(x)
out_ren=layers.Dense(512,activation='relu')(x)
out_col= layers.Dense(1,activation='sigmoid',name='col_out')(out_col)
out_ren=layers.Dense(1,activation='relu',name='ren_out')(out_ren)
multiOutputModel=tf.keras.models.Model(inputs=base_model.input, outputs=[out_col, out_ren])
Note in the VGG model I set pooling='max' so the output of VGG is a 1 dimentionsal tensor so you do not need a flatten layer.