3

Is it possible to use pre-trained model features from VGG-16 and pass to GlobalAveragePooling2D() layer of other model in Keras?

Sample code for storing offline features of VGG-16 network:

model = applications.VGG16(include_top=False, weights='imagenet')
bottleneck_features_train = model.predict(input)

Sample code for top model:

model = Sequential()
model.add(GlobalAveragePooling2D()) # Here I want to use pre-trained feature from VGG-16 net as input.

I can not use Flatten() layer as I want to predict multi-labels with multi-classes.

stop-cran
  • 4,229
  • 2
  • 30
  • 47
MUNEER AHMAD
  • 43
  • 7

1 Answers1

3

Sure, you definitely can. You've got a couple of options:

pooling kwarg

Use the pooling kwarg in the VGG16 constructor, which replaces the last pooling layer with the specified type. i.e.

model_base = keras.applications.vgg16.VGG16(include_top=False, input_shape=(*IMG_SIZE, 3), weights='imagenet', pooling="avg")

Adding layers to the output

You can also add more layers to the pretrained model: 

from keras.models import Model

model_base = keras.applications.vgg16.VGG16(include_top=False, input_shape=(*IMG_SIZE, 3), weights='imagenet')
output = model_base.output
output = GlobalAveragePooling2D()(output)
# Add any other layers you want to `output` here...
model = Model(model_base.input, output)
for layer in model_base.layers:
    layer.trainable = False

That last line freezes the pretrained layers so that you preserve the features of the pretrained model and just train the new layers.

I wrote a blog post that goes through the basics of working with pretrained models and extending them to work on various image classification problems; it's also got a link to some working code examples that might provide more context: http://innolitics.com/10x/pretrained-models-with-keras/

Reece Stevens
  • 444
  • 3
  • 7