Mastering TensorFlow 1.x

Image classification using retrained VGG16 in Keras

Let us use the COCO images dataset to retrain the model to fine-tune the classification task. We shall remove the last layer in the Keras model and add our own fully connected layer with softmax activation for 8 classes. We shall also demonstrate freezing the first few layers by setting the trainable attribute of the first 15 layers to False.

1. Start by importing the VGG16 model without the variables for the top layer, by setting the include_top to False:

# load the vgg modelfrom keras.applications import VGG16base_model=VGG16(weights='imagenet',include_top=False, input_shape=(224,224,3))

We also specify the input_shape in the above code otherwise Keras throws exceptions later.

2. Now we build ...