Again, we are going to replace the final fully connected layer of the pre-trained inception model and then use the rest of the inception model as a feature extractor. So, we first feed our raw images in the inception model, which will extract the features from them and then output our so-called transfer values.
After getting the transfer values of the extracted features from the inception model, you might need to save them to your desk because it will take some time if you did it on the fly, so it's useful to persist them to your desk to save you time. In TensorFlow tutorials, they use the term bottleneck values instead of transfer values, but it's just a different name for the exact same thing.
After getting the transfer values or loading them from the desk, we can feed them to any linear classifier that's customized to our new task. Here, we will feed the extracted transfer values to another neural network and then train for the new classes of CIFAR-10.
The following diagram, shows the general solution outline that we will be following: