By default this script will run 4,000 training steps. Each step chooses ten images at random from the training set, finds their bottlenecks from the cache, and feeds them into the final layer to get predictions. Those predictions are then compared against the actual labels to update the final layer's weights through the back-propagation process.
According to this paragraph from this article from tensorflow , once the bottlenecks features are calculated for newer classes, backpropagation algorithm is used to retrain the last layer given the output labels. My question is, are the bottleneck features of the 10,000 classes of imagenet present while this retraining progresses - or the retraining is done for the newer class labels only.