How to customize a pre-trained model with our own dataset?

Question

I am currently working on Object detection . I am using Amazon Workspace for training my model. I am using the model to detect cars and bikes . I have used a pre-trained model (which is Faster-RCNN-Inception-V2 model) and customized it with my own dataset for 2 labels namely car and bike. It took me 5 hours to complete the training. Now I want to modify my model for 2 more labels (keeping to old ones) namely bus and auto. But I don't want to do the training from scratch as my model is already trained for cars and bikes. So is there any way that I can train my model only with the dataset of bus and auto, and after training it will detect all 4 objects(car, bike, bus, and auto) ?

Answer 1

Load the saved model with pretrained weights. Remove last dense and 2 class softmax layer and add a new dense with 4 class softmax as your low level features are already trained. Now train this model with modified data.