I want to use a pre-trained model with a dataset that the model hasn’t been trained on. How can this approach be effective, and is it possible to apply?
That can work, but you’ll typically need to do additional training using your specific data, once you have added layers to customize the pretrained model to your problem. Of course it requires that your training data be of the same format, or at least convertible to the same format, as the original training data that the base model was built for.
There are various points in the courses and specializations here on DeepLearning.AI where Transfer Learning is discussed. The course that I’m familiar with that covers this topic is DLS Course 4, Week 2. If this is your first time attempting to use the Transfer Learning strategy, my suggestion would be to spend some time taking DLS Course 4 at least through Week 2 to get introduced to how it works and how to apply it in general. You may then need to do more research if the methods discussed in the course are not sufficiently specific to your situation.
Using a pre-trained model with a new dataset is an effective approach known as transfer learning. The pre-trained model has learned from a large dataset and captured meaningful features. By fine-tuning this model on a new dataset, we adapt its learned features to the specific task at hand. This saves time and computational resources compared to training from scratch. We choose a suitable pre-trained model, adjust its weights through fine-tuning, and evaluate its performance on the new dataset to ensure effective knowledge transfer.
1 Like