If the target dataset is large and different from the original:
- Remove the last fully connected layer and replace it with a fully connected layer that matches the number of classes in the target dataset
- Train the entire network from scratch with randomly initialized weights:
The Caffe library has ModelZoo, where one can share network weights.
Consider training from scratch when the dataset is large and completely different from the original dataset. In this case, we have enough data to train from scratch without the fear of overfitting. However, even in this case, it ...
