Fine-Tuning

Previously, you used transfer learning to leverage pre-trained models on your own dataset. You used the weights of state-of-the-art models that have been trained on large datasets such as ImageNet. These models learned the relevant parameters to recognize different patterns from images and helped you to achieve amazing results on different datasets.

But there is a catch with this approach. Transfer learning works well in general if the classes you are trying to predict belong to the same list as that of ImageNet. If this is the case, the weight learned from ImageNet will also be relevant to your dataset. For example, the cats and dogs classes from the preceding exercise are present in ImageNet, so its weights will also be relevant for this dataset.

However, if your dataset is very different from ImageNet, then the weights from these pre-trained models may not all be relevant. For example, if your dataset contains satellite images, and you are trying to determine whether...