In the previous section on gender classification using CNN, we saw that when we build a CNN model from scratch, we could encounter some of the following scenarios:
- The number of images that were passed is not sufficient for the model to learn
- Convolutions might not be learning all the features in our images when the images are big in size
The first problem could be tackled by performing our analysis on a large dataset. The second one could be tackled by training a larger network on the larger dataset for a longer number of epochs.
However, while we are able to perform all of this, more often than not, we do not have the amount of data that is needed to perform such an analysis. Transfer learning using pre-trained models comes to the rescue in such scenarios.
ImageNet is a popular competition...
