Table of contents

• 1. Introduction
• 2. What is transfer learning
• 3. Importing libraries and data
• 4. Utility functions
• 5. Create dataset
• 6. Creating Dataloaders
• 7. Creating metric
• 8. Transfer Learning
  • 8.1 Downloading model
  • 8.2 Freezing the model
  • 8.3 Creating final layer
• 9. Training
  • 9.1 Baseclass
  • 9.2 DataLoader
  • 9.3 Transfer learning model
  • 9.4 Training
• 10. Conclusion
• 11. Reference

1. Introduction

For us humans if we learn something, similar things become easy. This is because we can apply our existing knowledge to the new tasks. One example of this is learning to ride a bycycle. If we know know to ride a bycycle, we can easly learn how to ride a motor bike.

Similar to this there are many many CNN models availabe to us, which are pretrained in a tens of thousands of images. Instead of building everything from scratch, we can leverage the knowledge of these neural networks and modfy them according to our needs. This is called transfer learning.

2. What is transfer learning ?

Transfer learning is a machine learning technique where knowledge gained during training in one type of problem is used to train in other, similar types of problem.

Thus, instead of building your own deep neural networks, which can be a cumbersome task to say the least, you can find an existing neural network that accomplishes the same task you’re trying to solve and reuse the layers that are essential for pattern detection, while also making changes to the fully connected layer to suit your problem.