Transfer Learning for Images Using PyTorch: Essential Training
58mIntermediate2019-10-08
Authors
Jonathan Fernandes
Consultant focusing on data science, AI, and big data
Course details
After its debut in 2017, PyTorch quickly became the tool of choice for many deep learning researchers. In this course, Jonathan Fernandes shows you how to leverage this popular machine learning framework for a similarly buzzworthy technique: transfer learning. Using a hands-on approach, Jonathan explains the basics of transfer learning, which enables you to leverage the pretrained parameters of an existing deep-learning model for other tasks. He then shows how to implement transfer learning for images using PyTorch, including how to create a fixed feature extractor and freeze neural network layers. Plus, find out about using learning rates and differential learning rates.
Learning objectives
What is transfer learning?
Using autograd
Creating a fixed feature extractor
Training an extractor
Fine-tuning the ConvNet
Learning rates and differential learning rates
Learning objectives
What is transfer learning?
Using autograd
Creating a fixed feature extractor
Training an extractor
Fine-tuning the ConvNet
Learning rates and differential learning rates
Skills covered
PyTorchMachine LearningPythonEssential TrainingArtificial Intelligence (AI)Open Source
Concepts
0. Introduction
- 01 - Welcome
- 02 - What you should know before watching this course
1. What Is Transfer Learning
- 03 - What is transfer learning
- 04 - VGG16
- 05 - CIFAR-10 dataset
2. Transfer Learning - Fixed Feature Extractor
- 06 - Creating a fixed feature extractor
- 07 - Understanding loss - CrossEntropyLoss() and NLLLoss()
- 08 - Autograd
- 09 - Using autograd
- 10 - Training the fixed feature extractor
- 11 - Optimizers
- 12 - CPU to GPU
- 13 - Train the extractor
- 14 - Evaluate the network and viewing images
- 15 - Viewing images and normalization
- 16 - Accuracy of the model
3. Fine-Tuning the ConvNet
- 17 - Fine-tuning
- 18 - Using fine-tuning
- 19 - Training from the fully connected network onwards
- 20 - Unfreezing and training over the last CNN block onwards
- 21 - Unfreezing and training over the last two CNN block onwards
4. Further Techniques
- 22 - Learning rates
- 23 - Differential learning rates
Conclusion
- 24 - Next steps