Transfer Learning: Fine-tuning Networks

Transfer learning: ability to use a pre-trained model as a "shortcut" to learn patterns from data it was not originally trained on.

There are two types of transfer learning in the context of deep learning:

1. Transfer learning via feature extraction
2. Transfer learning via fine-tuning: modify the architecture of a network so we can re-train parts of the network

This repo will focus on the second method of transfer learning

Datasets

Flowers-17

Pictures of 17 category flower dataset with 80 images for each class. Dataset was collected by the Visual Geometry Group at the University of Oxford.

Technologies

• keras
• tensorflow
• scikit-learn

Deployment with Docker

1. Build the docker image:

$ cd docker
$ make build

2. Create a docker container based on the image:

$ make run

3. SSH to the docker container:

$ make dev

4. Show NN architecture (indexes and layers):

$ python src/inspect_model.py
[INFO] showing layers...
[INFO] 0	InputLayer
[INFO] 1	Conv2D
[INFO] 2	Conv2D
[INFO] 3	MaxPooling2D
[INFO] 4	Conv2D
[INFO] 5	Conv2D
[INFO] 6	MaxPooling2D
[INFO] 7	Conv2D
[INFO] 8	Conv2D
[INFO] 9	Conv2D
[INFO] 10	MaxPooling2D
[INFO] 11	Conv2D
[INFO] 12	Conv2D
[INFO] 13	Conv2D
[INFO] 14	MaxPooling2D
[INFO] 15	Conv2D
[INFO] 16	Conv2D
[INFO] 17	Conv2D
[INFO] 18	MaxPooling2D
[INFO] 19	Flatten
[INFO] 20	Dense
[INFO] 21	Dense
[INFO] 22	Dense

$ python src/inspect_model.py --include-top 0
[INFO] showing layers...
[INFO] 0	InputLayer
[INFO] 1	Conv2D
[INFO] 2	Conv2D
[INFO] 3	MaxPooling2D
[INFO] 4	Conv2D
[INFO] 5	Conv2D
[INFO] 6	MaxPooling2D
[INFO] 7	Conv2D
[INFO] 8	Conv2D
[INFO] 9	Conv2D
[INFO] 10	MaxPooling2D
[INFO] 11	Conv2D
[INFO] 12	Conv2D
[INFO] 13	Conv2D
[INFO] 14	MaxPooling2D
[INFO] 15	Conv2D
[INFO] 16	Conv2D
[INFO] 17	Conv2D
[INFO] 18	MaxPooling2D

5. Fine-tuning:

...

Resources

• Deep Learning for Computer Vision with Python by Dr. Adrian Rosebrock: https://www.pyimagesearch.com/deep-learning-computer-vision-python-book/