In this repository You'll find a machine learning task that consists in comparing three strategies for training CNNs on a image classification problem, with two of those strategies using Transfer Learning.
Transfer learning or inductive transfer is a research problem in machine learning that focuses on storing knowledge gained while solving one problem and applying it to a different but related problem.
Here, we use the the CIFAR-10 dataset, which consists of 60000 32x32 colour images in 10 classes, with 6000 images per class. There are 50000 training images and 10000 test images.
First, we split the dataset into two groups A and B. the latter contains classes 0 and 7 and the former contains the remaining classes. Then, we train a CNN from scratch using only the A dataset.
Then, we test three strategies to train a model using the B dataset:
- No Transfer Learning: Train the network from scratch, initialising the weights randomly.
- Fine-tuning in one layer: Use the weights of the network trained for part A and fine-tuning in the last Fully Connected layer.
- Fine-tuning in two layers: Use the weights of the network trained for part A and fine-tuning in both Fully Connected layers.
In this repo, You'll find both a jupyter notebook and a regular python scipt. You can choose which to use, given your application. However, the jupyter notebook contains analisys regarding the methods used, along with code and graphs.
Here are some examples of graphs created using code in this repository
For every model created, we can compare accuracy and loss for train/test:
Also, for every model, a Acc/Loss comparison graph is created:
We also plot graphs comparing all models. For train/test, we plot accuracy and loss for all models, as examplified below:
