Transfer Learning of the CIFAR-10 dataset using VGG16 with a SVM classifier.
- Python 3.6.4
- Please see requirement.txt for modules
- Shallow_classifier.ipynb
- Feature extractor: HoG
- Clasifier: SVM
- Dimensionality_reduction.ipynb
- t-SNE
- Model.ipynb
- Feature extractor: VGG16 bottleneck fc2
- Clasifier: SVM
- Presentation.ipynb
- Look here if you want to see the gist of the project without looking at a lot of code.
- download_data.py
- downloads data from https://www.cs.toronto.edu/~kriz/cifar.html and unpacks it
- extract_data.py
- Loads data form batches and saves into '.h5' files
- save_file function intakes 'proportion' argument which determines the size of the dataset. Use 0.1 if you do not have a CUDA enabled GPU, else use 1.
- features.py
- Taken from http://cs231n.github.io/assignments2016/assignment1/
- Used to extract the histogram of oriented gradients features of each image
- image_plotting.py
- Contains functions used to display images in the presentation jupyter notebook
- plot_cm.py
- plots confusion matrix
Presentation or here in this jupyter notebook.
10 random images form each class in the dataset.
- Code executed in this jupyter notebook.
- Test set accuracy = 51 %
| Accuracy | F1-score |
|---|---|
| 51 % | 51 % |
- I have chosen the second fully connected (FC2) layer to extract visual features.
- Code executed in this jupyter notebook.
- I have embedded them in 2D via the following dimensionality reduction techniques: PCA, t-SNE
- Code executed in this jupyter notebook.
- Test set accuracy (with bagging) = 80 %
| Accuracy | F1-score |
|---|---|
| 80 % | 80 % |