Project for the Lecture "Deep Learning" held at Hasso-Plattner-Institute that is about applying the concept of Class Activation Mapping to the CIFAR10 dataset which contains 60,000 32x32 images of 10 different categories.
In the following preview an overview of the influence of centercropping the input images before training is given as well as a first grasp of the impact of using pretrained models on the quality of the Class Activation Maps.
| Original Image | Alexnet (no Centercropping) | Alexnet (Centercropping) | VGG19 (no Centercropping) | VGG19 (Centercropping) |
|---|---|---|---|---|
 |
 |
 |
 |
 |
 |
 |
 |
 |
 |
| Original Image | Alexnet (not pretrained) | Alexnet (pretrained) | VGG19 (not pretrained) | VGG19 (pretrained) |
|---|---|---|---|---|
 |
 |
 |
 |
 |
 |
 |
 |
 |
 |
| File | Content |
|---|---|
data_loading.py |
Contains all functions that get the CIFAR10 dataset and preprocess it |
util.py |
Contains util functions that do not belong to a specific category |
config.py |
Contains parameters and the corresponding Getter Functions |
architecture.py |
This file contains the Class(es) representing the network structure of the neural network |
train.py |
Contains the main training function and the basic setup of the model |
test.py |
This file contains model evaluations functions AND all functionality regarding Class Activation mapping |
- Currently to get a random image plus its CAM execute the following command:
python test.py [-h] model outputmodel = {vgg19, alexnet}
output = output file name
- work in Progress