Repo for the project of NN 2020 based on the work Interpreting CNNs via Decision Trees.
Convolutional neural networks are nowadays widely used for different tasks in many fields, becomes thus important to understand what knowledge a CNN learns. In order to do so we have modified a VGG-16 network by adding a special mask layer at the end of the convolution, and trained it for an image classification task. Finally we have built a decision tree that could help us explain which object parts contributed the most to the final prediction and quantify these contributions. (Full report is available here)
- CNN: VGG16
- Benchmark dataset: PASCAL VOC 2010 Part Dataset + CUB200-2011
Virtual environment set up:
python3 -m venv --system-site-packages ./my_venv
# to enter the virtual env
source ./my_venv/bin/activate
# to exit the virtual env
deactivate
We developed this project with tensorflow for python 3.
- additional packages are indicated in the
requirements.txtfile, to install them using pip:
pip install --upgrade pip
pip install -r requirments.txt
- we exploited treelib as tree implementation.
- Giammarco D'Alessandro
- Maria Rosaria Fraraccio
- Luca Gioffrè (username changed from
balthier)