README.md

QC4RS

Quantum Classifiers for Remote Sensing

Figure 1: General scheme for hybrid systems

Framework for hybrid systems for the classification of satellite imagery. Pre- and postprocessing are performed classical while a parameterized quantum circuit is used for classification. The hybrid systems use 16 data qubits and the quantum systems are simulated on the GPU. Classifications were performed with EuroSAT and NWPU-RESISC45 data.

Figure 2: Preprocessing pipelines

Available preprocessing methods:

• VGG16 (without fully connected top layers for classification)
• Downscaling
• Principal component analysis
• Factor analysis
• Simple autoencoder
• Deep autoencoder
• Convolutional autoencoder
• Autoencoder created from restricted Boltzmann machines

Available quantum encodings:

• Basis encoding
• Angle encoding

Available quantum circuit architectures:

• FPQC
• GPQC

Available loss functions:

• Hinge loss
• Square hinge loss
• Binary cross-entropy loss

Available optimizers:

• Adam

Setup

Create a directory:

mkdir qc4rs_dir && cd qc4rs_dir

Clone the repository:

git clone https://github.com/tumbgd/qc4rs.git

Download and unzip a dataset, e.g. EuroSAT:

curl https://madm.dfki.de/files/sentinel/EuroSAT.zip -o EuroSAT.zip && unzip EuroSAT.zip

Build the container (The version in the keyring URL inside the Dockerfile is important, note this example is for Debian 10. You can find your version with the command 'lsb_release -a' in a workspace terminal):

cd qc4rs && docker build -t qc4rs .

Start the container:

docker run -it --rm -v /path/to/qc4rs_dir:/tf --gpus=all --name=qc4rs_container qc4rs

Default usage

Default training and evaluation of a hybrid system with EuroSAT data. By default, binary classification of the EuroSAT classes AnnualCrop and SeaLake is performed. The Dimensionality reduction is performed with a VGG16 combined with a deep autoencoder and the compressed data is angle encoded and classified by a parameterized quantum circuit with the FVQC as training layer.

python train.py

The training layer and all other parts can be easily adapted by several parameters. For a description of the parameter usage check:

python train.py --help

One-versus-rest multiclass classification can be performed. However, the script is limited since only classification of the EuroSAT dataset with preprocessing by a VGG16 combined with a deep autoencoder and the FPQC for classification can be performed with sufficent accuracy. For multiclass classification execute:

python train_ovr.py

Acknowledgements

• FPQC Farhi et al., 2018
• GPQC Grant et al., 2018
• VGG16 Simonyan et al., 2014
• EuroSAT Helber et al., 2019
• NWPU-RESISC45 Cheng et al., 2017
• Original RBM implementation eugenet12, 2020