Spectral Information Processing with Quantum Neural Networks [SIPwQNN] model samples from ibm_brisbane

Here we present the samples returned by our QML model ran on the ibm_brisbane device over 18 - 23 I 2025. It was used for pixel-level change detection on the OSCD dataset. We also present some auxiliary files and data we used. Below, we provide a brief description of each file.

Data

First, we describe all the samples we have. It's vital to note that the data we obtained using real device was gathered with an untrained model due to an error in the code. Therefore, the most important aggregated data pickle is the one containing the untrained outputs comparison.

• untrained_sim_final_analysis_data.pkl --- The most vital part of this dataset. This pickle file contains the aggregated outputs of the model with untrained weights (see auxiliary files description) ran on both simulator and real device (ibm_brisbane). These samples can be rightfully compared to assess the differences between the quality of outputs from the simulators and real device.

• untrained_sim_final_analysis_data.hdf5--- same data as above, stored in a hdf5 format.

• trained_sim_final_analysis_data.pkl --- A pickle file containing outputs of the trained model ran on the simulator and the untrained model ran on a real device (ibm_brisbane). It may be used to obtain some additional insights.

• trained_sim_final_analysis_data.hdf5 --- same data as above, stored in a hdf5 format.

• raw.zip --- This is an .zip archive containing all the raw backup (.txt) data from the real device run. It contains the value of model inputs, outputs and the time of a given sample processing.

• auxiliary_data.zip --- The result of processing the rest of the dataset test batches using trained and untrained model weights, both ran using noiseless simulator.

Scripts

Variants of the scripts we used to process the data. The scripts we used to obtain the data are available (in possibly different form) in the project repository.

• brisbane_data_preparation.py --- The scripts used to process the raw data.

• brisbane_analysis.py --- The scripts we used to analyze the presented data.

Auxiliary files

Additional files that may be of interest to the people accessing this dataset.

• mc_noiseless_default.pth --- That's the model we used to gather the data. One again, we highlight the fact that the quantum layer of the model had random weights during the data gathering. We include the used weights in the following pickle file.

• refac_mc_noiseless_default.pth --- The same model as above, but prepared for the latest version of the code.

• model_weights.pkl --- A pickle file containing weights of the model while it was ran on the real device.

• model_weights.hdf5 --- same data as above, stored in a hdf5 format.

• training_curve_data.txt --- A file containing training accuracy / loss and validation accuracy / loss during each epoch of the model training.

• pip_freeze.txt --- A results of pip freeze command, describing the python environment we used to gather the data. We made some changes to the pennylane-qiskit library. See the next section for details.

• computations_cost_estimation.xlsx --- A spreadsheet containing rough cost estimates of obtaining the samples we present.

• transpiled_model_example.txt --- An example of the model we run. The weights were set so that PennyLane could print it.

Modified penylane-qiskit files

To run the model on ibm_brisbane we had to make some changes in the pennylane-qiskit library.

• qiskit_device.py --- Our version of the pennlane_qiskit.qiskit_device script. It contains a hotfix that forbids qiskit_session closing after a single sample is sent for processing on the device.

• converter.py --- Our version of the pennylane-qiskit.converter. It contains a hotfix for ECR Gates compilation. In the end, we didn't have to use it, since we didn't use automatic transpilation of the model.

Acknowledgements

Supported by ESA under the contract No.~4000137375/22/NL/GLC/my.

Supported by PMW programme under the contract No. 5304/ESA/2023/0.

The simulations were carried out based on the IBM Quantum environment, access to which is financed within the framework of the project “Development of digital competences in quantum engineering in 2024-2025” carried out by the Poznan Supercomputing and Networking Center.