When building hybrid models, a single optimizer is usually handling all parameters ([1], [2], [3]), quantum and classical ones. This project aims to clarify if there are advantages of having individual optimizers for each part of the hybrid model.
With the current approach a full-size MNIST sample (28x28) is fed into a classical neural network. The output of this network is then embedded onto a Variational Quantum Circuit which performs the final classification.
Both parts have individual optimisers, of which the following types are planned to be tested:
- Adam (as a baseline comparison; shouldn't perform much different compared to a single Adam instance)
- SGD (different learning rates)
- NGD / QNG
- SPSA / QNSPSA [4]
Besides the overall performance, the convergence speed should be considered in evaluation.
This project is built using the Kedro Framework.
Using pip:
pip install -r src/requirements.txt
Without further configuration you can execute
kedro run
which will load MNIST, preprocess the data and start training the model.
If want an overview of the nodes and pipelines, you can execute
kedro viz
which will open Kedro`s dashboard in you browser.
The following parameters can be adjusted:
- Data Preprocessing Parameters:
conf/base/parameters/data_processing.yml- Batch size
- Seed
- Training Parameters:
conf/base/parameters/data_science.yml- Loss function
- Learning rate
- Number of epochs
- Optimizer approach
- Global Parameters:
conf/base/globals.yml- Training/ Test Size
- Number of classes
- Number of qubits
[1]: Hybrid Quantum Classical Graph Neural Networks for Particle Track Reconstruction
[2]: Quantum-classical convolutional neural networks in radiological image classification
[3]: Quantum classical hybrid neural networks for continuous variable prediction
[4]: Simultaneous Perturbation Stochastic Approximation of the Quantum Fisher Information