Photonic Quantum Generative Adversarial Networks for Classical Data

Paper Information

Title: Photonic quantum generative adversarial networks for classical data

Authors: Tigran Sedrakyan and Alexia Salavrakos

Published: Optica Quantum, vol. 2, no. 6, pp. 458-467 (2025)

DOI: 10.1364/OPTICAQ.530346 Paper URL: arXiv:2405.06023

Reproduction Status: 🚧 In Progress (MerLin QuantumLayer generator implemented; extended benchmarking and write-up ongoing)

Reproducer: Alexia Salavrakos & Cassandre Notton (cassandre.notton@quandela.com)

Abstract

This work introduces a photonic QGAN that learns classical image distributions with linear optical circuits in Fock-space encoding. The adversarial setup follows the classical GAN game: a generator synthesizes samples while a discriminator learns to separate real and generated data.

Unlike qubit-only QGAN proposals, the paper targets a near-term photonic path by using interferometers, photon-number states, and measurement-based readout. The reported experiments show end-to-end learning of image data, including hardware-compatible settings on a single-photon platform.

Significance

The paper closes an implementation gap between GAN theory and practical photonic hardware constraints. It demonstrates that adversarial training can be done with linear optics and Fock measurements, without requiring fault-tolerant gate-model devices.

For MerLin users, this is a strong benchmark problem because it combines:

variational photonic circuits,
data encoding in circuit parameters,
stochastic measurement outputs,
and a hybrid optimization loop (quantum generator plus classical discriminator).

MerLin Implementation

The current reproduced implementation (in merlinquantum/reproduced_papers) uses a MerLin-native photonic generator path. PatchGenerator is implemented as a torch.nn.Module and internally instantiates one ML.QuantumLayer per patch. The adversarial loop in lib/qgan.py is fully differentiable:

discriminator and generator are optimized with Adam,
BCE-with-logits loss is used for both adversaries,
optional label smoothing and configurable d_steps / g_steps are supported.

At this stage:

smoke mode is a lightweight wiring check,
digits and ideal train the MerLin QuantumLayer generator end-to-end,
hp_study performs successive-halving hyperparameter search on Adam-based training,
PatchGeneratorLegacy remains available as a reference for the previous sampler/SPSA workflow.

Key Contributions Reproduced

Project Packaging and Execution

Reproduction code is modularized into generator/discriminator/model/runner components.
Config-driven execution is available through configs/defaults.json.
digits and ideal workflows are scriptable from implementation.py.

MerLin QuantumLayer Generator

Patch-wise image generation is implemented with a ModuleList of ML.QuantumLayer objects.
Each patch layer is built from ParametrizedQuantumCircuit and receives latent features through named encoding parameters.
Fock-basis probabilities are regrouped and mapped into patch pixels, then concatenated into full images.

Differentiable GAN Training

Generator and discriminator are both trained with Adam in PyTorch.
Training supports configurable update ratios (d_steps, g_steps) and smoothed labels.
SSIM-based similarity/diversity metrics are computed during runs.

Implementation Details

The implemented PatchGenerator in lib/generators.py builds gen_count independent quantum layers and treats each as a patch generator. For a latent minibatch z:

# 1) Evaluate each patch quantum layer
raw_results_list = [m(z) for m in self.models]

# 2) Regroup measurement probabilities with output_map
gen_out = torch.zeros((B, self.bin_count), device=device, dtype=dtype)
gen_out.index_add_(1, idx, res)

# 3) Normalize and resize each patch to expected pixel count
gen_out = gen_out / (total_count + 1e-8)
img_gen = gen_out[:, surplus_half:surplus_half + expected_len]  # or zero-pad

# 4) Concatenate patch tensors into one image vector
return torch.cat(patches, dim=1)

PatchGeneratorLegacy remains in the same file as a reference implementation of the previous Perceval sampler path.

Experimental Results

The repository now provides MerLin-based training modes and artifacts, with ongoing work focused on broad benchmark coverage and final reporting.

Current Reproduction Status
Mode	Status	Notes
`smoke`	✅ Available	Fast pipeline validation; writes a completion marker.
`digits`	✅ Available (MerLin)	Optdigits subset training with the `QuantumLayer` patch generator.
`ideal`	✅ Available (MerLin)	Grid sweep over architecture/input-state configurations with differentiable training.
`hp_study`	✅ Available	Successive-halving hyperparameter search for Adam training parameters.
MerLin `QuantumLayer` backend	✅ Implemented	Active in `lib/generators.py` as the default `PatchGenerator` path.

Note

Result plots and generated samples are currently documented in the reproduced-papers project notebooks and output CSV artifacts.

How QuantumLayer Fits This Reproduction

Current behavior in `PatchGenerator`

PatchGenerator currently owns:

a ModuleList of ML.QuantumLayer blocks (one per patch),
named input parameter binding from latent noise to circuit encoding parameters,
trainable variational circuit parameters exposed as PyTorch parameters,
tensor-native post-processing from output probabilities to image patches.

Gradients now propagate through the quantum generator, so the GAN loop can use standard PyTorch optimizers.

Implemented `QuantumLayer` pattern

The active implementation in lib/generators.py follows this structure:

import merlin as ml
import torch
import torch.nn as nn

class PatchGeneratorMerlin(nn.Module):
    def __init__(self, gen_count, gen_arch, input_state):
        super().__init__()
        self.layers = nn.ModuleList()
        for _ in range(gen_count):
            pcvl_circuit = ParametrizedQuantumCircuit(len(input_state), gen_arch)
            layer = ml.QuantumLayer(
                input_size=len(pcvl_circuit.enc_param_names),
                circuit=pcvl_circuit.circuit,
                input_parameters=pcvl_circuit.enc_param_names,
                trainable_parameters=pcvl_circuit.var_param_names,
                input_state=input_state,
                measurement_strategy=ml.MeasurementStrategy.PROBABILITIES,
                computation_space=ml.ComputationSpace.FOCK,
            )
            self.layers.append(layer)

    def forward(self, z):
        patch_probs = [layer(z) for layer in self.layers]
        return torch.cat(patch_probs, dim=-1)

The file also keeps PatchGeneratorLegacy for backward comparison with the older sampler-based implementation.

With this implementation:

variational parameters become standard trainable PyTorch parameters,
generator updates use torch.optim (Adam in lib/qgan.py),
the default training loop no longer depends on SPSA.

Interactive Exploration

Notebook status in this docs repository: dedicated photonic_QGAN notebook page is not yet published under docs/source/notebooks/reproduced_papers.

Available notebooks in reproduced-papers project:

notebooks/qgan_digits.ipynb
notebooks/qgan_ideal.ipynb
notebooks/qgan_noisy.ipynb
notebooks/classical_gan.ipynb
notebooks/analyse.ipynb

Extensions and Future Work

The key MerLin-native generator path is now implemented. The next work items focus on validation and cleanup.

Planned migration steps

Finalize benchmark tables for digits, ideal, and hp_study in docs.
Update deprecated measurement API usage to MeasurementStrategy.probs(...) style.
Decide long-term status of PatchGeneratorLegacy (keep optional backend or deprecate).
Consolidate notebook narratives around the MerLin default path.

Code Access and Documentation

Reproduction Repository: merlinquantum/reproduced_papers (photonic_QGAN)

Original Paper Repository: Quandela/photonic-qgan

For command-line usage and mode details, see the project README: photonic_QGAN README.

Citation

@article{sedrakyan2025photonicqgan,
  title={Photonic quantum generative adversarial networks for classical data},
  author={Sedrakyan, Tigran and Salavrakos, Alexia},
  journal={Optica Quantum},
  volume={2},
  number={6},
  pages={458--467},
  year={2025},
  doi={10.1364/OPTICAQ.530346}
}