@inproceedings{hanruiwang2022quantumnas,
title = {Quantumnas: Noise-adaptive search for robust quantum circuits},
author = {Wang, Hanrui and Ding, Yongshan and Gu, Jiaqi and Li, Zirui and Lin, Yujun and Pan, David Z and Chong, Frederic T and Han, Song},
booktitle = {The 28th IEEE International Symposium on High-Performance Computer Architecture (HPCA-28)},
year = {2022}
}
A PyTorch-based hybrid classical-quantum dynamic neural networks framework.
- Welcome to contribute! Please contact us or post in the forum if you want to have new examples implemented by TorchQuantum or any other questions.
- Qmlsys website goes online: qmlsys.mit.edu
git clone https://github.com/mit-han-lab/torchquantum.git
cd torchquantum
pip install --editable .
python fix_qiskit_parameterization.py
- [HPCA'22] QuantumNAS: Noise-Adaptive Search for Robust Quantum Circuits
- [DAC'22] RobustQNN: Noise-Aware Training for Robust Quantum Neural Networks
- [DAC'22] On-Chip QNN: Towards Efficient On-Chip Training of Quantum Neural Networks
Construct quantum NN models as simple as constructing a normal pytorch model.
import torch.nn as nn
import torch.nn.functional as F
import torchquantum as tq
import torchquantum.functional as tqf
class QFCModel(nn.Module):
def __init__(self):
super().__init__()
self.n_wires = 4
self.q_device = tq.QuantumDevice(n_wires=self.n_wires)
self.measure = tq.MeasureAll(tq.PauliZ)
self.encoder_gates = [tqf.rx] * 4 + [tqf.ry] * 4 + \
[tqf.rz] * 4 + [tqf.rx] * 4
self.rx0 = tq.RX(has_params=True, trainable=True)
self.ry0 = tq.RY(has_params=True, trainable=True)
self.rz0 = tq.RZ(has_params=True, trainable=True)
self.crx0 = tq.CRX(has_params=True, trainable=True)
def forward(self, x):
bsz = x.shape[0]
# down-sample the image
x = F.avg_pool2d(x, 6).view(bsz, 16)
# reset qubit states
self.q_device.reset_states(bsz)
# encode the classical image to quantum domain
for k, gate in enumerate(self.encoder_gates):
gate(self.q_device, wires=k % self.n_wires, params=x[:, k])
# add some trainable gates (need to instantiate ahead of time)
self.rx0(self.q_device, wires=0)
self.ry0(self.q_device, wires=1)
self.rz0(self.q_device, wires=3)
self.crx0(self.q_device, wires=[0, 2])
# add some more non-parameterized gates (add on-the-fly)
tqf.hadamard(self.q_device, wires=3)
tqf.sx(self.q_device, wires=2)
tqf.cnot(self.q_device, wires=[3, 0])
tqf.qubitunitary(self.q_device0, wires=[1, 2], params=[[1, 0, 0, 0],
[0, 1, 0, 0],
[0, 0, 0, 1j],
[0, 0, -1j, 0]])
# perform measurement to get expectations (back to classical domain)
x = self.measure(self.q_device).reshape(bsz, 2, 2)
# classification
x = x.sum(-1).squeeze()
x = F.log_softmax(x, dim=1)
return x
- Easy construction of parameterized quantum circuits in PyTorch.
- Support batch mode inference and training on CPU/GPU.
- Support dynamic computation graph for easy debugging.
- Support easy deployment on real quantum devices such as IBMQ.
- Support more gates
- Support compile a unitary with descriptions to speedup training
- Support other measurements other than analytic method
- In einsum support multiple qubit sharing one letter. So that more than 26 qubit can be simulated.
- Support bmm based implementation to solve scalability issue
- Support conversion from torchquantum to qiskit
- Python >= 3.7
- PyTorch >= 1.8.0
- configargparse >= 0.14
- GPU model training requires NVIDIA GPUs
Train a quantum circuit to perform MNIST task and deploy on the real IBM Quito quantum computer as in mnist_example.py script:
cd examples/simple_mnist
python mnist_example.py
File | Description |
---|---|
devices.py | QuantumDevice class which stores the statevector |
encoding.py | Encoding layers to encode classical values to quantum domain |
functional.py | Quantum gate functions |
operators.py | Quantum gate classes |
layers.py | Layer templates such as RandomLayer |
measure.py | Measurement of quantum states to get classical values |
graph.py | Quantum gate graph used in static mode |
super_layer.py | Layer templates for SuperCircuits |
plugins/qiskit* | Convertors and processors for easy deployment on IBMQ |
examples/ | More examples for training QML and VQE models |
TorchQuantum Forum
Hanrui Wang hanrui@mit.edu