quantum-triplet-loss

How To Run

Create a hyperparameters.json in the main directory with the desired values. A good start would be:

{
    "dataset": "mnist_ae",
    "qubits": 12,
    "data_qubits": 8,
    "output_qubits": 2,
    "layers": 10,
    "train_size": 2500,
    "test_size": 500,
    "classes": [3, 6],
    "steps": 2000,
    "test_every": 50,
    "grads_every": 1,
    "start_stepsize": 0.005,
    "update_sz_every": 35000,
    "sz_factor": 0.1,
    "alpha": 1.0,
    "shots": null,
    "seed": 1234
}

To start a Triplet Loss Training run main.py. To start a Cross Entropy Training run classical_training.py.

Parameters explanation

Parameter	Explanation
dataset	Use mnist_ae or moons to reproduce the results from the paper.
qubits	Number of qubits of the PQC.
data_qubits	On how many qubits the data is encoded. Might be fixed by the dataset choice.
output_qubits	Number of measured qubits. For CE it will always be the number of classes.
layers	Number of layers of the PQC.
train_size	Number of training samples. Might be limited by the dataset choice.
test_size	Number of test samples. Might be limited by the dataset choice.
classes	Classes (digits) used for a training with MNIST dataset.
steps	Number of training steps.
test_every	After how many steps the test dataset is evaluated.
grads_every	After how many steps the gradients are printed.
start_stepsize	Initial stepsize of the optimizer.
update_sz_every	After how many steps the stepsize is updated (LR decay).
sz_factor	Which factor is used to update the stepsize (not used for the paper).
alpha	Alpha value of the Triplet Loss.
shots	Number of shots. null for analytic mode.
seed	Random seed.

Note that only a few classes/features combinations are provided for the MNIST dataset.

Using the mnist_ae dataset

The mnist_ae dataset was created by training an autoencoder on the popular MNIST dataset. Every created dataset that we used for the experiments of our paper are provided in the quantum-triplet-loss/datasets/minst_ae directory.

Lets look at the file Train_[3,4,6,9]_features_8.npz as an example. This dataset was created from the classes/digits 3, 4, 6 and 9, and every datapoint has 8 features. To use this dataset, you have to set "data_qubits": 8 and "classes": [3, 4, 6, 9]" in your hyperparameters.json file. The respective train and test set will be loaded automatically.

Fixed Parameters

• Circuit, measurement and embedding
• 12 qubits
  • 2 outputs for cross-entropy loss
  • 2 to 8 outputs for triplet loss
• LR 0.005, Alpha 1.0 - both constant
• Training steps
  • 2000 for triplet loss
  • 6000 for cross-entropy loss
• Triplet choice completely random

Tests

Tests with positive effect

• ZZ measurement on qubits 0,1 and 2,3

Tests without any effect

• StronglyEntanglingLayers
• Change of qubit- and layer count
• Changes of learning rate (dynamic and static) – 0.01 leads to best result
• Changes of alpha
  • Bigger alpha values lead to bigger variances of measurements
  • Changing the alpha value makes previous steps obsolete
• Measuring 3 qubits
• One separate circuit per measurement
• Data-Reuploading

Tests with negative effect

• Online mining of triplets (everything is mapped to one value)
• Usual training, followed by online mining
• Change of loss