mnist_2qubit_4class.py

"""
MIT License

Copyright (c) 2020-present TorchQuantum Authors

Permission is hereby granted, free of charge, to any person obtaining a copy
of this software and associated documentation files (the "Software"), to deal
in the Software without restriction, including without limitation the rights
to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
copies of the Software, and to permit persons to whom the Software is
furnished to do so, subject to the following conditions:

The above copyright notice and this permission notice shall be included in all
copies or substantial portions of the Software.

THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
SOFTWARE.
"""

"""
use 2 qubit to perform 4 class classification,
We can choose four different observables to measure the qubit state:
    1. XX
    2. YY
    3. ZZ
    4. XY
"""

import torch
import torch.nn.functional as F
import torch.optim as optim
import argparse

import torchquantum as tq
import torchquantum.functional as tqf

from torchquantum.measurement import expval_joint_analytical

from torchquantum.dataset import MNIST
from torch.optim.lr_scheduler import CosineAnnealingLR

import random
import numpy as np


class QFCModel(tq.QuantumModule):
    class QLayer(tq.QuantumModule):
        def __init__(self):
            super().__init__()
            self.n_wires = 2
            self.random_layer = tq.RandomLayer(
                n_ops=50, wires=list(range(self.n_wires))
            )

            # gates with trainable parameters
            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, qdev: tq.QuantumDevice):
            self.random_layer(qdev)

            # some trainable gates (instantiated ahead of time)
            self.rx0(qdev, wires=0)
            self.ry0(qdev, wires=1)
            self.rz0(qdev, wires=0)
            self.crx0(qdev, wires=[0, 1])

    def __init__(self):
        super().__init__()
        self.n_wires = 2
        # the encoder here is just for illustration purpose, may not be the best choice
        self.encoder = tq.GeneralEncoder(
            tq.encoder_op_list_name_dict["2x8_rxryrzrxryrzrxry"]
        )

        self.q_layer = self.QLayer()

    def forward(self, x, use_qiskit=False):
        qdev = tq.QuantumDevice(
            n_wires=self.n_wires, bsz=x.shape[0], device=x.device, record_op=True
        )

        bsz = x.shape[0]
        x = F.avg_pool2d(x, 6).view(bsz, 16)

        self.encoder(qdev, x)
        self.q_layer(qdev)
        obs_xx = expval_joint_analytical(qdev, "XX")
        obs_yy = expval_joint_analytical(qdev, "YY")
        obs_zz = expval_joint_analytical(qdev, "ZZ")
        obs_xy = expval_joint_analytical(qdev, "XY")

        x = torch.stack([obs_xx, obs_yy, obs_zz, obs_xy], dim=1)
        x = F.log_softmax(x, dim=1)

        return x


def train(dataflow, model, device, optimizer):
    for feed_dict in dataflow["train"]:
        inputs = feed_dict["image"].to(device)
        targets = feed_dict["digit"].to(device)

        outputs = model(inputs)
        loss = F.nll_loss(outputs, targets)
        optimizer.zero_grad()
        loss.backward()
        optimizer.step()
        print(f"loss: {loss.item()}", end="\r")


def valid_test(dataflow, split, model, device, qiskit=False):
    target_all = []
    output_all = []
    with torch.no_grad():
        for feed_dict in dataflow[split]:
            inputs = feed_dict["image"].to(device)
            targets = feed_dict["digit"].to(device)

            outputs = model(inputs, use_qiskit=qiskit)

            target_all.append(targets)
            output_all.append(outputs)
        target_all = torch.cat(target_all, dim=0)
        output_all = torch.cat(output_all, dim=0)

    _, indices = output_all.topk(1, dim=1)
    masks = indices.eq(target_all.view(-1, 1).expand_as(indices))
    size = target_all.shape[0]
    corrects = masks.sum().item()
    accuracy = corrects / size
    loss = F.nll_loss(output_all, target_all).item()

    print(f"{split} set accuracy: {accuracy}")
    print(f"{split} set loss: {loss}")


def main():
    parser = argparse.ArgumentParser()
    parser.add_argument(
        "--static", action="store_true", help="compute with " "static mode"
    )
    parser.add_argument("--pdb", action="store_true", help="debug with pdb")
    parser.add_argument(
        "--wires-per-block", type=int, default=2, help="wires per block int static mode"
    )
    parser.add_argument(
        "--epochs", type=int, default=5, help="number of training epochs"
    )

    args = parser.parse_args()

    if args.pdb:
        import pdb

        pdb.set_trace()

    seed = 0
    random.seed(seed)
    np.random.seed(seed)
    torch.manual_seed(seed)

    dataset = MNIST(
        root="./mnist_data",
        train_valid_split_ratio=[0.9, 0.1],
        digits_of_interest=[0, 1, 2, 3],
        n_test_samples=100,
    )

    dataflow = dict()

    for split in dataset:
        sampler = torch.utils.data.RandomSampler(dataset[split])
        dataflow[split] = torch.utils.data.DataLoader(
            dataset[split],
            batch_size=256,
            sampler=sampler,
            num_workers=8,
            pin_memory=True,
        )

    use_cuda = torch.cuda.is_available()
    device = torch.device("cuda" if use_cuda else "cpu")

    model = QFCModel().to(device)

    n_epochs = args.epochs
    optimizer = optim.Adam(model.parameters(), lr=5e-3, weight_decay=1e-4)
    scheduler = CosineAnnealingLR(optimizer, T_max=n_epochs)

    for epoch in range(1, n_epochs + 1):
        # train
        print(f"Epoch {epoch}:")
        train(dataflow, model, device, optimizer)
        print(optimizer.param_groups[0]["lr"])

        # valid
        valid_test(dataflow, "valid", model, device)
        scheduler.step()

    # test
    valid_test(dataflow, "test", model, device, qiskit=False)


if __name__ == "__main__":
    main()