mnist_example.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.
"""

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.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 = 4
            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)

        @tq.static_support
        def forward(self, q_device: tq.QuantumDevice):
            """
            1. To convert tq QuantumModule to qiskit or run in the static
            model, need to:
                (1) add @tq.static_support before the forward
                (2) make sure to add
                    static=self.static_mode and
                    parent_graph=self.graph
                    to all the tqf functions, such as tqf.hadamard below
            """
            self.q_device = q_device

            self.random_layer(self.q_device)

            # some trainable gates (instantiated 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, static=self.static_mode, parent_graph=self.graph
            )
            tqf.sx(
                self.q_device, wires=2, static=self.static_mode, parent_graph=self.graph
            )
            tqf.cnot(
                self.q_device,
                wires=[3, 0],
                static=self.static_mode,
                parent_graph=self.graph,
            )

    def __init__(self):
        super().__init__()
        self.n_wires = 4
        self.q_device = tq.QuantumDevice(n_wires=self.n_wires)
        self.encoder = tq.AmplitudeEncoder()

        self.q_layer = self.QLayer()
        self.measure = tq.MeasureAll(tq.PauliZ)

    def forward(self, x, use_qiskit=False):
        bsz = x.shape[0]
        x = F.avg_pool2d(x, 6).view(bsz, 16)

        self.encoder(self.q_device, x)
        self.q_layer(self.q_device)
        x = self.measure(self.q_device)

        x = x.reshape(bsz, 2, 2).sum(-1).squeeze()
        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=[3, 6],
        n_test_samples=75,
    )
    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)

    if args.static:
        # optionally to switch to the static mode, which can bring speedup
        # on training
        model.q_layer.static_on(wires_per_block=args.wires_per_block)

    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()