fabric_example.py

"""
This example illustrates how SwiftLoader can be used with Fabric from lightning

"""

# Copyright The Lightning AI team.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
#     http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
"""Here are 4 easy steps to use Fabric in your PyTorch code.

1. Create the Lightning Fabric object at the beginning of your script.

2. Remove all ``.to`` and ``.cuda`` calls since Fabric will take care of it.

3. Apply ``setup`` over each model and optimizers pair, ``setup_dataloaders`` on all your dataloaders,
and replace ``loss.backward()`` with ``self.backward(loss)``.

4. Run the script from the terminal using ``lightning run model path/to/train.py``

Accelerate your training loop by setting the ``--accelerator``, ``--strategy``, ``--devices`` options directly from
the command line. See ``lightning run model --help`` or learn more from the documentation:
https://lightning.ai/docs/fabric.

"""

import argparse
import os
from os import path

import torch
import torch.nn as nn
import torch.nn.functional as F
import torch.optim as optim
import torchvision.transforms as T
from lightning.fabric import Fabric, seed_everything
from torch.optim.lr_scheduler import StepLR
from torchmetrics.classification import Accuracy
from torchvision.datasets import MNIST
from swift_loader import SwiftLoader

DATASETS_PATH = path.join(path.dirname(__file__), "datasets")
os.makedirs(DATASETS_PATH, exist_ok=True)


class Net(nn.Module):
    def __init__(self) -> None:
        super().__init__()
        self.conv1 = nn.Conv2d(1, 32, 3, 1)
        self.conv2 = nn.Conv2d(32, 64, 3, 1)
        self.dropout1 = nn.Dropout(0.25)
        self.dropout2 = nn.Dropout(0.5)
        self.fc1 = nn.Linear(9216, 128)
        self.fc2 = nn.Linear(128, 10)

    def forward(self, x):
        x = self.conv1(x)
        x = F.relu(x)
        x = self.conv2(x)
        x = F.relu(x)
        x = F.max_pool2d(x, 2)
        x = self.dropout1(x)
        x = torch.flatten(x, 1)
        x = self.fc1(x)
        x = F.relu(x)
        x = self.dropout2(x)
        x = self.fc2(x)
        return F.log_softmax(x, dim=1)


def run(hparams):
    # Create the Lightning Fabric object. The parameters like accelerator, strategy, devices etc. will be proided
    # by the command line. See all options: `lightning run model --help`
    fabric = Fabric()

    seed_everything(hparams.seed)  # instead of torch.manual_seed(...)

    transform = T.Compose([T.ToTensor(), T.Normalize((0.1307,), (0.3081,))])

    # we still keep the below code to ensure that data is downlaoded before any other processes try to access it
    # Let rank 0 download the data first, then everyone will load MNIST
    with fabric.rank_zero_first(
        local=False
    ):  # set `local=True` if your filesystem is not shared between machines
        train_dataset = MNIST(
            DATASETS_PATH,
            download=fabric.is_global_zero,
            train=True,
            transform=transform,
        )
        test_dataset = MNIST(
            DATASETS_PATH,
            download=fabric.is_global_zero,
            train=False,
            transform=transform,
        )

    # train_loader = torch.utils.data.DataLoader(
    #     train_dataset,
    #     batch_size=hparams.batch_size,
    # )
    # test_loader = torch.utils.data.DataLoader(
    #     test_dataset, batch_size=hparams.batch_size
    # )

    # # don't forget to call `setup_dataloaders` to prepare for dataloaders for distributed training.
    # train_loader, test_loader = fabric.setup_dataloaders(train_loader, test_loader)

    """
    Instead of using torch loader with fabric, we can use the SwiftLoader here

    When using SwiftLoader, we need to know how many GPUs in total (nb_consumer)
    One easy way is to pass this number to the script via an argument
    However, we can also determine it automatically like the following
    """

    if "SWIFT_LOADER_CONSUMER" not in os.environ:
        nb_consumer = max(1, torch.cuda.device_count())
        os.environ["SWIFT_LOADER_CONSUMER"] = nb_consumer
    else:
        nb_consumer = int(os.environ["SWIFT_LOADER_CONSUMER"])

    # now create SwiftLoader
    train_loader = SwiftLoader(
        dataset_class=MNIST,
        dataset_kwargs={"root": DATASETS_PATH, "train": True, "transform": transform},
        batch_size=hparams.batch_size,
        nb_consumer=nb_consumer,
        worker_per_consumer=hparams.worker_per_consumer,
        shuffle=True,
    )
    test_loader = SwiftLoader(
        dataset_class=MNIST,
        dataset_kwargs={"root": DATASETS_PATH, "train": False, "transform": transform},
        batch_size=hparams.batch_size,
        nb_consumer=nb_consumer,
        worker_per_consumer=hparams.worker_per_consumer,
        shuffle=True,  # if shuffle doesnt affect the result, shuffle=True is desirable
    )

    model = Net()  # remove call to .to(device)
    optimizer = optim.Adadelta(model.parameters(), lr=hparams.lr)

    # don't forget to call `setup` to prepare for model / optimizer for distributed training.
    # the model is moved automatically to the right device.
    model, optimizer = fabric.setup(model, optimizer)

    scheduler = StepLR(optimizer, step_size=1, gamma=hparams.gamma)

    # use torchmetrics instead of manually computing the accuracy
    test_acc = Accuracy(task="multiclass", num_classes=10).to(fabric.device)

    """
    Before consuming the worker,
    we should call SwiftLoader.start(worker_index: int, device: torch.device, to_device: callable=None)
    """

    train_loader.start(
        consumer_index=fabric.global_rank,
        device=fabric.device,
    )
    test_loader.start(
        consumer_index=fabric.global_rank,
        device=fabric.device,
    )

    # EPOCH LOOP
    for epoch in range(1, hparams.epochs + 1):
        # TRAINING LOOP
        model.train()
        for batch_idx, (data, target) in enumerate(train_loader):
            # NOTE: no need to call `.to(device)` on the data, target
            optimizer.zero_grad()
            output = model(data)
            loss = F.nll_loss(output, target)
            fabric.backward(loss)  # instead of loss.backward()

            optimizer.step()
            if (batch_idx == 0) or ((batch_idx + 1) % hparams.log_interval == 0):
                print(
                    "Train Epoch: {} [{}/{} ({:.0f}%)]\tLoss: {:.6f}".format(
                        epoch,
                        batch_idx * len(data),
                        len(train_loader.dataset),
                        100.0 * batch_idx / len(train_loader),
                        loss.item(),
                    )
                )
                if hparams.dry_run:
                    break

        scheduler.step()

        # TESTING LOOP
        model.eval()
        test_loss = 0
        with torch.no_grad():
            for data, target in test_loader:
                # NOTE: no need to call `.to(device)` on the data, target
                output = model(data)
                test_loss += F.nll_loss(output, target, reduction="sum").item()

                # WITHOUT TorchMetrics
                # pred = output.argmax(dim=1, keepdim=True)  # get the index of the max log-probability
                # correct += pred.eq(target.view_as(pred)).sum().item()

                # WITH TorchMetrics
                test_acc(output, target)

                if hparams.dry_run:
                    break

        # all_gather is used to aggregated the value across processes
        test_loss = fabric.all_gather(test_loss).sum() / len(test_loader.dataset)

        print(
            f"\nTest set: Average loss: {test_loss:.4f}, Accuracy: ({100 * test_acc.compute():.0f}%)\n"
        )
        test_acc.reset()

        if hparams.dry_run:
            break

    # When using distributed training, use `fabric.save`
    # to ensure the current process is allowed to save a checkpoint
    if hparams.save_model:
        fabric.save(model.state_dict(), "mnist_cnn.pt")


if __name__ == "__main__":
    # Arguments can be passed in through the CLI as normal and will be parsed here
    # Example:
    # lightning run model image_classifier.py accelerator=cuda --epochs=3
    parser = argparse.ArgumentParser(description="Fabric MNIST Example")
    parser.add_argument(
        "--batch-size",
        type=int,
        default=64,
        metavar="N",
        help="input batch size for training (default: 64)",
    )
    parser.add_argument(
        "--epochs",
        type=int,
        default=14,
        metavar="N",
        help="number of epochs to train (default: 14)",
    )
    parser.add_argument(
        "--lr",
        type=float,
        default=1.0,
        metavar="LR",
        help="learning rate (default: 1.0)",
    )
    parser.add_argument(
        "--gamma",
        type=float,
        default=0.7,
        metavar="M",
        help="Learning rate step gamma (default: 0.7)",
    )
    parser.add_argument(
        "--dry-run",
        action="store_true",
        default=False,
        help="quickly check a single pass",
    )
    parser.add_argument(
        "--seed", type=int, default=1, metavar="S", help="random seed (default: 1)"
    )
    parser.add_argument(
        "--log-interval",
        type=int,
        default=10,
        metavar="N",
        help="how many batches to wait before logging training status",
    )
    parser.add_argument(
        "--save-model",
        action="store_true",
        default=False,
        help="For Saving the current Model",
    )
    hparams = parser.parse_args()

    run(hparams)