Ray Tune 超参数调优

Hyperparameter tuning can make the difference between an average model and a highly accurate one. Often simple things like choosing a different learning rate or changing a network layer size can have a dramatic impact on your model performance.

Fortunately, there are tools that help with finding the best combination of parameters. Ray Tune is an industry standard tool for distributed hyperparameter tuning. Ray Tune includes the latest hyperparameter search algorithms, integrates with TensorBoard and other analysis libraries, and natively supports distributed training through Ray’s distributed machine learning engine.

In this tutorial, we will show you how to integrate Ray Tune into your PyTorch training workflow. We will extend this tutorial from the PyTorch documentation for training a CIFAR10 image classifier.

As you will see, we only need to add some slight modifications. In particular, we need to

1. wrap data loading and training in functions,

2. make some network parameters configurable,

3. add checkpointing (optional),

4. and define the search space for the model tuning

To run this tutorial, please make sure the following packages are installed:

• ray[tune]: Distributed hyperparameter tuning library

• torchvision: For the data transformers

Setup / Imports

Let’s start with the imports:

from functools import partial
import os
import tempfile
from pathlib import Path
import torch
import torch.nn as nn
import torch.nn.functional as F
import torch.optim as optim
from torch.utils.data import random_split
import torchvision
import torchvision.transforms as transforms
from ray import tune
from ray import train
from ray.train import Checkpoint, get_checkpoint
from ray.tune.schedulers import ASHAScheduler
import ray.cloudpickle as pickle

Most of the imports are needed for building the PyTorch model. Only the last imports are for Ray Tune.

Data loaders

We wrap the data loaders in their own function and pass a global data directory. This way we can share a data directory between different trials.

def load_data(data_dir="./data"):
    transform = transforms.Compose(
        [transforms.ToTensor(), transforms.Normalize((0.5, 0.5, 0.5), (0.5, 0.5, 0.5))]
    )

    trainset = torchvision.datasets.CIFAR10(
        root=data_dir, train=True, download=True, transform=transform
    )

    testset = torchvision.datasets.CIFAR10(
        root=data_dir, train=False, download=True, transform=transform
    )

    return trainset, testset

Configurable neural network

We can only tune those parameters that are configurable. In this example, we can specify the layer sizes of the fully connected layers:

class Net(nn.Module):
    def __init__(self, l1=120, l2=84):
        super(Net, self).__init__()
        self.conv1 = nn.Conv2d(3, 6, 5)
        self.pool = nn.MaxPool2d(2, 2)
        self.conv2 = nn.Conv2d(6, 16, 5)
        self.fc1 = nn.Linear(16 * 5 * 5, l1)
        self.fc2 = nn.Linear(l1, l2)
        self.fc3 = nn.Linear(l2, 10)

    def forward(self, x):
        x = self.pool(F.relu(self.conv1(x)))
        x = self.pool(F.relu(self.conv2(x)))
        x = torch.flatten(x, 1)  # flatten all dimensions except batch
        x = F.relu(self.fc1(x))
        x = F.relu(self.fc2(x))
        x = self.fc3(x)
        return x

The train function

Now it gets interesting, because we introduce some changes to the example from the PyTorch documentation.

We wrap the training script in a function train_cifar(config, data_dir=None). The config parameter will receive the hyperparameters we would like to train with. The data_dir specifies the directory where we load and store the data, so that multiple runs can share the same data source. We also load the model and optimizer state at the start of the run, if a checkpoint is provided. Further down in this tutorial you will find information on how to save the checkpoint and what it is used for.

net = Net(config["l1"], config["l2"])

checkpoint = get_checkpoint()
if checkpoint:
    with checkpoint.as_directory() as checkpoint_dir:
        data_path = Path(checkpoint_dir) / "data.pkl"
        with open(data_path, "rb") as fp:
            checkpoint_state = pickle.load(fp)
        start_epoch = checkpoint_state["epoch"]
        net.load_state_dict(checkpoint_state["net_state_dict"])
        optimizer.load_state_dict(checkpoint_state["optimizer_state_dict"])
else:
    start_epoch = 0

The learning rate of the optimizer is made configurable, too:

optimizer = optim.SGD(net.parameters(), lr=config["lr"], momentum=0.9)

We also split the training data into a training and validation subset. We thus train on 80% of the data and calculate the validation loss on the remaining 20%. The batch sizes with which we iterate through the training and test sets are configurable as well.

Adding (multi) GPU support with DataParallel

Image classification benefits largely from GPUs. Luckily, we can continue to use PyTorch’s abstractions in Ray Tune. Thus, we can wrap our model in nn.DataParallel to support data parallel training on multiple GPUs:

device = "cpu"
if torch.cuda.is_available():
    device = "cuda:0"
    if torch.cuda.device_count() > 1:
        net = nn.DataParallel(net)
net.to(device)

By using a device variable we make sure that training also works when we have no GPUs available. PyTorch requires us to send our data to the GPU memory explicitly, like this:

for i, data in enumerate(trainloader, 0):
    inputs, labels = data
    inputs, labels = inputs.to(device), labels.to(device)

The code now supports training on CPUs, on a single GPU, and on multiple GPUs. Notably, Ray also supports fractional GPUs so we can share GPUs among trials, as long as the model still fits on the GPU memory. We’ll come back to that later.

Communicating with Ray Tune

The most interesting part is the communication with Ray Tune:

checkpoint_data = {
    "epoch": epoch,
    "net_state_dict": net.state_dict(),
    "optimizer_state_dict": optimizer.state_dict(),
}
with tempfile.TemporaryDirectory() as checkpoint_dir:
    data_path = Path(checkpoint_dir) / "data.pkl"
    with open(data_path, "wb") as fp:
        pickle.dump(checkpoint_data, fp)

    checkpoint = Checkpoint.from_directory(checkpoint_dir)
    train.report(
        {"loss": val_loss / val_steps, "accuracy": correct / total},
        checkpoint=checkpoint,
    )

Here we first save a checkpoint and then report some metrics back to Ray Tune. Specifically, we send the validation loss and accuracy back to Ray Tune. Ray Tune can then use these metrics to decide which hyperparameter configuration lead to the best results. These metrics can also be used to stop bad performing trials early in order to avoid wasting resources on those trials.

The checkpoint saving is optional, however, it is necessary if we wanted to use advanced schedulers like Population Based Training. Also, by saving the checkpoint we can later load the trained models and validate them on a test set. Lastly, saving checkpoints is useful for fault tolerance, and it allows us to interrupt training and continue training later.

Full training function

The full code example looks like this:

def train_cifar(config, data_dir=None):
    net = Net(config["l1"], config["l2"])

    device = "cpu"
    if torch.cuda.is_available():
        device = "cuda:0"
        if torch.cuda.device_count() > 1:
            net = nn.DataParallel(net)
    net.to(device)

    criterion = nn.CrossEntropyLoss()
    optimizer = optim.SGD(net.parameters(), lr=config["lr"], momentum=0.9)

    checkpoint = get_checkpoint()
    if checkpoint:
        with checkpoint.as_directory() as checkpoint_dir:
            data_path = Path(checkpoint_dir) / "data.pkl"
            with open(data_path, "rb") as fp:
                checkpoint_state = pickle.load(fp)
            start_epoch = checkpoint_state["epoch"]
            net.load_state_dict(checkpoint_state["net_state_dict"])
            optimizer.load_state_dict(checkpoint_state["optimizer_state_dict"])
    else:
        start_epoch = 0

    trainset, testset = load_data(data_dir)

    test_abs = int(len(trainset) * 0.8)
    train_subset, val_subset = random_split(
        trainset, [test_abs, len(trainset) - test_abs]
    )

    trainloader = torch.utils.data.DataLoader(
        train_subset, batch_size=int(config["batch_size"]), shuffle=True, num_workers=8
    )
    valloader = torch.utils.data.DataLoader(
        val_subset, batch_size=int(config["batch_size"]), shuffle=True, num_workers=8
    )

    for epoch in range(start_epoch, 10):  # loop over the dataset multiple times
        running_loss = 0.0
        epoch_steps = 0
        for i, data in enumerate(trainloader, 0):
            # get the inputs; data is a list of [inputs, labels]
            inputs, labels = data
            inputs, labels = inputs.to(device), labels.to(device)

            # zero the parameter gradients
            optimizer.zero_grad()

            # forward + backward + optimize
            outputs = net(inputs)
            loss = criterion(outputs, labels)
            loss.backward()
            optimizer.step()

            # print statistics
            running_loss += loss.item()
            epoch_steps += 1
            if i % 2000 == 1999:  # print every 2000 mini-batches
                print(
                    "[%d, %5d] loss: %.3f"
                    % (epoch + 1, i + 1, running_loss / epoch_steps)
                )
                running_loss = 0.0

        # Validation loss
        val_loss = 0.0
        val_steps = 0
        total = 0
        correct = 0
        for i, data in enumerate(valloader, 0):
            with torch.no_grad():
                inputs, labels = data
                inputs, labels = inputs.to(device), labels.to(device)

                outputs = net(inputs)
                _, predicted = torch.max(outputs.data, 1)
                total += labels.size(0)
                correct += (predicted == labels).sum().item()

                loss = criterion(outputs, labels)
                val_loss += loss.cpu().numpy()
                val_steps += 1

        checkpoint_data = {
            "epoch": epoch,
            "net_state_dict": net.state_dict(),
            "optimizer_state_dict": optimizer.state_dict(),
        }
        with tempfile.TemporaryDirectory() as checkpoint_dir:
            data_path = Path(checkpoint_dir) / "data.pkl"
            with open(data_path, "wb") as fp:
                pickle.dump(checkpoint_data, fp)

            checkpoint = Checkpoint.from_directory(checkpoint_dir)
            train.report(
                {"loss": val_loss / val_steps, "accuracy": correct / total},
                checkpoint=checkpoint,
            )

    print("Finished Training")

As you can see, most of the code is adapted directly from the original example.

Test set accuracy

Commonly the performance of a machine learning model is tested on a hold-out test set with data that has not been used for training the model. We also wrap this in a function:

def test_accuracy(net, device="cpu"):
    trainset, testset = load_data()

    testloader = torch.utils.data.DataLoader(
        testset, batch_size=4, shuffle=False, num_workers=2
    )

    correct = 0
    total = 0
    with torch.no_grad():
        for data in testloader:
            images, labels = data
            images, labels = images.to(device), labels.to(device)
            outputs = net(images)
            _, predicted = torch.max(outputs.data, 1)
            total += labels.size(0)
            correct += (predicted == labels).sum().item()

    return correct / total

The function also expects a device parameter, so we can do the test set validation on a GPU.

Configuring the search space

Lastly, we need to define Ray Tune’s search space. Here is an example:

config = {
    "l1": tune.choice([2 ** i for i in range(9)]),
    "l2": tune.choice([2 ** i for i in range(9)]),
    "lr": tune.loguniform(1e-4, 1e-1),
    "batch_size": tune.choice([2, 4, 8, 16])
}

The tune.choice() accepts a list of values that are uniformly sampled from. In this example, the l1 and l2 parameters should be powers of 2 between 4 and 256, so either 4, 8, 16, 32, 64, 128, or 256. The lr (learning rate) should be uniformly sampled between 0.0001 and 0.1. Lastly, the batch size is a choice between 2, 4, 8, and 16.

At each trial, Ray Tune will now randomly sample a combination of parameters from these search spaces. It will then train a number of models in parallel and find the best performing one among these. We also use the ASHAScheduler which will terminate bad performing trials early.

We wrap the train_cifar function with functools.partial to set the constant data_dir parameter. We can also tell Ray Tune what resources should be available for each trial:

gpus_per_trial = 2
# ...
result = tune.run(
    partial(train_cifar, data_dir=data_dir),
    resources_per_trial={"cpu": 8, "gpu": gpus_per_trial},
    config=config,
    num_samples=num_samples,
    scheduler=scheduler,
    checkpoint_at_end=True)

You can specify the number of CPUs, which are then available e.g. to increase the num_workers of the PyTorch DataLoader instances. The selected number of GPUs are made visible to PyTorch in each trial. Trials do not have access to GPUs that haven’t been requested for them - so you don’t have to care about two trials using the same set of resources.

Here we can also specify fractional GPUs, so something like gpus_per_trial=0.5 is completely valid. The trials will then share GPUs among each other. You just have to make sure that the models still fit in the GPU memory.

After training the models, we will find the best performing one and load the trained network from the checkpoint file. We then obtain the test set accuracy and report everything by printing.

The full main function looks like this:

def main(num_samples=10, max_num_epochs=10, gpus_per_trial=2):
    data_dir = os.path.abspath("./data")
    load_data(data_dir)
    config = {
        "l1": tune.choice([2**i for i in range(9)]),
        "l2": tune.choice([2**i for i in range(9)]),
        "lr": tune.loguniform(1e-4, 1e-1),
        "batch_size": tune.choice([2, 4, 8, 16]),
    }
    scheduler = ASHAScheduler(
        metric="loss",
        mode="min",
        max_t=max_num_epochs,
        grace_period=1,
        reduction_factor=2,
    )
    result = tune.run(
        partial(train_cifar, data_dir=data_dir),
        resources_per_trial={"cpu": 2, "gpu": gpus_per_trial},
        config=config,
        num_samples=num_samples,
        scheduler=scheduler,
    )

    best_trial = result.get_best_trial("loss", "min", "last")
    print(f"Best trial config: {best_trial.config}")
    print(f"Best trial final validation loss: {best_trial.last_result['loss']}")
    print(f"Best trial final validation accuracy: {best_trial.last_result['accuracy']}")

    best_trained_model = Net(best_trial.config["l1"], best_trial.config["l2"])
    device = "cpu"
    if torch.cuda.is_available():
        device = "cuda:0"
        if gpus_per_trial > 1:
            best_trained_model = nn.DataParallel(best_trained_model)
    best_trained_model.to(device)

    best_checkpoint = result.get_best_checkpoint(trial=best_trial, metric="accuracy", mode="max")
    with best_checkpoint.as_directory() as checkpoint_dir:
        data_path = Path(checkpoint_dir) / "data.pkl"
        with open(data_path, "rb") as fp:
            best_checkpoint_data = pickle.load(fp)

        best_trained_model.load_state_dict(best_checkpoint_data["net_state_dict"])
        test_acc = test_accuracy(best_trained_model, device)
        print("Best trial test set accuracy: {}".format(test_acc))


if __name__ == "__main__":
    # You can change the number of GPUs per trial here:
    main(num_samples=10, max_num_epochs=10, gpus_per_trial=0)

If you run the code, an example output could look like this:

Number of trials: 10/10 (10 TERMINATED)
+-----+--------------+------+------+-------------+--------+---------+------------+
| ... |   batch_size |   l1 |   l2 |          lr |   iter |    loss |   accuracy |
|-----+--------------+------+------+-------------+--------+---------+------------|
| ... |            2 |    1 |  256 | 0.000668163 |      1 | 2.31479 |     0.0977 |
| ... |            4 |   64 |    8 | 0.0331514   |      1 | 2.31605 |     0.0983 |
| ... |            4 |    2 |    1 | 0.000150295 |      1 | 2.30755 |     0.1023 |
| ... |           16 |   32 |   32 | 0.0128248   |     10 | 1.66912 |     0.4391 |
| ... |            4 |    8 |  128 | 0.00464561  |      2 | 1.7316  |     0.3463 |
| ... |            8 |  256 |    8 | 0.00031556  |      1 | 2.19409 |     0.1736 |
| ... |            4 |   16 |  256 | 0.00574329  |      2 | 1.85679 |     0.3368 |
| ... |            8 |    2 |    2 | 0.00325652  |      1 | 2.30272 |     0.0984 |
| ... |            2 |    2 |    2 | 0.000342987 |      2 | 1.76044 |     0.292  |
| ... |            4 |   64 |   32 | 0.003734    |      8 | 1.53101 |     0.4761 |
+-----+--------------+------+------+-------------+--------+---------+------------+

Best trial config: {'l1': 64, 'l2': 32, 'lr': 0.0037339984519545164, 'batch_size': 4}
Best trial final validation loss: 1.5310075663924216
Best trial final validation accuracy: 0.4761
Best trial test set accuracy: 0.4737

Most trials have been stopped early in order to avoid wasting resources. The best performing trial achieved a validation accuracy of about 47%, which could be confirmed on the test set.

So that’s it! You can now tune the parameters of your PyTorch models.