%load_ext d2lbook.tab
tab.interact_select(["pytorch"])

Hyperparameter Optimization API

:label:sec_api_hpo

Before we dive into the methodology, we will first discuss a basic code structure that allows us to efficiently implement various HPO algorithms. In general, all HPO algorithms considered here need to implement two decision making primitives, searching and scheduling. First, they need to sample new hyperparameter configurations, which often involves some kind of search over the configuration space. Second, for each configuration, an HPO algorithm needs to schedule its evaluation and decide how many resources to allocate for it. Once we start to evaluate a configuration, we will refer to it as a trial. We map these decisions to two classes, HPOSearcher and HPOScheduler. On top of that, we also provide a HPOTuner class that executes the optimization process.

This concept of scheduler and searcher is also implemented in popular HPO libraries, such as Syne Tune :cite:salinas-automl22, Ray Tune :cite:liaw-arxiv18 or Optuna :cite:akiba-sigkdd19.

%%tab pytorch
import time
from d2l import torch as d2l
from scipy import stats

Searcher

Below we define a base class for searchers, which provides a new candidate configuration through the sample_configuration function. A simple way to implement this function would be to sample configurations uniformly at random, as we did for random search in :numref:sec_what_is_hpo. More sophisticated algorithms, such as Bayesian optimization, will make these decisions based on the performance of previous trials. As a result, these algorithms are able to sample more promising candidates over time. We add the update function in order to update the history of previous trials, which can then be exploited to improve our sampling distribution.

%%tab pytorch
class HPOSearcher(d2l.HyperParameters):  #@save
    def sample_configuration() -> dict:
        raise NotImplementedError

    def update(self, config: dict, error: float, additional_info=None):
        pass

The following code shows how to implement our random search optimizer from the previous section in this API. As a slight extension, we allow the user to prescribe the first configuration to be evaluated via initial_config, while subsequent ones are drawn at random.

%%tab pytorch
class RandomSearcher(HPOSearcher):  #@save
    def __init__(self, config_space: dict, initial_config=None):
        self.save_hyperparameters()

    def sample_configuration(self) -> dict:
        if self.initial_config is not None:
            result = self.initial_config
            self.initial_config = None
        else:
            result = {
                name: domain.rvs()
                for name, domain in self.config_space.items()
            }
        return result

Scheduler

Beyond sampling configurations for new trials, we also need to decide when and for how long to run a trial. In practice, all these decisions are done by the HPOScheduler, which delegates the choice of new configurations to a HPOSearcher. The suggest method is called whenever some resource for training becomes available. Apart from invoking sample_configuration of a searcher, it may also decide upon parameters like max_epochs (i.e., how long to train the model for). The update method is called whenever a trial returns a new observation.

%%tab pytorch
class HPOScheduler(d2l.HyperParameters):  #@save
    def suggest(self) -> dict:
        raise NotImplementedError

    def update(self, config: dict, error: float, info=None):
        raise NotImplementedError

To implement random search, but also other HPO algorithms, we only need a basic scheduler that schedules a new configuration every time new resources become available.

%%tab pytorch
class BasicScheduler(HPOScheduler):  #@save
    def __init__(self, searcher: HPOSearcher):
        self.save_hyperparameters()

    def suggest(self) -> dict:
        return self.searcher.sample_configuration()

    def update(self, config: dict, error: float, info=None):
        self.searcher.update(config, error, additional_info=info)

Tuner

Finally, we need a component that runs the scheduler/searcher and does some book-keeping of the results. The following code implements a sequential execution of the HPO trials that evaluates one training job after the next and will serve as a basic example. We will later use Syne Tune for more scalable distributed HPO cases.

%%tab pytorch
class HPOTuner(d2l.HyperParameters):  #@save
    def __init__(self, scheduler: HPOScheduler, objective: callable):
        self.save_hyperparameters()
        # Bookeeping results for plotting
        self.incumbent = None
        self.incumbent_error = None
        self.incumbent_trajectory = []
        self.cumulative_runtime = []
        self.current_runtime = 0
        self.records = []

    def run(self, number_of_trials):
        for i in range(number_of_trials):
            start_time = time.time()
            config = self.scheduler.suggest()
            print(f"Trial {i}: config = {config}")
            error = self.objective(**config)
            error = float(d2l.numpy(error.cpu()))
            self.scheduler.update(config, error)
            runtime = time.time() - start_time
            self.bookkeeping(config, error, runtime)
            print(f"    error = {error}, runtime = {runtime}")

Bookkeeping the Performance of HPO Algorithms

With any HPO algorithm, we are mostly interested in the best performing configuration (called incumbent) and its validation error after a given wall-clock time. This is why we track runtime per iteration, which includes both the time to run an evaluation (call of objective) and the time to make a decision (call of scheduler.suggest). In the sequel, we will plot cumulative_runtime against incumbent_trajectory in order to visualize the any-time performance of the HPO algorithm defined in terms of scheduler (and searcher). This allows us to quantify not only how well the configuration found by an optimizer works, but also how quickly an optimizer is able to find it.

%%tab pytorch
@d2l.add_to_class(HPOTuner)  #@save
def bookkeeping(self, config: dict, error: float, runtime: float):
    self.records.append({"config": config, "error": error, "runtime": runtime})
    # Check if the last hyperparameter configuration performs better 
    # than the incumbent
    if self.incumbent is None or self.incumbent_error > error:
        self.incumbent = config
        self.incumbent_error = error
    # Add current best observed performance to the optimization trajectory
    self.incumbent_trajectory.append(self.incumbent_error)
    # Update runtime
    self.current_runtime += runtime
    self.cumulative_runtime.append(self.current_runtime)

Example: Optimizing the Hyperparameters of a Convolutional Neural Network

We now use our new implementation of random search to optimize the batch size and learning rate of the LeNet convolutional neural network from :numref:sec_lenet. We being by defining the objective function, which will once more be validation error.

%%tab pytorch
def hpo_objective_lenet(learning_rate, batch_size, max_epochs=10):  #@save
    model = d2l.LeNet(lr=learning_rate, num_classes=10)
    trainer = d2l.HPOTrainer(max_epochs=max_epochs, num_gpus=1)
    data = d2l.FashionMNIST(batch_size=batch_size)
    model.apply_init([next(iter(data.get_dataloader(True)))[0]], d2l.init_cnn)
    trainer.fit(model=model, data=data)
    validation_error = trainer.validation_error()
    return validation_error

We also need to define the configuration space. Moreover, the first configuration to be evaluated is the default setting used in :numref:sec_lenet.

config_space = {
    "learning_rate": stats.loguniform(1e-2, 1),
    "batch_size": stats.randint(32, 256),
}
initial_config = {
    "learning_rate": 0.1,
    "batch_size": 128,
}

Now we can start our random search:

searcher = RandomSearcher(config_space, initial_config=initial_config)
scheduler = BasicScheduler(searcher=searcher)
tuner = HPOTuner(scheduler=scheduler, objective=hpo_objective_lenet)
tuner.run(number_of_trials=5)

Below we plot the optimization trajectory of the incumbent to get the any-time performance of random search:

board = d2l.ProgressBoard(xlabel="time", ylabel="error")
for time_stamp, error in zip(
    tuner.cumulative_runtime, tuner.incumbent_trajectory
):
    board.draw(time_stamp, error, "random search", every_n=1)

Comparing HPO Algorithms

Just as with training algorithms or model architectures, it is important to understand how to best compare different HPO algorithms. Each HPO run depends on two major sources of randomness: the random effects of the training process, such as random weight initialization or mini-batch ordering, and the intrinsic randomness of the HPO algorithm itself, such as the random sampling of random search. Hence, when comparing different algorithms, it is crucial to run each experiment several times and report statistics, such as mean or median, across a population of multiple repetitions of an algorithm based on different seeds of the random number generator.

To illustrate this, we compare random search (see :numref:sec_rs) and Bayesian optimization :cite:snoek-nips12 on tuning the hyperparameters of a feed-forward neural network. Each algorithm was evaluated \(50\) times with a different random seed. The solid line indicates the average performance of the incumbent across these \(50\) repetitions and the dashed line the standard deviation. We can see that random search and Bayesian optimization perform roughly the same up to ~1000 seconds, but Bayesian optimization can make use of the past observation to identify better configurations and thus quickly outperforms random search afterwards.

:label:example_anytime_performance

Summary

This section laid out a simple, yet flexible interface to implement various HPO algorithms that we will look at in this chapter. Similar interfaces can be found in popular open-source HPO frameworks. We also looked at how we can compare HPO algorithms, and potential pitfall one needs to be aware.

Exercises

1. The goal of this exercise is to implement the objective function for a slightly more challenging HPO problem, and to run more realistic experiments. We will use the two hidden layer MLP DropoutMLP implemented in :numref:sec_dropout.
   1. Code up the objective function, which should depend on all hyperparameters of the model and batch_size. Use max_epochs=50. GPUs do not help here, so num_gpus=0. Hint: Modify hpo_objective_lenet.
   2. Choose a sensible search space, where num_hiddens_1, num_hiddens_2 are integers in \([8, 1024]\), and dropout values lie in \([0, 0.95]\), while batch_size lies in \([16, 384]\). Provide code for config_space, using sensible distributions from scipy.stats.
   3. Run random search on this example with number_of_trials=20 and plot the results. Make sure to first evaluate the default configuration of :numref:sec_dropout, which is initial_config = {'num_hiddens_1': 256, 'num_hiddens_2': 256, 'dropout_1': 0.5, 'dropout_2': 0.5, 'lr': 0.1, 'batch_size': 256}.
2. In this exercise, you will implement a new searcher (subclass of HPOSearcher) which makes decisions based on past data. It depends on parameters probab_local, num_init_random. Its sample_configuration method works as follows. For the first num_init_random calls, do the same as RandomSearcher.sample_configuration. Otherwise, with probability 1 - probab_local, do the same as RandomSearcher.sample_configuration. Otherwise, pick the configuration which attained the smallest validation error so far, select one of its hyperparameters at random, and sample its value randomly like in RandomSearcher.sample_configuration, but leave all other values the same. Return this configuration, which is identical to the best configuration so far, except in this one hyperparameter.
   1. Code up this new LocalSearcher. Hint: Your searcher requires config_space as argument at construction. Feel free to use a member of type RandomSearcher. You will also have to implement the update method.
   2. Re-run the experiment from the previous exercise, but using your new searcher instead of RandomSearcher. Experiment with different values for probab_local, num_init_random. However, note that a proper comparison between different HPO methods requires repeating experiments several times, and ideally considering a number of benchmark tasks.

:begin_tab:pytorch Discussions :end_tab:

---

最后更新: November 25, 2023
创建日期: November 25, 2023