Hyperparameter Tuning With Determined¶
Hyperparameter tuning is a common machine learning workflow that involves appropriately configuring the data, model architecture, and learning algorithm to yield an effective model. Hyperparameter tuning is a challenging problem in deep learning given the potentially large number of hyperparameters to consider. Determined provides support for hyperparameter search as a first-class workflow that is tightly integrated with Determined’s job scheduler, which allows for efficient execution of state-of-the-art early-stopping based approaches as well as seamless parallelization of these methods.
Adaptive Search¶
Our default recommended search method is adaptive search, an early-stopping based technique that speeds up traditional techniques like random search by periodically abandoning low-performing hyperparameter configurations in a principled fashion. Adaptive search builds on the two prototypical adaptive downsampling approaches, Successive Halving (SHA) and Hyperband, while also enabling seamless parallelization and a simpler user interface.
Determined offers two adaptive searchers that employ the same underlying algorithm but differ in their level of configurability.
Adaptive (Simple) is easier to configure and provides sensible default settings for most situations. This searcher requires just two intuitive settings: the number of configurations to evaluate, and the maximum allowed resource budget per configuration. We recommend starting with this searcher.
Adaptive (Advanced) allows users to control the adaptive search behavior in a more fine-grained way.
Experimental¶
Adaptive (Asynchronous) is an asynchronous version of the above adaptive searchers (ASHA) that is more suitable for large-scale HP search experiments in the distributed setting.
Other Supported Methods¶
Determined also supports other common hyperparameter search algorithms:
Single is appropriate for manual hyperparameter tuning, as it trains a single hyperparameter configuration.
Grid brute force evaluates all possible hyperparameter configurations and returns the best.
Random evaluates a set of hyperparameter configurations chosen at random and returns the best.
Population-based training (PBT) begins as random search but periodically replaces low-performing hyperparameter configurations with ones near the high-performing points in the hyperparameter space.
Handling Trial Errors and Early Stopping Requests¶
When a trial encounters an error or fails unexpectedly, Determined will
restart it from the latest checkpoint unless we have done so
max_restarts times, which is configured in the experiment
configuration. Once we have reached max_restarts, any further trials
that fail will be marked as errored and will not be restarted. For
search methods that adapt to validation metric values (Adaptive
(Simple), Adaptive
(Advanced), and
Population-based training (PBT)),
we do not continue training errored trials, even if the search method
would typically call for us to continue training. This behavior is
useful when some parts of the hyperparameter space result in models that
cannot be trained successfully (e.g., the search explores a range of
batch sizes and some of those batch sizes cause GPU OOM errors). An
experiment can complete successfully as long as at least one of the
trials within it completes successfully.
Trial code can also request that training be stopped early, e.g., via a framework
callback such as tf.keras.callbacks.EarlyStopping
or manually by calling determined.TrialContext.set_stop_requested(). When
early stopping is requested, Determined will finish the current training or
validation step and checkpoint the trial. Trials that are stopped early are
considered to be “completed”, whereas trials that fail are marked as “errored”.