from __future__ import absolute_import from __future__ import division from __future__ import print_function from datetime import datetime import gzip import io import os import pickle import shutil import tempfile import time import uuid import ray from ray.tune import TuneError from ray.tune.logger import UnifiedLogger from ray.tune.result import DEFAULT_RESULTS_DIR from ray.tune.trial import Resources class Trainable(object): """Abstract class for trainable models, functions, etc. A call to ``train()`` on a trainable will execute one logical iteration of training. As a rule of thumb, the execution time of one train call should be large enough to avoid overheads (i.e. more than a few seconds), but short enough to report progress periodically (i.e. at most a few minutes). Calling ``save()`` should save the training state of a trainable to disk, and ``restore(path)`` should restore a trainable to the given state. Generally you only need to implement ``_train``, ``_save``, and ``_restore`` here when subclassing Trainable. Note that, if you don't require checkpoint/restore functionality, then instead of implementing this class you can also get away with supplying just a `my_train(config, reporter)` function and calling: ``register_trainable("my_func", train)`` to register it for use with Tune. The function will be automatically converted to this interface (sans checkpoint functionality). Attributes: config (obj): The hyperparam configuration for this trial. logdir (str): Directory in which training outputs should be placed. registry (obj): Tune object registry which holds user-registered classes and objects by name. """ def __init__(self, config=None, registry=None, logger_creator=None): """Initialize an Trainable. Subclasses should prefer defining ``_setup()`` instead of overriding ``__init__()`` directly. Args: config (dict): Trainable-specific configuration data. registry (obj): Object registry for user-defined envs, models, etc. If unspecified, the default registry will be used. logger_creator (func): Function that creates a ray.tune.Logger object. If unspecified, a default logger is created. """ if registry is None: from ray.tune.registry import get_registry registry = get_registry() self._initialize_ok = False self._experiment_id = uuid.uuid4().hex self.config = config or {} self.registry = registry if logger_creator: self._result_logger = logger_creator(self.config) self.logdir = self._result_logger.logdir else: logdir_prefix = datetime.today().strftime("%Y-%m-%d_%H-%M-%S") if not os.path.exists(DEFAULT_RESULTS_DIR): os.makedirs(DEFAULT_RESULTS_DIR) self.logdir = tempfile.mkdtemp( prefix=logdir_prefix, dir=DEFAULT_RESULTS_DIR) self._result_logger = UnifiedLogger(self.config, self.logdir, None) self._iteration = 0 self._time_total = 0.0 self._timesteps_total = 0 self._setup() self._initialize_ok = True self._local_ip = ray.services.get_node_ip_address() @classmethod def default_resource_request(cls, config): """Returns the resource requirement for the given configuration. This can be overriden by sub-classes to set the correct trial resource allocation, so the user does not need to. """ return Resources(cpu=1, gpu=0) @classmethod def resource_help(cls, config): """Returns a help string for configuring this trainable's resources.""" return "" def train(self): """Runs one logical iteration of training. Subclasses should override ``_train()`` instead to return results. This method auto-fills many fields, so only ``timesteps_this_iter`` is requied to be present. Returns: A TrainingResult that describes training progress. """ if not self._initialize_ok: raise ValueError( "Trainable initialization failed, see previous errors") start = time.time() result = self._train() self._iteration += 1 if result.time_this_iter_s is not None: time_this_iter = result.time_this_iter_s else: time_this_iter = time.time() - start if result.timesteps_this_iter is None: raise TuneError("Must specify timesteps_this_iter in result", result) self._time_total += time_this_iter self._timesteps_total += result.timesteps_this_iter # Include the negative loss to use as a stopping condition if result.mean_loss is not None: neg_loss = -result.mean_loss else: neg_loss = result.neg_mean_loss now = datetime.today() result = result._replace( experiment_id=self._experiment_id, date=now.strftime("%Y-%m-%d_%H-%M-%S"), timestamp=int(time.mktime(now.timetuple())), training_iteration=self._iteration, timesteps_total=self._timesteps_total, time_this_iter_s=time_this_iter, time_total_s=self._time_total, neg_mean_loss=neg_loss, pid=os.getpid(), hostname=os.uname()[1], node_ip=self._local_ip, config=self.config) self._result_logger.on_result(result) return result def save(self, checkpoint_dir=None): """Saves the current model state to a checkpoint. Subclasses should override ``_save()`` instead to save state. This method dumps additional metadata alongside the saved path. Args: checkpoint_dir (str): Optional dir to place the checkpoint. Returns: Checkpoint path that may be passed to restore(). """ checkpoint_path = self._save(checkpoint_dir or self.logdir) pickle.dump([ self._experiment_id, self._iteration, self._timesteps_total, self._time_total ], open(checkpoint_path + ".tune_metadata", "wb")) return checkpoint_path def save_to_object(self): """Saves the current model state to a Python object. It also saves to disk but does not return the checkpoint path. Returns: Object holding checkpoint data. """ tmpdir = tempfile.mkdtemp("save_to_object", dir=self.logdir) checkpoint_prefix = self.save(tmpdir) data = {} base_dir = os.path.dirname(checkpoint_prefix) for path in os.listdir(base_dir): path = os.path.join(base_dir, path) if path.startswith(checkpoint_prefix): data[os.path.basename(path)] = open(path, "rb").read() out = io.BytesIO() with gzip.GzipFile(fileobj=out, mode="wb") as f: compressed = pickle.dumps({ "checkpoint_name": os.path.basename(checkpoint_prefix), "data": data, }) if len(compressed) > 10e6: # getting pretty large print("Checkpoint size is {} bytes".format(len(compressed))) f.write(compressed) shutil.rmtree(tmpdir) return out.getvalue() def restore(self, checkpoint_path): """Restores training state from a given model checkpoint. These checkpoints are returned from calls to save(). Subclasses should override ``_restore()`` instead to restore state. This method restores additional metadata saved with the checkpoint. """ self._restore(checkpoint_path) metadata = pickle.load(open(checkpoint_path + ".tune_metadata", "rb")) self._experiment_id = metadata[0] self._iteration = metadata[1] self._timesteps_total = metadata[2] self._time_total = metadata[3] def restore_from_object(self, obj): """Restores training state from a checkpoint object. These checkpoints are returned from calls to save_to_object(). """ out = io.BytesIO(obj) info = pickle.loads(gzip.GzipFile(fileobj=out, mode="rb").read()) data = info["data"] tmpdir = tempfile.mkdtemp("restore_from_object", dir=self.logdir) checkpoint_path = os.path.join(tmpdir, info["checkpoint_name"]) for file_name, file_contents in data.items(): with open(os.path.join(tmpdir, file_name), "wb") as f: f.write(file_contents) self.restore(checkpoint_path) shutil.rmtree(tmpdir) def stop(self): """Releases all resources used by this trainable.""" if self._initialize_ok: self._result_logger.close() self._stop() def _train(self): """Subclasses should override this to implement train().""" raise NotImplementedError def _save(self, checkpoint_dir): """Subclasses should override this to implement save().""" raise NotImplementedError def _restore(self, checkpoint_path): """Subclasses should override this to implement restore().""" raise NotImplementedError def _setup(self): """Subclasses should override this for custom initialization.""" pass def _stop(self): """Subclasses should override this for any cleanup on stop.""" pass def wrap_function(train_func): from ray.tune.function_runner import FunctionRunner class WrappedFunc(FunctionRunner): def _trainable_func(self): return train_func return WrappedFunc