def testFailImportingRemoteFunction(self): ray.init(start_ray_local=True, num_workers=2, driver_mode=ray.SILENT_MODE) # This example is somewhat contrived. It should be successfully pickled, and # then it should throw an exception when it is unpickled. This may depend a # bit on the specifics of our pickler. def reducer(*args): raise Exception("There is a problem here.") class Foo(object): def __init__(self): self.__name__ = "Foo_object" self.func_doc = "" self.__globals__ = {} def __reduce__(self): return reducer, () def __call__(self): return ray.remote(Foo()) for _ in range(100): # Retry if we need to wait longer. if len(ray.task_info()["failed_remote_function_imports"]) >= 1: break time.sleep(0.1) self.assertTrue("There is a problem here." in ray.task_info()["failed_remote_function_imports"][0]["error_message"]) ray.worker.cleanup()
def test_remote_training_loss(ray_start_regular): net = ray.remote(TrainActor).remote() net_values = TrainActor().values loss, variables, _, sess, grads, train, placeholders = net_values before_acc = sess.run( loss, feed_dict=dict(zip(placeholders, [[2] * 100, [4] * 100]))) for _ in range(3): gradients_list = ray.get([ net.training_step.remote(variables.get_weights()) for _ in range(2) ]) mean_grads = [ sum(gradients[i] for gradients in gradients_list) / len(gradients_list) for i in range(len(gradients_list[0])) ] feed_dict = { grad[0]: mean_grad for (grad, mean_grad) in zip(grads, mean_grads) } sess.run(train, feed_dict=feed_dict) after_acc = sess.run( loss, feed_dict=dict(zip(placeholders, [[2] * 100, [4] * 100]))) assert before_acc < after_acc
def _setup_runner(self): self.status = Trial.RUNNING trainable_cls = get_registry().get( TRAINABLE_CLASS, self.trainable_name) cls = ray.remote( num_cpus=self.resources.driver_cpu_limit, num_gpus=self.resources.driver_gpu_limit)(trainable_cls) if not self.result_logger: if not os.path.exists(self.local_dir): os.makedirs(self.local_dir) self.logdir = tempfile.mkdtemp( prefix="{}_{}".format( str(self)[:MAX_LEN_IDENTIFIER], date_str()), dir=self.local_dir) self.result_logger = UnifiedLogger( self.config, self.logdir, self.upload_dir) remote_logdir = self.logdir def logger_creator(config): # Set the working dir in the remote process, for user file writes if not os.path.exists(remote_logdir): os.makedirs(remote_logdir) os.chdir(remote_logdir) return NoopLogger(config, remote_logdir) # Logging for trials is handled centrally by TrialRunner, so # configure the remote runner to use a noop-logger. self.runner = cls.remote( config=self.config, registry=get_registry(), logger_creator=logger_creator)
def _init(self): self.local_evaluator = DQNEvaluator( self.registry, self.env_creator, self.config, self.logdir, 0) remote_cls = ray.remote( num_cpus=1, num_gpus=self.config["num_gpus_per_worker"])( DQNEvaluator) self.remote_evaluators = [ remote_cls.remote( self.registry, self.env_creator, self.config, self.logdir, i) for i in range(self.config["num_workers"])] if self.config["force_evaluators_remote"]: self.remote_evaluators = drop_colocated(self.remote_evaluators) for k in OPTIMIZER_SHARED_CONFIGS: if k not in self.config["optimizer_config"]: self.config["optimizer_config"][k] = self.config[k] self.optimizer = getattr(optimizers, self.config["optimizer_class"])( self.config["optimizer_config"], self.local_evaluator, self.remote_evaluators) self.saver = tf.train.Saver(max_to_keep=None) self.last_target_update_ts = 0 self.num_target_updates = 0
def testBasic(self): ray.init(num_cpus=4) local = _MockEvaluator() remotes = ray.remote(_MockEvaluator) remote_evaluators = [remotes.remote() for i in range(5)] test_optimizer = AsyncOptimizer( {"grads_per_step": 10}, local, remote_evaluators) test_optimizer.step() self.assertTrue(all(local.get_weights() == 0))
def test_simple_class(self): cls = ray.remote(cyth.simple_class) a1 = cls.remote() a2 = cls.remote() result1 = ray.get(a1.increment.remote()) result2 = ray.get(a2.increment.remote()) result3 = ray.get(a2.increment.remote()) self.assertEqual(result1, 1) self.assertEqual(result2, 1) self.assertEqual(result3, 2)
def example6(): """Cython simple class""" ray.init() cls = ray.remote(cyth.simple_class) a1 = cls.remote() a2 = cls.remote() result1 = ray.get(a1.increment.remote()) result2 = ray.get(a2.increment.remote()) print(result1, result2)
def test_network_driver_worker_independent(ray_start_regular): # Create a network on the driver locally. sess1 = tf.Session() loss1, init1, _, _ = make_linear_network() ray.experimental.TensorFlowVariables(loss1, sess1) sess1.run(init1) net2 = ray.remote(NetActor).remote() weights2 = ray.get(net2.get_weights.remote()) new_weights2 = ray.get( net2.set_and_get_weights.remote(net2.get_weights.remote())) assert weights2 == new_weights2
def run_func(func, *args, **kwargs): """Helper function for running examples""" ray.init() func = ray.remote(func) # NOTE: kwargs not allowed for now result = ray.get(func.remote(*args)) # Inspect the stack to get calling example caller = inspect.stack()[1][3] print("%s: %s" % (caller, str(result))) return result
def testNetworkDriverWorkerIndependent(self): ray.init(num_workers=1) # Create a network on the driver locally. sess1 = tf.Session() loss1, init1, _, _ = make_linear_network() ray.experimental.TensorFlowVariables(loss1, sess1) sess1.run(init1) net2 = ray.remote(NetActor).remote() weights2 = ray.get(net2.get_weights.remote()) new_weights2 = ray.get(net2.set_and_get_weights.remote( net2.get_weights.remote())) self.assertEqual(weights2, new_weights2)
def make( cls, evaluator_cls, evaluator_args, num_workers, optimizer_config): """Create evaluators and an optimizer instance using those evaluators. Args: evaluator_cls (class): Python class of the evaluators to create. evaluator_args (list): List of constructor args for the evaluators. num_workers (int): Number of remote evaluators to create in addition to a local evaluator. This can be zero or greater. optimizer_config (dict): Keyword arguments to pass to the optimizer class constructor. """ local_evaluator = evaluator_cls(*evaluator_args) remote_cls = ray.remote(num_cpus=1)(evaluator_cls) remote_evaluators = [ remote_cls.remote(*evaluator_args) for _ in range(num_workers)] return cls(optimizer_config, local_evaluator, remote_evaluators)
def _init(self): self.global_step = 0 self.kl_coeff = self.config["kl_coeff"] self.local_evaluator = PPOEvaluator( self.registry, self.env_creator, self.config, self.logdir, False) RemotePPOEvaluator = ray.remote( **self.config["worker_resources"])(PPOEvaluator) self.remote_evaluators = [ RemotePPOEvaluator.remote( self.registry, self.env_creator, self.config, self.logdir, True) for _ in range(self.config["num_workers"])] self.start_time = time.time() if self.config["write_logs"]: self.file_writer = tf.summary.FileWriter( self.logdir, self.local_evaluator.sess.graph) else: self.file_writer = None self.saver = tf.train.Saver(max_to_keep=None)
def _setup_runner(self, trial): cls = ray.remote( num_cpus=trial.resources.cpu, num_gpus=trial.resources.gpu)(trial._get_trainable_cls()) trial.init_logger() # We checkpoint metadata here to try mitigating logdir duplication self.try_checkpoint_metadata(trial) remote_logdir = trial.logdir def logger_creator(config): # Set the working dir in the remote process, for user file writes if not os.path.exists(remote_logdir): os.makedirs(remote_logdir) os.chdir(remote_logdir) return NoopLogger(config, remote_logdir) # Logging for trials is handled centrally by TrialRunner, so # configure the remote runner to use a noop-logger. return cls.remote(config=trial.config, logger_creator=logger_creator)
def __init__(self, env_fns, spaces=None): """ envs: list of gym environments to run in subprocesses """ self.waiting = False self.closed = False self.task_pool = TaskPool(timeout=10) nenvs = len(env_fns) self.actors = [] self.actor_to_i = {} remote_actor = ray.remote(Actor) for i in range(nenvs): actor = remote_actor.remote(i, env_fns[i]) self.actors.append(actor) self.actor_to_i[actor] = i observation_space, action_space = ray.get(self.actors[0].get_spaces.remote()) VecEnv.__init__(self, len(env_fns), observation_space, action_space) self.results = [([0] * OBSERVATION_SPACE, 0, False, {"bad": True})] * self.num_envs
def testSynchronize(self): """Synchronize applies filter buffer onto own filter""" filt1 = MeanStdFilter(()) for i in range(10): filt1(i) self.assertEqual(filt1.rs.n, 10) filt1.clear_buffer() self.assertEqual(filt1.buffer.n, 0) RemoteEvaluator = ray.remote(_MockEvaluator) remote_e = RemoteEvaluator.remote(sample_count=10) remote_e.sample.remote() FilterManager.synchronize({ "obs_filter": filt1, "rew_filter": filt1.copy() }, [remote_e]) filters = ray.get(remote_e.get_filters.remote()) obs_f = filters["obs_filter"] self.assertEqual(filt1.rs.n, 20) self.assertEqual(filt1.buffer.n, 0) self.assertEqual(obs_f.rs.n, filt1.rs.n) self.assertEqual(obs_f.buffer.n, filt1.buffer.n)
def testRemoteTrainingLoss(self): ray.init(num_workers=2) net = ray.remote(TrainActor).remote() net_values = TrainActor().values loss, variables, _, sess, grads, train, placeholders = net_values before_acc = sess.run(loss, feed_dict=dict(zip(placeholders, [[2] * 100, [4] * 100]))) for _ in range(3): gradients_list = ray.get( [net.training_step.remote(variables.get_weights()) for _ in range(2)]) mean_grads = [sum([gradients[i] for gradients in gradients_list]) / len(gradients_list) for i in range(len(gradients_list[0]))] feed_dict = {grad[0]: mean_grad for (grad, mean_grad) in zip(grads, mean_grads)} sess.run(train, feed_dict=feed_dict) after_acc = sess.run(loss, feed_dict=dict(zip(placeholders, [[2] * 100, [4] * 100]))) self.assertTrue(before_acc < after_acc)
def __init__(self, model_creator, data_creator, optimizer_creator, loss_creator, train_function=None, validation_function=None, initialization_hook=None, config=None, num_replicas=1, use_gpu=False, batch_size=16, backend="auto"): """Sets up the PyTorch trainer. Args: model_creator (dict -> torch.nn.Module): creates the model using the config. data_creator (int, dict -> DataLoader, DataLoader): Function that takes in (batch_size, config) and returns two Torch DataLoader objects. optimizer_creator (torch.nn.Module, dict -> optimizer): creates the loss and optimizer using the model and the config. loss_creator (dict -> loss): Creates the loss function/criterion using the config. train_function: Trains a model for a epoch. This takes in ( model, train_dataloader, criterion, optimizer, config), and returns a dict of training stats. validation_function: Runs validation. This takes in ( model, val_dataloader, criterion, config) and returns a dict of validation stats. config (dict): configuration passed to "model_creator", "data_creator", "optimizer_creator", and "loss_creator". num_replicas (int): the number of workers used in distributed training. use_gpu (bool): Sets resource allocation for workers to 1 GPU if true. batch_size (int): batch size for an update. backend (string): backend used by distributed PyTorch. """ # TODO: add support for mixed precision # TODO: add support for callbacks if num_replicas > 1 and not dist.is_available(): raise ValueError( ("Distributed PyTorch is not supported on macOS. " "To run without distributed PyTorch, set 'num_replicas=1'. " "For more information, see " "https://github.com/pytorch/examples/issues/467.")) self.model_creator = model_creator self.train_function = train_function self.validation_function = validation_function self.config = {} if config is None else config self.optimizer_timer = utils.TimerStat(window_size=1) if backend == "auto": backend = "nccl" if use_gpu else "gloo" logger.info("Using {} as backend.".format(backend)) if num_replicas == 1: # Generate actor class Runner = ray.remote(num_cpus=1, num_gpus=int(use_gpu))(PyTorchRunner) # Start workers self.workers = [ Runner.remote(model_creator, data_creator, optimizer_creator, loss_creator, train_function=train_function, validation_function=validation_function, config=self.config, batch_size=batch_size) ] if initialization_hook: self.apply_all_workers(initialization_hook) # Get setup tasks in order to throw errors on failure ray.get(self.workers[0].setup.remote()) else: # Generate actor class Runner = ray.remote( num_cpus=1, num_gpus=int(use_gpu))(DistributedPyTorchRunner) # Compute batch size per replica batch_size_per_replica = batch_size // num_replicas if batch_size % num_replicas > 0: new_batch_size = batch_size_per_replica * num_replicas logger.warning( ("Changing batch size from {old_batch_size} to " "{new_batch_size} to evenly distribute batches across " "{num_replicas} replicas.").format( old_batch_size=batch_size, new_batch_size=new_batch_size, num_replicas=num_replicas)) # Start workers self.workers = [ Runner.remote(model_creator, data_creator, optimizer_creator, loss_creator, backend=backend, train_function=train_function, validation_function=validation_function, config=self.config, batch_size=batch_size_per_replica) for i in range(num_replicas) ] if initialization_hook: self.apply_all_workers(initialization_hook) # Compute URL for initializing distributed PyTorch ip = ray.get(self.workers[0].get_node_ip.remote()) port = ray.get(self.workers[0].find_free_port.remote()) address = "tcp://{ip}:{port}".format(ip=ip, port=port) # Get setup tasks in order to throw errors on failure ray.get([ worker.setup.remote(address, i, len(self.workers)) for i, worker in enumerate(self.workers) ])
def __init__(self, model_creator, num_workers, devices_per_worker, gpu=True, strategy="ps", grad_shard_bytes=10000000, all_reduce_alg="simple"): if num_workers == 1 and strategy == "ps": logger.warning( "The parameter server strategy does not make sense for single " "worker operation, falling back to simple mode.") strategy = "simple" if strategy == "ps": use_plasma_op = True elif strategy == "simple": use_plasma_op = False grad_shard_bytes = 0 # tensor fusion doesn't make sense else: raise ValueError("strategy must be one of 'ps', 'simple'") self.strategy = strategy self.model_creator = model_creator if gpu: requests = {"num_gpus": devices_per_worker} else: requests = {"num_cpus": devices_per_worker} RemoteSGDWorker = ray.remote(**requests)(SGDWorker) self.workers = [] logger.info( "Creating SGD workers ({} total, {} devices per worker)".format( num_workers, devices_per_worker)) for worker_index in range(num_workers): self.workers.append( RemoteSGDWorker.remote( worker_index, model_creator, num_devices=devices_per_worker, plasma_op=use_plasma_op, gpu=gpu, max_bytes=grad_shard_bytes, all_reduce_alg=all_reduce_alg)) logger.info("Waiting for gradient configuration") shard_shapes = ray.get(self.workers[0].shard_shapes.remote()) logger.info("Waiting for actors to start") ray.get([w.shard_shapes.remote() for w in self.workers]) if strategy == "ps": logger.info("Starting parameter servers ({} shards)".format( len(shard_shapes))) self.ps_list = [ ParameterServer.remote(len(self.workers), i) for i, s in enumerate(shard_shapes) ] ray.get([ ps.initialize.remote(s) for ps, s in zip(self.ps_list, shard_shapes) ]) logger.info("Parameter servers started") else: self.ps_list = []
def testRemoteTrainingStep(self): ray.init(num_workers=1) net = ray.remote(TrainActor).remote() ray.get(net.training_step.remote(net.get_weights.remote()))
def __init__( self, *, model_creator, optimizer_creator, loss_creator=None, scheduler_creator=None, training_operator_cls=TrainingOperator, initialization_hook=None, config=None, scheduler_step_freq="batch", use_tqdm=False, backend="torch_distributed", workers_per_node=1): # todo remove ray_ctx to run on workers ray_ctx = RayContext.get() if not (isinstance(model_creator, types.FunctionType) and isinstance(optimizer_creator, types.FunctionType)): # Torch model is also callable. raise ValueError( "Must provide a function for both model_creator and optimizer_creator") self.model_creator = model_creator self.optimizer_creator = optimizer_creator self.loss_creator = loss_creator self.scheduler_creator = scheduler_creator self.training_operator_cls = training_operator_cls self.scheduler_step_freq = scheduler_step_freq self.use_tqdm = use_tqdm if not training_operator_cls and not loss_creator: raise ValueError("If a loss_creator is not provided, you must " "provide a custom training operator.") self.initialization_hook = initialization_hook self.config = {} if config is None else config worker_config = self.config.copy() params = dict( model_creator=self.model_creator, optimizer_creator=self.optimizer_creator, loss_creator=self.loss_creator, scheduler_creator=self.scheduler_creator, training_operator_cls=self.training_operator_cls, scheduler_step_freq=self.scheduler_step_freq, use_tqdm=self.use_tqdm, config=worker_config) if backend == "torch_distributed": cores_per_node = ray_ctx.ray_node_cpu_cores // workers_per_node num_nodes = ray_ctx.num_ray_nodes * workers_per_node RemoteRunner = ray.remote(num_cpus=cores_per_node)(TorchRunner) self.remote_workers = [ RemoteRunner.remote(**params) for i in range(num_nodes) ] ray.get([ worker.setup.remote(cores_per_node) for i, worker in enumerate(self.remote_workers) ]) head_worker = self.remote_workers[0] address = ray.get(head_worker.setup_address.remote()) logger.info(f"initializing pytorch process group on {address}") ray.get([ worker.setup_torch_distribute.remote(address, i, num_nodes) for i, worker in enumerate(self.remote_workers) ]) elif backend == "horovod": from zoo.orca.learn.horovod.horovod_ray_runner import HorovodRayRunner self.horovod_runner = HorovodRayRunner(ray_ctx, worker_cls=TorchRunner, worker_param=params, workers_per_node=workers_per_node) self.remote_workers = self.horovod_runner.remote_workers cores_per_node = self.horovod_runner.cores_per_node ray.get([ worker.setup.remote(cores_per_node) for i, worker in enumerate(self.remote_workers) ]) ray.get([ worker.setup_horovod.remote() for i, worker in enumerate(self.remote_workers) ]) else: raise Exception("Only \"torch_distributed\" and \"horovod\" are supported " "values of backend, but got {}".format(backend)) self.num_workers = len(self.remote_workers)
def generate_fake_x_y_data(num_data, seed=0): # Seed numpy to make the script deterministic. np.random.seed(seed) x = np.random.rand(num_data) y = x * 0.1 + 0.3 return x, y # Generate some training data. batch_ids = [generate_fake_x_y_data.remote(BATCH_SIZE, seed=i) for i in range(NUM_BATCHES)] x_ids = [x_id for x_id, y_id in batch_ids] y_ids = [y_id for x_id, y_id in batch_ids] # Generate some test data. x_test, y_test = ray.get(generate_fake_x_y_data.remote(BATCH_SIZE, seed=NUM_BATCHES)) # Create actors to store the networks. remote_network = ray.remote(Network) actor_list = [remote_network.remote(x_ids[i], y_ids[i]) for i in range(NUM_BATCHES)] local_network = Network(x_test, y_test) # Get initial weights of local network. weights = local_network.get_weights() # Do some steps of training. for iteration in range(NUM_ITERS): # Put the weights in the object store. This is optional. We could instead pass # the variable weights directly into step.remote, in which case it would be # placed in the object store under the hood. However, in that case multiple # copies of the weights would be put in the object store, so this approach is # more efficient. weights_id = ray.put(weights) # Call the remote function multiple times in parallel.
def RayFuncWrapFunc(func): return ray.remote(func)
def get_ray_result(cython_func, *args): func = ray.remote(cython_func) return ray.get(func.remote(*args))
def train(envMaker, policy, optPolicy, baseline=None, optBaselineMaker=None, saver=None, iterations: int = 100, batchSize: int = BATCH_SIZE, gamma: float = GAMMA, lmbd: float = LAMBDA, maxDKL: float = MAX_DKL, beta: float = BETA, maxEpisodeLength: int = MAX_LENGTH, pBatchMem: float = 1.0, nTests: int = TESTS, testFreq: int = TEST_FREQ, testSteps: int = MAX_LENGTH, device=DEVICE_DEFT, nWorkers=NCPUS, workerSeeds: list = [], testSeed: int = 69, logger=None, **kwargs): assert (gamma <= 1) and (gamma >= 0), "Gamma must be in the interval [0,1] " assert nWorkers > 0, "nWorkers must be greater than 0" assert batchSize > 32, "Just 'case" gae = kwargs.get("gae", False) print("Gae status", gae) # init ray if needed if nWorkers > 1: try: import ray RAY = True nWorkers = nWorkers if nWorkers <= NCPUS else NCPUS ray.init(num_cpus=nWorkers) except: RAY = False else: RAY = False # Test variables testRewardRes, testVar, testStepsRes = [], [], [] envTest = envMaker(seed=testSeed) if testSeed > 0: torch.manual_seed(testSeed) # Finishing training def closeTent(): if RAY: ray.shutdown() return (testRewardRes, testVar, testStepsRes) # Create and load the optimizers optPolicy = optPolicy(policy, **kwargs) optBaseline = optBaselineMaker( baseline.parameters()) if baseline is not None else None # Creating crawler if RAY: diffSeeds = len(workerSeeds) - (nWorkers - 1) if diffSeeds < 0: for _ in range(-diffSeeds): workerSeeds += [-1] crawler = ray.remote(Crawler) batchPerCrw = ceil(batchSize / (nWorkers - 1)) crawlers = [ crawler.remote(envMaker, policy.clone().to(DEVICE_DEFT), baseline.clone().to(DEVICE_DEFT) if baseline is not None else None, gamma, maxEpisodeLength, batchPerCrw, pBatchMem, gae=gae, lmbd=lmbd, seed=workerSeeds[i]) for i in range(nWorkers - 1) ] else: crawler = Crawler(envMaker, policy.clone(), baseline.clone() if baseline is not None else None, gamma, maxEpisodeLength, batchSize, pBatchMem, gae=gae, lmbd=lmbd, device=device, seed=workerSeeds[0]) # iterations loop bar = tqdm(range(iterations), unit="updates", desc="Training Policy") for it in bar: # Checking saver if saver is not None: saver.check() # Checking and executing test if it % testFreq == 0: meanAcc, var, meanSteps = testRun(envTest, policy, nTests=nTests, testSteps=testSteps, logger=logger) testRewardRes += [meanAcc] testVar += [var] testStepsRes += [meanSteps] bar.write( "Test Results: meanGt {:.3f}, var {:.3f} meanEpSteps {:.3f}". format(meanAcc, var, meanSteps)) if kwargs.get("desiredPerformance", False): upper = meanAcc + 0.5 * var**0.5 if upper >= kwargs["desiredPerformance"]: return closeTent() # Produce and get trajectories batches if not RAY: trajectories = [crawler.getBatch()] else: trajectories = ray.get([crw.getBatch.remote() for crw in crawlers]) # Update policy parameters s = optPolicy.updateParams(*trajectories) bar.write(s) if s is not None else None if logger is not None and s is not None: logger.logr(s) # Update baseline parameters if baseline is not None: states, returns = optPolicy.states, optPolicy.returns states.detach_().to(device) returns = returns.detach_().to(device) # Doing mini batches - Information already scrambled n = returns.shape[0] for i in range(0, n, 32): s = i + 32 s = s if s < n else n states_b, returns_b = states[i:s], returns[i:s] baseline_b = baseline.forward(states_b).squeeze() optBaseline.zero_grad() lossBaseline = F.mse_loss(baseline_b, returns_b) lossBaseline.backward() optBaseline.step() # Update crawlers if RAY: sdPi = policy.getState(cpu=True, lst=True) ray.get([cwr.updatePi.remote(sdPi) for cwr in crawlers]) if baseline is not None: sdB = baseline.getState(cpu=True, lst=True) ray.get([cwr.updateBasline.remote(sdB) for cwr in crawlers]) ray.get([cwr.clearMem.remote() for cwr in crawlers]) else: crawler.updatePi(policy.getState(lst=True)) if baseline is not None: crawler.updateBasline(baseline.getState(lst=True)) crawler.clearMem() return closeTent()
self.replay_batch_size, beta=self.prioritized_replay_beta) return MultiAgentBatch(samples, self.replay_batch_size) def update_priorities(self, prio_dict: Dict) -> None: with self.update_priorities_timer: for policy_id, (batch_indexes, td_errors) in prio_dict.items(): new_priorities = (np.abs(td_errors) + self.prioritized_replay_eps) self.replay_buffers[policy_id].update_priorities( batch_indexes, new_priorities) def stats(self, debug: bool = False) -> Dict: stat = { "add_batch_time_ms": round(1000 * self.add_batch_timer.mean, 3), "replay_time_ms": round(1000 * self.replay_timer.mean, 3), "update_priorities_time_ms": round(1000 * self.update_priorities_timer.mean, 3), } for policy_id, replay_buffer in self.replay_buffers.items(): stat.update({ "policy_{}".format(policy_id): replay_buffer.stats(debug=debug) }) return stat ReplayActor = ray.remote(num_cpus=0)(LocalReplayBuffer)
analysis = tune.run( train_convnet, name="pbt_test", scheduler=scheduler, metric="mean_accuracy", mode="max", verbose=1, stop=stopper, export_formats=[ExportFormat.MODEL], checkpoint_score_attr="mean_accuracy", keep_checkpoints_num=4, num_samples=4, config={ "lr": tune.uniform(0.001, 1), "momentum": tune.uniform(0.001, 1), }, ) # __tune_end__ if args.server_address: # If using Ray Client, we want to make sure checkpoint access # happens on the server. So we wrap `test_best_model` in a Ray task. # We have to make sure it gets executed on the same node that # ``tune.run`` is called on. from ray.util.ml_utils.node import force_on_current_node remote_fn = force_on_current_node(ray.remote(test_best_model)) ray.get(remote_fn.remote(analysis)) else: test_best_model(analysis)
def compute_gradients(self, samples): """Returns critic, actor gradients.""" return self.model.compute_gradients(samples) def apply_gradients(self, grads): """Applies gradients to evaluator weights.""" self.model.apply_gradients(grads) def compute_apply(self, samples): grads, _ = self.compute_gradients(samples) self.apply_gradients(grads) def get_weights(self): """Returns model weights.""" return self.model.get_weights() def set_weights(self, weights): """Sets model weights.""" self.model.set_weights(weights) def get_completed_rollout_metrics(self): """Returns metrics on previously completed rollouts. Calling this clears the queue of completed rollout metrics. """ return self.sampler.get_metrics() RemoteDDPGEvaluator = ray.remote(DDPGEvaluator)
def __init__(self): self._job_info_client = JobInfoStorageClient() self._log_client = JobLogStorageClient() self._supervisor_actor_cls = ray.remote(JobSupervisor) self._recover_running_jobs()
def __init__(self, env_fn: Callable[[], gym.Env]) -> None: super().__init__(env_fn) self.env = ray.remote(gym.Wrapper).options(num_cpus=0).remote(env_fn())
def get(self, trainable_cls): """Gets the wrapped trainable_cls, otherwise calls ray.remote.""" if trainable_cls not in self._cache: remote_cls = ray.remote(trainable_cls) self._cache[trainable_cls] = remote_cls return self._cache[trainable_cls]
def as_remote(cls, num_cpus=None, num_gpus=None, resources=None): return ray.remote( num_cpus=num_cpus, num_gpus=num_gpus, resources=resources)(cls)
def save(self): torch.save(self.model.state_dict(), "mnist_cnn.pt") net = Network() net.train() # __torch_net_end__ # yapf: enable # yapf: disable # __torch_ray_start__ import ray ray.init() RemoteNetwork = ray.remote(Network) # Use the below instead of `ray.remote(network)` to leverage the GPU. # RemoteNetwork = ray.remote(num_gpus=1)(Network) # __torch_ray_end__ # yapf: enable # yapf: disable # __torch_actor_start__ NetworkActor = RemoteNetwork.remote() NetworkActor2 = RemoteNetwork.remote() ray.get([NetworkActor.train.remote(), NetworkActor2.train.remote()]) # __torch_actor_end__ # yapf: enable # yapf: disable
def setup(self, config: PartialAlgorithmConfigDict): # Call super's setup to validate config, create RolloutWorkers # (train and eval), etc.. num_gpus_saved = config["num_gpus"] config["num_gpus"] = min(config["num_gpus"], 1) super().setup(config) self.config["num_gpus"] = num_gpus_saved # - Create n policy learner actors (@ray.remote-converted Policies) on # one or more GPU nodes. # - On each such node, also locate one replay buffer shard. ma_cfg = self.config["multiagent"] # By default, set max_num_policies_to_train to the number of policy IDs # provided in the multiagent config. if self.config["max_num_policies_to_train"] is None: self.config["max_num_policies_to_train"] = len( self.workers.local_worker().get_policies_to_train() ) # Single CPU replay shard (co-located with GPUs so we can place the # policies on the same machine(s)). num_gpus = ( 0.01 if (self.config["num_gpus"] and not self.config["_fake_gpus"]) else 0 ) ReplayActor = ray.remote( num_cpus=1, num_gpus=num_gpus, )(MixInMultiAgentReplayBuffer) # Setup remote replay buffer shards and policy learner actors # (located on any GPU machine in the cluster): replay_actor_args = [ self.config["replay_buffer_capacity"], self.config["replay_buffer_replay_ratio"], ] # Create a DistributedLearners utility object and set it up with # the initial first n learnable policies (found in the config). distributed_learners = DistributedLearners( config=self.config, max_num_policies_to_train=self.config["max_num_policies_to_train"], replay_actor_class=ReplayActor, replay_actor_args=replay_actor_args, ) for pid, policy_spec in ma_cfg["policies"].items(): if pid in self.workers.local_worker().get_policies_to_train(): distributed_learners.add_policy(pid, policy_spec) # Store distributed_learners on all RolloutWorkers # so they know, to which replay shard to send samples to. def _set_policy_learners(worker): worker._distributed_learners = distributed_learners ray.get( [ w.apply.remote(_set_policy_learners) for w in self.workers.remote_workers() ] ) self.distributed_learners = distributed_learners self._sampling_actor_manager = AsyncRequestsManager( self.workers.remote_workers(), max_remote_requests_in_flight_per_worker=self.config[ "max_requests_in_flight_per_sampler_worker" ], ray_wait_timeout_s=self.config["timeout_s_sampler_manager"], ) policy_actors = [policy_actor for _, policy_actor, _ in distributed_learners] self._learner_worker_manager = AsyncRequestsManager( workers=policy_actors, max_remote_requests_in_flight_per_worker=self.config[ "max_requests_in_flight_per_learner_worker" ], ray_wait_timeout_s=self.config["timeout_s_learner_manager"], )
def setup(self): self.square = ray.remote(resources={"foo": 1})(square)
def as_remote(cls, num_cpus=None, num_gpus=None): return ray.remote(num_cpus=num_cpus, num_gpus=num_gpus)(cls)
def send_dir_to_head(local_dir: str, remote_dir: str): import ray ip = ray.get(ray.remote(_get_head_ip).remote()) return send_dir_to_node(ip, local_dir, remote_dir)
async def router(serve_instance): q = ray.remote(Router).remote(serve_instance._controller) yield q ray.kill(q)
def apply_gradients(self, grads): self.policy.apply_gradients(grads) def get_weights(self): return self.policy.get_weights() def set_weights(self, params): self.policy.set_weights(params) def save(self): weights = self.get_weights() return pickle.dumps({ "weights": weights}) def restore(self, objs): objs = pickle.loads(objs) self.set_weights(objs["weights"]) def get_metrics(self): completed = [] while True: try: completed.append(self.metrics_queue.get_nowait()) except queue.Empty: break return completed RemoteBCEvaluator = ray.remote(BCEvaluator) GPURemoteBCEvaluator = ray.remote(num_gpus=1)(BCEvaluator)
def apply(self, fn: Any, remote_args: dict, blocks: Iterable[Block]) -> Iterable[ObjectRef[Block]]: map_bar = ProgressBar("Map Progress", total=len(blocks)) class BlockWorker: def ready(self): return "ok" @ray.method(num_returns=2) def process_block(self, block: Block, meta: BlockMetadata) -> (Block, BlockMetadata): new_block = fn(block) accessor = BlockAccessor.for_block(new_block) new_metadata = BlockMetadata(num_rows=accessor.num_rows(), size_bytes=accessor.size_bytes(), schema=accessor.schema(), input_files=meta.input_files) return new_block, new_metadata if not remote_args: remote_args["num_cpus"] = 1 BlockWorker = ray.remote(**remote_args)(BlockWorker) self.workers = [BlockWorker.remote()] metadata_mapping = {} tasks = {w.ready.remote(): w for w in self.workers} ready_workers = set() blocks_in = [(b, m) for (b, m) in zip(blocks, blocks.get_metadata())] blocks_out = [] while len(blocks_out) < len(blocks): ready, _ = ray.wait(list(tasks), timeout=0.01, num_returns=1, fetch_local=False) if not ready: if len(ready_workers) / len(self.workers) > 0.8: w = BlockWorker.remote() self.workers.append(w) tasks[w.ready.remote()] = w map_bar.set_description( "Map Progress ({} actors {} pending)".format( len(ready_workers), len(self.workers) - len(ready_workers))) continue [obj_id] = ready worker = tasks[obj_id] del tasks[obj_id] # Process task result. if worker in ready_workers: blocks_out.append(obj_id) map_bar.update(1) else: ready_workers.add(worker) # Schedule a new task. if blocks_in: block_ref, meta_ref = worker.process_block.remote( *blocks_in.pop()) metadata_mapping[block_ref] = meta_ref tasks[block_ref] = worker new_metadata = ray.get([metadata_mapping[b] for b in blocks_out]) map_bar.close() return BlockList(blocks_out, new_metadata)
def test_remote_training_step(ray_start_regular): net = ray.remote(TrainActor).remote() ray.get(net.training_step.remote(net.get_weights.remote()))
def setup(self): self.square = ray.remote(num_cpus=1)(square)
def run_experiments( experiments: Union[Experiment, Mapping, Sequence[Union[Experiment, Mapping]]], scheduler: Optional[TrialScheduler] = None, server_port: Optional[int] = None, verbose: Union[int, Verbosity] = Verbosity.V3_TRIAL_DETAILS, progress_reporter: Optional[ProgressReporter] = None, resume: bool = False, reuse_actors: bool = False, trial_executor: Optional[RayTrialExecutor] = None, raise_on_failed_trial: bool = True, concurrent: bool = True, # Deprecated args. queue_trials: Optional[bool] = None, callbacks: Optional[Sequence[Callback]] = None, _remote: Optional[bool] = None): """Runs and blocks until all trials finish. Examples: >>> experiment_spec = Experiment("experiment", my_func) >>> run_experiments(experiments=experiment_spec) >>> experiment_spec = {"experiment": {"run": my_func}} >>> run_experiments(experiments=experiment_spec) Returns: List of Trial objects, holding data for each executed trial. """ # To be removed in 1.9. if queue_trials is not None: raise DeprecationWarning( "`queue_trials` has been deprecated and is replaced by " "the `TUNE_MAX_PENDING_TRIALS_PG` environment variable. " "Per default at least one Trial is queued at all times, " "so you likely don't need to change anything other than " "removing this argument from your call to `tune.run()`") if _remote is None: _remote = ray.util.client.ray.is_connected() if _remote is True and trial_executor: raise ValueError("cannot use custom trial executor") if not trial_executor or isinstance(trial_executor, RayTrialExecutor): _ray_auto_init() if _remote: remote_run = ray.remote(num_cpus=0)(run_experiments) # Make sure tune.run_experiments is run on the server node. remote_run = force_on_current_node(remote_run) return ray.get( remote_run.remote(experiments, scheduler, server_port, verbose, progress_reporter, resume, reuse_actors, trial_executor, raise_on_failed_trial, concurrent, callbacks, _remote=False)) # This is important to do this here # because it schematize the experiments # and it conducts the implicit registration. experiments = convert_to_experiment_list(experiments) if concurrent: return run(experiments, server_port=server_port, verbose=verbose, progress_reporter=progress_reporter, resume=resume, reuse_actors=reuse_actors, trial_executor=trial_executor, raise_on_failed_trial=raise_on_failed_trial, scheduler=scheduler, callbacks=callbacks).trials else: trials = [] for exp in experiments: trials += run(exp, server_port=server_port, verbose=verbose, progress_reporter=progress_reporter, resume=resume, reuse_actors=reuse_actors, trial_executor=trial_executor, raise_on_failed_trial=raise_on_failed_trial, scheduler=scheduler, callbacks=callbacks).trials return trials
# This will usually run on the head node @ray.remote def _get_head_ip(): return ray.util.get_node_ip_address() ip = ray.get(_get_head_ip.remote()) remote_tune_script = "/tmp/_tune_script.py" print( f"Sending tune script to remote node {ip} ({remote_tune_script})" ) send_local_file_to_remote_file(TUNE_SCRIPT, remote_tune_script, ip) print("Starting remote cloud test using Ray client") _run_test_remote = ray.remote(resources={f"node:{ip}": 0.01}, num_cpus=0)(_run_test) result = ray.get( _run_test_remote.remote( args.variant, args.trainable, run_time, bucket, args.cpus_per_trial, remote_tune_script, )) except Exception as e: err = e result = {} if bucket: try:
def run( run_or_experiment: Union[str, Callable, Type], name: Optional[str] = None, metric: Optional[str] = None, mode: Optional[str] = None, stop: Union[None, Mapping, Stopper, Callable[[str, Mapping], bool]] = None, time_budget_s: Union[None, int, float, datetime.timedelta] = None, config: Optional[Dict[str, Any]] = None, resources_per_trial: Union[None, Mapping[str, Union[float, int, Mapping]], PlacementGroupFactory] = None, num_samples: int = 1, local_dir: Optional[str] = None, search_alg: Optional[Union[Searcher, SearchAlgorithm, str]] = None, scheduler: Optional[Union[TrialScheduler, str]] = None, keep_checkpoints_num: Optional[int] = None, checkpoint_score_attr: Optional[str] = None, checkpoint_freq: int = 0, checkpoint_at_end: bool = False, verbose: Union[int, Verbosity] = Verbosity.V3_TRIAL_DETAILS, progress_reporter: Optional[ProgressReporter] = None, log_to_file: bool = False, trial_name_creator: Optional[Callable[[Trial], str]] = None, trial_dirname_creator: Optional[Callable[[Trial], str]] = None, sync_config: Optional[SyncConfig] = None, export_formats: Optional[Sequence] = None, max_failures: int = 0, fail_fast: bool = False, restore: Optional[str] = None, server_port: Optional[int] = None, resume: bool = False, reuse_actors: bool = False, trial_executor: Optional[RayTrialExecutor] = None, raise_on_failed_trial: bool = True, callbacks: Optional[Sequence[Callback]] = None, max_concurrent_trials: Optional[int] = None, # Deprecated args queue_trials: Optional[bool] = None, loggers: Optional[Sequence[Type[Logger]]] = None, _remote: Optional[bool] = None, ) -> ExperimentAnalysis: """Executes training. When a SIGINT signal is received (e.g. through Ctrl+C), the tuning run will gracefully shut down and checkpoint the latest experiment state. Sending SIGINT again (or SIGKILL/SIGTERM instead) will skip this step. Many aspects of Tune, such as the frequency of global checkpointing, maximum pending placement group trials and the path of the result directory be configured through environment variables. Refer to :ref:`tune-env-vars` for a list of environment variables available. Examples: .. code-block:: python # Run 10 trials (each trial is one instance of a Trainable). Tune runs # in parallel and automatically determines concurrency. tune.run(trainable, num_samples=10) # Run 1 trial, stop when trial has reached 10 iterations tune.run(my_trainable, stop={"training_iteration": 10}) # automatically retry failed trials up to 3 times tune.run(my_trainable, stop={"training_iteration": 10}, max_failures=3) # Run 1 trial, search over hyperparameters, stop after 10 iterations. space = {"lr": tune.uniform(0, 1), "momentum": tune.uniform(0, 1)} tune.run(my_trainable, config=space, stop={"training_iteration": 10}) # Resumes training if a previous machine crashed tune.run(my_trainable, config=space, local_dir=<path/to/dir>, resume=True) # Rerun ONLY failed trials after an experiment is finished. tune.run(my_trainable, config=space, local_dir=<path/to/dir>, resume="ERRORED_ONLY") Args: run_or_experiment (function | class | str | :class:`Experiment`): If function|class|str, this is the algorithm or model to train. This may refer to the name of a built-on algorithm (e.g. RLLib's DQN or PPO), a user-defined trainable function or class, or the string identifier of a trainable function or class registered in the tune registry. If Experiment, then Tune will execute training based on Experiment.spec. If you want to pass in a Python lambda, you will need to first register the function: ``tune.register_trainable("lambda_id", lambda x: ...)``. You can then use ``tune.run("lambda_id")``. metric (str): Metric to optimize. This metric should be reported with `tune.report()`. If set, will be passed to the search algorithm and scheduler. mode (str): Must be one of [min, max]. Determines whether objective is minimizing or maximizing the metric attribute. If set, will be passed to the search algorithm and scheduler. name (str): Name of experiment. stop (dict | callable | :class:`Stopper`): Stopping criteria. If dict, the keys may be any field in the return result of 'train()', whichever is reached first. If function, it must take (trial_id, result) as arguments and return a boolean (True if trial should be stopped, False otherwise). This can also be a subclass of ``ray.tune.Stopper``, which allows users to implement custom experiment-wide stopping (i.e., stopping an entire Tune run based on some time constraint). time_budget_s (int|float|datetime.timedelta): Global time budget in seconds after which all trials are stopped. Can also be a ``datetime.timedelta`` object. config (dict): Algorithm-specific configuration for Tune variant generation (e.g. env, hyperparams). Defaults to empty dict. Custom search algorithms may ignore this. resources_per_trial (dict|PlacementGroupFactory): Machine resources to allocate per trial, e.g. ``{"cpu": 64, "gpu": 8}``. Note that GPUs will not be assigned unless you specify them here. Defaults to 1 CPU and 0 GPUs in ``Trainable.default_resource_request()``. This can also be a PlacementGroupFactory object wrapping arguments to create a per-trial placement group. num_samples (int): Number of times to sample from the hyperparameter space. Defaults to 1. If `grid_search` is provided as an argument, the grid will be repeated `num_samples` of times. If this is -1, (virtually) infinite samples are generated until a stopping condition is met. local_dir (str): Local dir to save training results to. Defaults to ``~/ray_results``. search_alg (Searcher|SearchAlgorithm|str): Search algorithm for optimization. You can also use the name of the algorithm. scheduler (TrialScheduler|str): Scheduler for executing the experiment. Choose among FIFO (default), MedianStopping, AsyncHyperBand, HyperBand and PopulationBasedTraining. Refer to ray.tune.schedulers for more options. You can also use the name of the scheduler. keep_checkpoints_num (int): Number of checkpoints to keep. A value of `None` keeps all checkpoints. Defaults to `None`. If set, need to provide `checkpoint_score_attr`. checkpoint_score_attr (str): Specifies by which attribute to rank the best checkpoint. Default is increasing order. If attribute starts with `min-` it will rank attribute in decreasing order, i.e. `min-validation_loss`. checkpoint_freq (int): How many training iterations between checkpoints. A value of 0 (default) disables checkpointing. This has no effect when using the Functional Training API. checkpoint_at_end (bool): Whether to checkpoint at the end of the experiment regardless of the checkpoint_freq. Default is False. This has no effect when using the Functional Training API. verbose (Union[int, Verbosity]): 0, 1, 2, or 3. Verbosity mode. 0 = silent, 1 = only status updates, 2 = status and brief trial results, 3 = status and detailed trial results. Defaults to 3. progress_reporter (ProgressReporter): Progress reporter for reporting intermediate experiment progress. Defaults to CLIReporter if running in command-line, or JupyterNotebookReporter if running in a Jupyter notebook. log_to_file (bool|str|Sequence): Log stdout and stderr to files in Tune's trial directories. If this is `False` (default), no files are written. If `true`, outputs are written to `trialdir/stdout` and `trialdir/stderr`, respectively. If this is a single string, this is interpreted as a file relative to the trialdir, to which both streams are written. If this is a Sequence (e.g. a Tuple), it has to have length 2 and the elements indicate the files to which stdout and stderr are written, respectively. trial_name_creator (Callable[[Trial], str]): Optional function for generating the trial string representation. trial_dirname_creator (Callable[[Trial], str]): Function for generating the trial dirname. This function should take in a Trial object and return a string representing the name of the directory. The return value cannot be a path. sync_config (SyncConfig): Configuration object for syncing. See tune.SyncConfig. export_formats (list): List of formats that exported at the end of the experiment. Default is None. max_failures (int): Try to recover a trial at least this many times. Ray will recover from the latest checkpoint if present. Setting to -1 will lead to infinite recovery retries. Setting to 0 will disable retries. Defaults to 0. fail_fast (bool | str): Whether to fail upon the first error. If fail_fast='raise' provided, Tune will automatically raise the exception received by the Trainable. fail_fast='raise' can easily leak resources and should be used with caution (it is best used with `ray.init(local_mode=True)`). restore (str): Path to checkpoint. Only makes sense to set if running 1 trial. Defaults to None. server_port (int): Port number for launching TuneServer. resume (str|bool): One of "LOCAL", "REMOTE", "PROMPT", "ERRORED_ONLY", or bool. LOCAL/True restores the checkpoint from the local experiment directory, determined by ``name`` and ``local_dir``. REMOTE restores the checkpoint from ``upload_dir`` (as passed to ``sync_config``). PROMPT provides CLI feedback. False forces a new experiment. ERRORED_ONLY resets and reruns ERRORED trials upon resume - previous trial artifacts will be left untouched. If resume is set but checkpoint does not exist, ValueError will be thrown. reuse_actors (bool): Whether to reuse actors between different trials when possible. This can drastically speed up experiments that start and stop actors often (e.g., PBT in time-multiplexing mode). This requires trials to have the same resource requirements. trial_executor (TrialExecutor): Manage the execution of trials. raise_on_failed_trial (bool): Raise TuneError if there exists failed trial (of ERROR state) when the experiments complete. callbacks (list): List of callbacks that will be called at different times in the training loop. Must be instances of the ``ray.tune.callback.Callback`` class. If not passed, `LoggerCallback` and `SyncerCallback` callbacks are automatically added. max_concurrent_trials (int): Maximum number of trials to run concurrently. Must be non-negative. If None or 0, no limit will be applied. This is achieved by wrapping the ``search_alg`` in a :class:`ConcurrencyLimiter`, and thus setting this argument will raise an exception if the ``search_alg`` is already a :class:`ConcurrencyLimiter`. Defaults to None. _remote (bool): Whether to run the Tune driver in a remote function. This is disabled automatically if a custom trial executor is passed in. This is enabled by default in Ray client mode. Returns: ExperimentAnalysis: Object for experiment analysis. Raises: TuneError: Any trials failed and `raise_on_failed_trial` is True. """ # To be removed in 1.9. if queue_trials is not None: raise DeprecationWarning( "`queue_trials` has been deprecated and is replaced by " "the `TUNE_MAX_PENDING_TRIALS_PG` environment variable. " "Per default at least one Trial is queued at all times, " "so you likely don't need to change anything other than " "removing this argument from your call to `tune.run()`") # NO CODE IS TO BE ADDED ABOVE THIS COMMENT # remote_run_kwargs must be defined before any other # code is ran to ensure that at this point, # `locals()` is equal to args and kwargs remote_run_kwargs = locals().copy() remote_run_kwargs.pop("_remote") if _remote is None: _remote = ray.util.client.ray.is_connected() if _remote is True and trial_executor: raise ValueError("cannot use custom trial executor") if not trial_executor or isinstance(trial_executor, RayTrialExecutor): _ray_auto_init() if _remote: remote_run = ray.remote(num_cpus=0)(run) # Make sure tune.run is called on the sever node. remote_run = force_on_current_node(remote_run) # JupyterNotebooks don't work with remote tune runs out of the box # (e.g. via Ray client) as they don't have access to the main # process stdout. So we introduce a queue here that accepts # callables, which will then be executed on the driver side. if isinstance(progress_reporter, JupyterNotebookReporter): execute_queue = Queue(actor_options={ "num_cpus": 0, **force_on_current_node(None) }) progress_reporter.set_output_queue(execute_queue) def get_next_queue_item(): try: return execute_queue.get(block=False) except Empty: return None else: # If we don't need a queue, use this dummy get fn instead of # scheduling an unneeded actor def get_next_queue_item(): return None def _handle_execute_queue(): execute_item = get_next_queue_item() while execute_item: if isinstance(execute_item, Callable): execute_item() execute_item = get_next_queue_item() remote_future = remote_run.remote(_remote=False, **remote_run_kwargs) # ray.wait(...)[1] returns futures that are not ready, yet while ray.wait([remote_future], timeout=0.2)[1]: # Check if we have items to execute _handle_execute_queue() # Handle queue one last time _handle_execute_queue() return ray.get(remote_future) del remote_run_kwargs all_start = time.time() if loggers: # Raise DeprecationWarning in 1.9, remove in 1.10/1.11 warnings.warn( "The `loggers` argument is deprecated. Please pass the respective " "`LoggerCallback` classes to the `callbacks` argument instead. " "See https://docs.ray.io/en/latest/tune/api_docs/logging.html") if mode and mode not in ["min", "max"]: raise ValueError( "The `mode` parameter passed to `tune.run()` has to be one of " "['min', 'max']") set_verbosity(verbose) config = config or {} sync_config = sync_config or SyncConfig() set_sync_periods(sync_config) if num_samples == -1: num_samples = sys.maxsize result_buffer_length = None # Create scheduler here as we need access to some of its properties if isinstance(scheduler, str): # importing at top level causes a recursive dependency from ray.tune.schedulers import create_scheduler scheduler = create_scheduler(scheduler) scheduler = scheduler or FIFOScheduler() if not scheduler.supports_buffered_results: # Result buffering with e.g. a Hyperband scheduler is a bad idea, as # hyperband tries to stop trials when processing brackets. With result # buffering, we might trigger this multiple times when evaluating # a single trial, which leads to unexpected behavior. env_result_buffer_length = os.getenv("TUNE_RESULT_BUFFER_LENGTH", "") if env_result_buffer_length: warnings.warn( f"You are using a {type(scheduler)} scheduler, but " f"TUNE_RESULT_BUFFER_LENGTH is set " f"({env_result_buffer_length}). This can lead to undesired " f"and faulty behavior, so the buffer length was forcibly set " f"to 1 instead.") result_buffer_length = 1 trial_executor = trial_executor or RayTrialExecutor( reuse_actors=reuse_actors, result_buffer_length=result_buffer_length) if isinstance(run_or_experiment, list): experiments = run_or_experiment else: experiments = [run_or_experiment] for i, exp in enumerate(experiments): if not isinstance(exp, Experiment): experiments[i] = Experiment( name=name, run=exp, stop=stop, time_budget_s=time_budget_s, config=config, resources_per_trial=resources_per_trial, num_samples=num_samples, local_dir=local_dir, sync_config=sync_config, trial_name_creator=trial_name_creator, trial_dirname_creator=trial_dirname_creator, log_to_file=log_to_file, checkpoint_freq=checkpoint_freq, checkpoint_at_end=checkpoint_at_end, keep_checkpoints_num=keep_checkpoints_num, checkpoint_score_attr=checkpoint_score_attr, export_formats=export_formats, max_failures=max_failures, restore=restore) else: logger.debug("Ignoring some parameters passed into tune.run.") if fail_fast and max_failures != 0: raise ValueError("max_failures must be 0 if fail_fast=True.") if isinstance(search_alg, str): # importing at top level causes a recursive dependency from ray.tune.suggest import create_searcher search_alg = create_searcher(search_alg) # if local_mode=True is set during ray.init(). is_local_mode = ray.worker._mode() == ray.worker.LOCAL_MODE if is_local_mode: max_concurrent_trials = 1 if not search_alg: search_alg = BasicVariantGenerator( max_concurrent=max_concurrent_trials or 0) elif max_concurrent_trials: if isinstance(search_alg, ConcurrencyLimiter): if search_alg.max_concurrent != max_concurrent_trials: raise ValueError( "You have specified `max_concurrent_trials=" f"{max_concurrent_trials}`, but the `search_alg` is " "already a `ConcurrencyLimiter` with `max_concurrent=" f"{search_alg.max_concurrent}. FIX THIS by setting " "`max_concurrent_trials=None`.") else: logger.warning( "You have specified `max_concurrent_trials=" f"{max_concurrent_trials}`, but the `search_alg` is " "already a `ConcurrencyLimiter`. `max_concurrent_trials` " "will be ignored.") else: if max_concurrent_trials < 1: raise ValueError( "`max_concurrent_trials` must be greater or equal than 1, " f"got {max_concurrent_trials}.") if isinstance(search_alg, Searcher): search_alg = ConcurrencyLimiter( search_alg, max_concurrent=max_concurrent_trials) elif not is_local_mode: logger.warning( "You have passed a `SearchGenerator` instance as the " "`search_alg`, but `max_concurrent_trials` requires a " "`Searcher` instance`. `max_concurrent_trials` " "will be ignored.") if isinstance(search_alg, Searcher): search_alg = SearchGenerator(search_alg) if config and not set_search_properties_backwards_compatible( search_alg.set_search_properties, metric, mode, config, ** experiments[0].public_spec): if has_unresolved_values(config): raise ValueError( "You passed a `config` parameter to `tune.run()` with " "unresolved parameters, but the search algorithm was already " "instantiated with a search space. Make sure that `config` " "does not contain any more parameter definitions - include " "them in the search algorithm's search space if necessary.") if not scheduler.set_search_properties(metric, mode): raise ValueError( "You passed a `metric` or `mode` argument to `tune.run()`, but " "the scheduler you are using was already instantiated with their " "own `metric` and `mode` parameters. Either remove the arguments " "from your scheduler or from your call to `tune.run()`") # Create syncer callbacks callbacks = create_default_callbacks(callbacks, sync_config, metric=metric, loggers=loggers) runner = TrialRunner( search_alg=search_alg, scheduler=scheduler, local_checkpoint_dir=experiments[0].checkpoint_dir, remote_checkpoint_dir=experiments[0].remote_checkpoint_dir, sync_config=sync_config, stopper=experiments[0].stopper, resume=resume, server_port=server_port, fail_fast=fail_fast, trial_executor=trial_executor, callbacks=callbacks, metric=metric, # Driver should only sync trial checkpoints if # checkpoints are not synced to cloud driver_sync_trial_checkpoints=not bool(sync_config.upload_dir)) if not runner.resumed: for exp in experiments: search_alg.add_configurations([exp]) else: logger.info("TrialRunner resumed, ignoring new add_experiment but " "updating trial resources.") if resources_per_trial: runner.update_pending_trial_resources(resources_per_trial) progress_reporter = progress_reporter or detect_reporter() if not progress_reporter.set_search_properties(metric, mode): raise ValueError( "You passed a `metric` or `mode` argument to `tune.run()`, but " "the reporter you are using was already instantiated with their " "own `metric` and `mode` parameters. Either remove the arguments " "from your reporter or from your call to `tune.run()`") progress_reporter.set_total_samples(search_alg.total_samples) # Calls setup on callbacks runner.setup_experiments(experiments=experiments, total_num_samples=search_alg.total_samples) # User Warning for GPUs if trial_executor.has_gpus(): if isinstance(resources_per_trial, dict) and "gpu" in resources_per_trial: # "gpu" is manually set. pass elif _check_default_resources_override(experiments[0].run_identifier): # "default_resources" is manually overridden. pass else: logger.warning("Tune detects GPUs, but no trials are using GPUs. " "To enable trials to use GPUs, set " "tune.run(resources_per_trial={'gpu': 1}...) " "which allows Tune to expose 1 GPU to each trial. " "You can also override " "`Trainable.default_resource_request` if using the " "Trainable API.") original_handler = signal.getsignal(signal.SIGINT) state = {signal.SIGINT: False} def sigint_handler(sig, frame): logger.warning( "SIGINT received (e.g. via Ctrl+C), ending Ray Tune run. " "This will try to checkpoint the experiment state one last time. " "Press CTRL+C one more time (or send SIGINT/SIGKILL/SIGTERM) " "to skip. ") state[signal.SIGINT] = True # Restore original signal handler to react to future SIGINT signals signal.signal(signal.SIGINT, original_handler) if not int(os.getenv("TUNE_DISABLE_SIGINT_HANDLER", "0")): signal.signal(signal.SIGINT, sigint_handler) tune_start = time.time() progress_reporter.set_start_time(tune_start) while not runner.is_finished() and not state[signal.SIGINT]: runner.step() if has_verbosity(Verbosity.V1_EXPERIMENT): _report_progress(runner, progress_reporter) tune_taken = time.time() - tune_start try: runner.checkpoint(force=True) except Exception as e: logger.warning(f"Trial Runner checkpointing failed: {str(e)}") if has_verbosity(Verbosity.V1_EXPERIMENT): _report_progress(runner, progress_reporter, done=True) wait_for_sync() runner.cleanup() incomplete_trials = [] for trial in runner.get_trials(): if trial.status != Trial.TERMINATED: incomplete_trials += [trial] if incomplete_trials: if raise_on_failed_trial and not state[signal.SIGINT]: raise TuneError("Trials did not complete", incomplete_trials) else: logger.error("Trials did not complete: %s", incomplete_trials) all_taken = time.time() - all_start if has_verbosity(Verbosity.V1_EXPERIMENT): logger.info(f"Total run time: {all_taken:.2f} seconds " f"({tune_taken:.2f} seconds for the tuning loop).") if state[signal.SIGINT]: logger.warning( "Experiment has been interrupted, but the most recent state was " "saved. You can continue running this experiment by passing " "`resume=True` to `tune.run()`") trials = runner.get_trials() return ExperimentAnalysis(runner.checkpoint_file, trials=trials, default_metric=metric, default_mode=mode)
async def test_task_runner_custom_method_batch(serve_instance): q = ray.remote(Router).remote() await q.setup.remote("") @serve.accept_batch class Batcher: def a(self, _): return ["a-{}".format(i) for i in range(serve.context.batch_size)] def b(self, _): return ["b-{}".format(i) for i in range(serve.context.batch_size)] def error_different_size(self, _): return [""] * (serve.context.batch_size * 2) def error_non_iterable(self, _): return 42 def return_np_array(self, _): return np.array([1] * serve.context.batch_size).astype(np.int32) CONSUMER_NAME = "runner" PRODUCER_NAME = "producer" backend_config = BackendConfig( { "max_batch_size": 4, "batch_wait_timeout": 2 }, accepts_batches=True) worker = setup_worker( CONSUMER_NAME, Batcher, backend_config=backend_config) await q.set_traffic.remote(PRODUCER_NAME, TrafficPolicy({ CONSUMER_NAME: 1.0 })) await q.set_backend_config.remote(CONSUMER_NAME, backend_config) def make_request_param(call_method): return RequestMetadata( PRODUCER_NAME, context.TaskContext.Python, call_method=call_method) a_query_param = make_request_param("a") b_query_param = make_request_param("b") futures = [q.enqueue_request.remote(a_query_param) for _ in range(2)] futures += [q.enqueue_request.remote(b_query_param) for _ in range(2)] await q.add_new_worker.remote(CONSUMER_NAME, "replica1", worker) gathered = await asyncio.gather(*futures) assert set(gathered) == {"a-0", "a-1", "b-0", "b-1"} with pytest.raises(RayServeException, match="doesn't preserve batch size"): different_size = make_request_param("error_different_size") await q.enqueue_request.remote(different_size) with pytest.raises(RayServeException, match="iterable"): non_iterable = make_request_param("error_non_iterable") await q.enqueue_request.remote(non_iterable) np_array = make_request_param("return_np_array") result_np_value = await q.enqueue_request.remote(np_array) assert isinstance(result_np_value, np.int32)
def test_gpu_ids(shutdown_only): num_gpus = 10 ray.init(num_cpus=10, num_gpus=num_gpus) def get_gpu_ids(num_gpus_per_worker): time.sleep(0.1) gpu_ids = ray.get_gpu_ids() assert len(gpu_ids) == num_gpus_per_worker assert (os.environ["CUDA_VISIBLE_DEVICES"] == ",".join( [str(i) for i in gpu_ids])) for gpu_id in gpu_ids: assert gpu_id in range(num_gpus) return gpu_ids f0 = ray.remote(num_gpus=0)(lambda: get_gpu_ids(0)) f1 = ray.remote(num_gpus=1)(lambda: get_gpu_ids(1)) f2 = ray.remote(num_gpus=2)(lambda: get_gpu_ids(2)) f4 = ray.remote(num_gpus=4)(lambda: get_gpu_ids(4)) f5 = ray.remote(num_gpus=5)(lambda: get_gpu_ids(5)) # Wait for all workers to start up. @ray.remote def f(): time.sleep(0.1) return os.getpid() start_time = time.time() while True: if len(set(ray.get([f.remote() for _ in range(10)]))) == 10: break if time.time() > start_time + 10: raise RayTestTimeoutException( "Timed out while waiting for workers to start " "up.") list_of_ids = ray.get([f0.remote() for _ in range(10)]) assert list_of_ids == 10 * [[]] list_of_ids = ray.get([f1.remote() for _ in range(10)]) set_of_ids = {tuple(gpu_ids) for gpu_ids in list_of_ids} assert set_of_ids == {(i, ) for i in range(10)} list_of_ids = ray.get([f2.remote(), f4.remote(), f4.remote()]) all_ids = [gpu_id for gpu_ids in list_of_ids for gpu_id in gpu_ids] assert set(all_ids) == set(range(10)) # There are only 10 GPUs, and each task uses 5 GPUs, so there should only # be 2 tasks scheduled at a given time. t1 = time.time() ray.get([f5.remote() for _ in range(20)]) assert time.time() - t1 >= 10 * 0.1 # Test that actors have CUDA_VISIBLE_DEVICES set properly. @ray.remote class Actor0(object): def __init__(self): gpu_ids = ray.get_gpu_ids() assert len(gpu_ids) == 0 assert (os.environ["CUDA_VISIBLE_DEVICES"] == ",".join( [str(i) for i in gpu_ids])) # Set self.x to make sure that we got here. self.x = 1 def test(self): gpu_ids = ray.get_gpu_ids() assert len(gpu_ids) == 0 assert (os.environ["CUDA_VISIBLE_DEVICES"] == ",".join( [str(i) for i in gpu_ids])) return self.x @ray.remote(num_gpus=1) class Actor1(object): def __init__(self): gpu_ids = ray.get_gpu_ids() assert len(gpu_ids) == 1 assert (os.environ["CUDA_VISIBLE_DEVICES"] == ",".join( [str(i) for i in gpu_ids])) # Set self.x to make sure that we got here. self.x = 1 def test(self): gpu_ids = ray.get_gpu_ids() assert len(gpu_ids) == 1 assert (os.environ["CUDA_VISIBLE_DEVICES"] == ",".join( [str(i) for i in gpu_ids])) return self.x a0 = Actor0.remote() ray.get(a0.test.remote()) a1 = Actor1.remote() ray.get(a1.test.remote())
def sync_filters(self, new_filters): """Changes self's filter to given and rebases any accumulated delta. Args: new_filters (dict): Filters with new state to update local copy. """ assert all(k in new_filters for k in self.filters) for k in self.filters: self.filters[k].sync(new_filters[k]) def get_filters(self, flush_after=False): """Returns a snapshot of filters. Args: flush_after (bool): Clears the filter buffer state. Returns: return_filters (dict): Dict for serializable filters """ return_filters = {} for k, f in self.filters.items(): return_filters[k] = f.as_serializable() if flush_after: f.clear_buffer() return return_filters RemoteA3CEvaluator = ray.remote(A3CEvaluator) GPURemoteA3CEvaluator = ray.remote(num_gpus=1)(A3CEvaluator)
gamma=self.config["gamma"], use_gae=False) return samples def get_completed_rollout_metrics(self): """Returns metrics on previously completed rollouts. Calling this clears the queue of completed rollout metrics. """ return self.sampler.get_metrics() def compute_gradients(self, samples): """ Returns gradient w.r.t. samples.""" gradient, info = self.policy.compute_gradients(samples) return gradient def apply_gradients(self, grads): """Applies gradients to evaluator weights.""" self.policy.apply_gradients(grads) def get_weights(self): """Returns model weights.""" return self.policy.get_weights() def set_weights(self, weights): """Sets model weights.""" return self.policy.set_weights(weights) RemotePGEvaluator = ray.remote(PGEvaluator)
def apply(self, fn: Any, remote_args: dict, blocks: BlockList) -> BlockList: context = DatasetContext.get_current() blocks_in = list(blocks.iter_blocks_with_metadata()) orig_num_blocks = len(blocks_in) results = [] map_bar = ProgressBar("Map Progress", total=orig_num_blocks) class BlockWorker: def ready(self): return "ok" def map_block_split(self, block: Block, input_files: List[str]) -> BlockPartition: return _map_block_split(block, fn, input_files) @ray.method(num_returns=2) def map_block_nosplit( self, block: Block, input_files: List[str]) -> Tuple[Block, BlockMetadata]: return _map_block_nosplit(block, fn, input_files) if not remote_args: remote_args["num_cpus"] = 1 BlockWorker = ray.remote(**remote_args)(BlockWorker) self.workers = [BlockWorker.remote()] tasks = {w.ready.remote(): w for w in self.workers} metadata_mapping = {} ready_workers = set() while len(results) < orig_num_blocks: ready, _ = ray.wait(list(tasks), timeout=0.01, num_returns=1, fetch_local=False) if not ready: if len(ready_workers) / len(self.workers) > 0.8: w = BlockWorker.remote() self.workers.append(w) tasks[w.ready.remote()] = w map_bar.set_description( "Map Progress ({} actors {} pending)".format( len(ready_workers), len(self.workers) - len(ready_workers))) continue [obj_id] = ready worker = tasks[obj_id] del tasks[obj_id] # Process task result. if worker in ready_workers: results.append(obj_id) map_bar.update(1) else: ready_workers.add(worker) # Schedule a new task. if blocks_in: block, meta = blocks_in.pop() if context.block_splitting_enabled: ref = worker.map_block_split.remote( block, meta.input_files) else: ref, meta_ref = worker.map_block_nosplit.remote( block, meta.input_files) metadata_mapping[ref] = meta_ref tasks[ref] = worker map_bar.close() new_blocks, new_metadata = [], [] if context.block_splitting_enabled: for result in ray.get(results): for block, metadata in result: new_blocks.append(block) new_metadata.append(metadata) else: for block in results: new_blocks.append(block) new_metadata.append(metadata_mapping[block]) return BlockList(new_blocks, new_metadata)
def repartition(self, num_partitions: int, batch_ms: int = 0) -> "ParallelIterator[T]": """Returns a new ParallelIterator instance with num_partitions shards. The new iterator contains the same data in this instance except with num_partitions shards. The data is split in round-robin fashion for the new ParallelIterator. Args: num_partitions (int): The number of shards to use for the new ParallelIterator batch_ms (int): Batches items for batch_ms milliseconds on each shard before retrieving it. Increasing batch_ms increases latency but improves throughput. Returns: A ParallelIterator with num_partitions number of shards and the data of this ParallelIterator split round-robin among the new number of shards. Examples: >>> it = from_range(8, 2) >>> it = it.repartition(3) >>> list(it.get_shard(0)) [0, 4, 3, 7] >>> list(it.get_shard(1)) [1, 5] >>> list(it.get_shard(2)) [2, 6] """ # initialize the local iterators for all the actors all_actors = [] for actor_set in self.actor_sets: actor_set.init_actors() all_actors.extend(actor_set.actors) def base_iterator(num_partitions, partition_index, timeout=None): futures = {} for a in all_actors: futures[a.par_iter_slice_batch.remote( step=num_partitions, start=partition_index, batch_ms=batch_ms)] = a while futures: pending = list(futures) if timeout is None: # First try to do a batch wait for efficiency. ready, _ = ray.wait( pending, num_returns=len(pending), timeout=0) # Fall back to a blocking wait. if not ready: ready, _ = ray.wait(pending, num_returns=1) else: ready, _ = ray.wait( pending, num_returns=len(pending), timeout=timeout) for obj_ref in ready: actor = futures.pop(obj_ref) try: batch = ray.get(obj_ref) futures[actor.par_iter_slice_batch.remote( step=num_partitions, start=partition_index, batch_ms=batch_ms)] = actor for item in batch: yield item except StopIteration: pass # Always yield after each round of wait with timeout. if timeout is not None: yield _NextValueNotReady() def make_gen_i(i): return lambda: base_iterator(num_partitions, i) name = self.name + f".repartition[num_partitions={num_partitions}]" generators = [make_gen_i(s) for s in range(num_partitions)] worker_cls = ray.remote(ParallelIteratorWorker) actors = [worker_cls.remote(g, repeat=False) for g in generators] # need explicit reference to self so actors in this instance do not die return ParallelIterator( [_ActorSet(actors, [])], name, parent_iterators=[self])
def test_options(): """General test of option keywords in Ray.""" import re from ray._private import ray_option_utils def f(): return 1 class A: x = 1 task_defaults = { k: v.default_value for k, v in ray_option_utils.task_options.items() } task_defaults_for_options = task_defaults.copy() task_defaults_for_options.pop("max_calls") ray.remote(f).options(**task_defaults_for_options) ray.remote(**task_defaults)(f).options(**task_defaults_for_options) with pytest.raises( ValueError, match=re.escape( "Setting 'max_calls' is not supported in '.options()'."), ): ray.remote(f).options(max_calls=1) actor_defaults = { k: v.default_value for k, v in ray_option_utils.actor_options.items() } actor_defaults_for_options = actor_defaults.copy() actor_defaults_for_options.pop("concurrency_groups") ray.remote(A).options(**actor_defaults_for_options) ray.remote(**actor_defaults)(A).options(**actor_defaults_for_options) with pytest.raises( ValueError, match=re.escape( "Setting 'concurrency_groups' is not supported in '.options()'." ), ): ray.remote(A).options(concurrency_groups=[]) unique_object = type("###", (), {})() for k, v in ray_option_utils.task_options.items(): v.validate(k, v.default_value) with pytest.raises(TypeError): v.validate(k, unique_object) for k, v in ray_option_utils.actor_options.items(): v.validate(k, v.default_value) with pytest.raises(TypeError): v.validate(k, unique_object) # test updating each namespace of "_metadata" independently assert ray_option_utils.update_options( { "_metadata": { "ns1": { "a1": 1, "b1": 2, "c1": 3 }, "ns2": { "a2": 1 } }, "num_cpus": 1, "xxx": { "x": 2 }, "zzz": 42, }, { "_metadata": { "ns1": { "b1": 22 }, "ns3": { "b3": 2 } }, "num_cpus": 2, "xxx": { "y": 2 }, "yyy": 3, }, ) == { "_metadata": { "ns1": { "a1": 1, "b1": 22, "c1": 3 }, "ns2": { "a2": 1 }, "ns3": { "b3": 2 }, }, "num_cpus": 2, "xxx": { "y": 2 }, "yyy": 3, "zzz": 42, } # test options for other Ray libraries. namespace = "namespace" class mock_options: def __init__(self, **options): self.options = {"_metadata": {namespace: options}} def keys(self): return ("_metadata", ) def __getitem__(self, key): return self.options[key] def __call__(self, f): f._default_options.update(self.options) return f @mock_options(a=1, b=2) @ray.remote(num_gpus=2) def foo(): pass assert foo._default_options == { "_metadata": { "namespace": { "a": 1, "b": 2 } }, "num_gpus": 2, } f2 = foo.options(num_cpus=1, num_gpus=1, **mock_options(a=11, c=3)) # TODO(suquark): The current implementation of `.options()` is so bad that we # cannot even access its options from outside. Here we hack the closures to # achieve our goal. Need futher efforts to clean up the tech debt. assert f2.remote.__closure__[1].cell_contents == { "_metadata": { "namespace": { "a": 11, "b": 2, "c": 3 } }, "num_cpus": 1, "num_gpus": 1, } class mock_options2(mock_options): def __init__(self, **options): self.options = {"_metadata": {namespace + "2": options}} f3 = foo.options(num_cpus=1, num_gpus=1, **mock_options2(a=11, c=3)) assert f3.remote.__closure__[1].cell_contents == { "_metadata": { "namespace": { "a": 1, "b": 2 }, "namespace2": { "a": 11, "c": 3 } }, "num_cpus": 1, "num_gpus": 1, } with pytest.raises(TypeError): # Ensure only a single "**option" per ".options()". # Otherwise it would be confusing. foo.options( num_cpus=1, num_gpus=1, **mock_options(a=11, c=3), **mock_options2(a=11, c=3), )
def _build_model(self): # To look clearly, whether I use bias. use_bias = self._config.use_bias self.observ = tf.placeholder(tf.float32, (None, 4), name='observ') self.target = tf.placeholder(tf.float32, name='target') x = tf.layers.dense(self.observ, 1, use_bias=use_bias, kernel_initializer=tf.zeros_initializer) self.logits = x def _set_loss(self): losses = tf.losses.mean_squared_error(labels=self.target, predictions=self.logits) self.loss = tf.reduce_mean(losses) def predict(self, observ): baseline = self._sess.run([self.logits], feed_dict={self.observ: observ}) return baseline def apply(self, observ, target): _, loss = self._sess.run([self.train_op, self.loss], feed_dict={self.observ: observ, self.target: target}) return loss def get_weights(self): return self.variables.get_weights() # Remote actor(Policy) function enable be called in distribution. RemotePolicy = ray.remote(Policy)
def __init__(self): self._status_client = JobStatusStorageClient() self._log_client = JobLogStorageClient() self._supervisor_actor_cls = ray.remote(JobSupervisor)
def main(args=None, model=None) -> GenerativeQAModule: parser = argparse.ArgumentParser() parser = pl.Trainer.add_argparse_args(parser) parser = GenerativeQAModule.add_model_specific_args(parser, os.getcwd()) parser = GenerativeQAModule.add_retriever_specific_args(parser) args = args or parser.parse_args() Path(args.output_dir).mkdir(exist_ok=True) named_actors = [] if args.distributed_retriever == "ray" and args.gpus > 1: if not is_ray_available(): raise RuntimeError("Please install Ray to use the Ray " "distributed retriever.") # Connect to an existing Ray cluster. try: ray.init(address=args.ray_address) except (ConnectionError, ValueError): logger.warning( "Connection to Ray cluster failed. Make sure a Ray" "cluster is running by either using Ray's cluster " "launcher (`ray up`) or by manually starting Ray on " "each node via `ray start --head` for the head node " "and `ray start --address='<ip address>:6379'` for " "additional nodes. See " "https://docs.ray.io/en/master/cluster/index.html " "for more info.") raise # Create Ray actors only for rank 0. if ("LOCAL_RANK" not in os.environ or int(os.environ["LOCAL_RANK"]) == 0) and ("NODE_RANK" not in os.environ or int(os.environ["NODE_RANK"]) == 0): remote_cls = ray.remote(RayRetriever) named_actors = [ remote_cls.options( name="retrieval_worker_{}".format(i)).remote() for i in range(args.num_retrieval_workers) ] else: logger.info( "Getting named actors for NODE_RANK {}, LOCAL_RANK {}".format( os.environ["NODE_RANK"], os.environ["LOCAL_RANK"])) named_actors = [ ray.get_actor("retrieval_worker_{}".format(i)) for i in range(args.num_retrieval_workers) ] args.actor_handles = named_actors assert args.actor_handles == named_actors if model is None: model: GenerativeQAModule = GenerativeQAModule(args) dataset = Path(args.data_dir).name if (args.logger_name == "default" or args.fast_dev_run or str(args.output_dir).startswith("/tmp") or str(args.output_dir).startswith("/var")): training_logger = True # don't pollute wandb logs unnecessarily elif args.logger_name == "wandb": from pytorch_lightning.loggers import WandbLogger project = os.environ.get("WANDB_PROJECT", dataset) training_logger = WandbLogger(name=model.output_dir.name, project=project) elif args.logger_name == "wandb_shared": from pytorch_lightning.loggers import WandbLogger training_logger = WandbLogger(name=model.output_dir.name, project=f"hf_{dataset}") es_callback = (get_early_stopping_callback(model.val_metric, args.early_stopping_patience) if args.early_stopping_patience >= 0 else False) trainer: pl.Trainer = generic_train( model, args, logging_callback=Seq2SeqLoggingCallback(), checkpoint_callback=get_checkpoint_callback(args.output_dir, model.val_metric), early_stopping_callback=es_callback, logger=training_logger, custom_ddp_plugin=CustomDDP() if args.gpus > 1 else None, profiler=pl.profiler.AdvancedProfiler() if args.profile else None, ) pickle_save(model.hparams, model.output_dir / "hparams.pkl") if not args.do_predict: return model # test() without a model tests using the best checkpoint automatically trainer.test() return model
def setup(self): self.square = ray.remote(square)