def _finalize_experiment( experiment_json, metric, app_id, run_id, state, duration, logdir, best_logdir, optimization_key, ): """Attaches the experiment outcome as xattr metadata to the app directory. """ outputs = _build_summary_json(logdir) if outputs: hopshdfs.dump(outputs, logdir + "/.summary.json") if best_logdir: experiment_json["bestDir"] = best_logdir[len(hopshdfs.project_path()):] experiment_json["optimizationKey"] = optimization_key experiment_json["metric"] = metric experiment_json["state"] = state experiment_json["duration"] = duration experiment_utils._attach_experiment_xattr(app_id, run_id, experiment_json, "REPLACE")
def _exit_handler(): """ Handles jobs killed by the user. """ try: global running global experiment_json if running and experiment_json is not None: experiment_json["status"] = "KILLED" experiment_utils._attach_experiment_xattr( app_id, run_id, experiment_json, "REPLACE" ) except Exception as err: util._log(err)
def _exit_handler(): """ Returns: """ try: global running global experiment_json if running and experiment_json != None: experiment_json['state'] = "KILLED" exp_ml_id = app_id + "_" + str(run_id) experiment_utils._attach_experiment_xattr(exp_ml_id, experiment_json, 'FULL_UPDATE') except Exception as err: print(err) pass
def _exit_handler(): """ Returns: """ try: global running global experiment_json if running and experiment_json != None: experiment_json['state'] = "KILLED" experiment_utils._attach_experiment_xattr(app_id, run_id, experiment_json, 'REPLACE') except Exception as err: print(err) pass
def _exception_handler(duration): """ Handles exceptions during execution of an experiment :param duration: duration of the experiment until exception in milliseconds :type duration: int """ try: global running global experiment_json if running and experiment_json is not None: experiment_json["state"] = "FAILED" experiment_json["duration"] = duration experiment_utils._attach_experiment_xattr( app_id, run_id, experiment_json, "REPLACE" ) except Exception as err: util._log(err)
def mirrored(train_fn, name='no-name', local_logdir=False, description=None, evaluator=False, metric_key=None): """ *Distributed Training* Example usage: >>> from hops import experiment >>> def mirrored_training(): >>> # Do all imports in the function >>> import tensorflow >>> # Put all code inside the train_fn function >>> from hops import tensorboard >>> from hops import devices >>> logdir = tensorboard.logdir() >>> ...MirroredStrategy()... >>> experiment.mirrored(mirrored_training, local_logdir=True) Args: :train_fn: contains the code where you are using MirroredStrategy. :name: name of the experiment :local_logdir: True if *tensorboard.logdir()* should be in the local filesystem, otherwise it is in HDFS :description: a longer description for the experiment :evaluator: whether to run one of the workers as an evaluator :metric_key: If returning a dict with multiple return values, this key should match the name of the key in the dict for the metric you want to associate with the experiment Returns: HDFS path in your project where the experiment is stored and return value from the process running as chief """ num_ps = util.num_param_servers() assert num_ps == 0, "number of parameter servers should be 0" global running if running: raise RuntimeError("An experiment is currently running.") num_workers = util.num_executors() if evaluator: assert num_workers > 2, "number of workers must be atleast 3 if evaluator is set to True" start = time.time() sc = util._find_spark().sparkContext try: global app_id global experiment_json global run_id app_id = str(sc.applicationId) _start_run() experiment_utils._create_experiment_dir(app_id, run_id) experiment_json = experiment_utils._populate_experiment( name, 'mirrored', 'DISTRIBUTED_TRAINING', None, description, app_id, None, None) experiment_json = experiment_utils._attach_experiment_xattr( app_id, run_id, experiment_json, 'CREATE') logdir, return_dict = mirrored_impl._run(sc, train_fn, run_id, local_logdir=local_logdir, name=name, evaluator=evaluator) duration = experiment_utils._seconds_to_milliseconds(time.time() - start) metric = experiment_utils._get_metric(return_dict, metric_key) experiment_utils._finalize_experiment(experiment_json, metric, app_id, run_id, 'FINISHED', duration, logdir, None, None) return logdir, return_dict except: _exception_handler( experiment_utils._seconds_to_milliseconds(time.time() - start)) raise finally: _end_run(sc)
def grid_search(train_fn, grid_dict, direction=Direction.MAX, name='no-name', local_logdir=False, description=None, optimization_key='metric'): """ *Parallel Experiment* Run grid search evolution to explore a predefined set of hyperparameter combinations. The function is treated as a blackbox that returns a metric for some given hyperparameter combination. The returned metric is used to evaluate how 'good' the hyperparameter combination was. Example usage: >>> from hops import experiment >>> grid_dict = {'learning_rate': [0.1, 0.3], 'layers': [2, 9], 'dropout': [0.1,0.9]} >>> def train_nn(learning_rate, layers, dropout): >>> # Do all imports in the function >>> import tensorflow >>> # Put all code inside the train_fn function >>> return network.evaluate(learning_rate, layers, dropout) >>> experiment.grid_search(train_nn, grid_dict, direction=Direction.MAX) Returning multiple outputs, including images and logs: >>> from hops import experiment >>> grid_dict = {'learning_rate': [0.1, 0.3], 'layers': [2, 9], 'dropout': [0.1,0.9]} >>> def train_nn(learning_rate, layers, dropout): >>> # Do all imports in the function >>> import tensorflow >>> # Put all code inside the train_fn function >>> from PIL import Image >>> f = open('logfile.txt', 'w') >>> f.write('Starting training...') >>> accuracy, loss = network.evaluate(learning_rate, layers, dropout) >>> img = Image.new(.....) >>> img.save('diagram.png') >>> return {'accuracy': accuracy, 'loss': loss, 'logfile': 'logfile.txt', 'diagram': 'diagram.png'} >>> # Important! Remember: optimization_key must be set when returning multiple outputs >>> experiment.grid_search(train_nn, grid_dict, direction=Direction.MAX, optimization_key='accuracy') Args: :train_fn: the function to run, must return a metric :grid_dict: a dict with a key for each argument with a corresponding value being a list containing the hyperparameters to test, internally all possible combinations will be generated and run as separate Experiments :direction: Direction.MAX to maximize the returned metric, Direction.MIN to minize the returned metric :name: name of the experiment :local_logdir: True if *tensorboard.logdir()* should be in the local filesystem, otherwise it is in HDFS :description: a longer description for the experiment :optimization_key: When returning a dict, the key name of the metric to maximize or minimize in the dict should be set as this value Returns: HDFS path in your project where the experiment is stored, dict with best hyperparameters and return dict with best metrics """ num_ps = util.num_param_servers() assert num_ps == 0, "number of parameter servers should be 0" global running if running: raise RuntimeError("An experiment is currently running.") start = time.time() sc = util._find_spark().sparkContext try: global app_id global experiment_json global run_id app_id = str(sc.applicationId) _start_run() experiment_utils._create_experiment_dir(app_id, run_id) experiment_json = experiment_utils._populate_experiment( name, 'grid_search', 'PARALLEL_EXPERIMENTS', json.dumps(grid_dict), description, app_id, direction, optimization_key) experiment_json = experiment_utils._attach_experiment_xattr( app_id, run_id, experiment_json, 'CREATE') grid_params = experiment_utils.grid_params(grid_dict) logdir, best_param, best_metric, return_dict = grid_search_impl._run( sc, train_fn, run_id, grid_params, direction=direction, local_logdir=local_logdir, name=name, optimization_key=optimization_key) duration = experiment_utils._seconds_to_milliseconds(time.time() - start) experiment_utils._finalize_experiment( experiment_json, best_metric, app_id, run_id, 'FINISHED', duration, experiment_utils._get_logdir(app_id, run_id), logdir, optimization_key) return logdir, best_param, return_dict except: _exception_handler( experiment_utils._seconds_to_milliseconds(time.time() - start)) raise finally: _end_run(sc)
def launch(train_fn, args_dict=None, name='no-name', local_logdir=False, description=None, metric_key=None): """ *Experiment* or *Parallel Experiment* Run an Experiment contained in *train_fn* one time with no arguments or multiple times with different arguments if *args_dict* is specified. Example usage: >>> from hops import experiment >>> def train_nn(): >>> # Do all imports in the function >>> import tensorflow >>> # Put all code inside the train_fn function >>> accuracy, loss = network.evaluate(learning_rate, layers, dropout) >>> experiment.launch(train_nn) Returning multiple outputs, including images and logs: >>> from hops import experiment >>> def train_nn(): >>> # Do all imports in the function >>> import tensorflow >>> # Put all code inside the train_fn function >>> from PIL import Image >>> f = open('logfile.txt', 'w') >>> f.write('Starting training...') >>> accuracy, loss = network.evaluate(learning_rate, layers, dropout) >>> img = Image.new(.....) >>> img.save('diagram.png') >>> return {'accuracy': accuracy, 'loss': loss, 'logfile': 'logfile.txt', 'diagram': 'diagram.png'} >>> experiment.launch(train_nn) Args: :train_fn: The function to run :args_dict: If specified will run the same function multiple times with different arguments, {'a':[1,2], 'b':[5,3]} would run the function two times with arguments (1,5) and (2,3) provided that the function signature contains two arguments like *def func(a,b):* :name: name of the experiment :local_logdir: True if *tensorboard.logdir()* should be in the local filesystem, otherwise it is in HDFS :description: A longer description for the experiment :metric_key: If returning a dict with multiple return values, this key should match the name of the key in the dict for the metric you want to associate with the experiment Returns: HDFS path in your project where the experiment is stored """ num_ps = util.num_param_servers() assert num_ps == 0, "number of parameter servers should be 0" global running if running: raise RuntimeError( "An experiment is currently running. Please call experiment.end() to stop it." ) start = time.time() sc = util._find_spark().sparkContext try: global app_id global experiment_json global run_id app_id = str(sc.applicationId) _start_run() experiment_utils._create_experiment_dir(app_id, run_id) experiment_json = None if args_dict: experiment_json = experiment_utils._populate_experiment( name, 'launch', 'EXPERIMENT', json.dumps(args_dict), description, app_id, None, None) else: experiment_json = experiment_utils._populate_experiment( name, 'launch', 'EXPERIMENT', None, description, app_id, None, None) experiment_json = experiment_utils._attach_experiment_xattr( app_id, run_id, experiment_json, 'CREATE') logdir, return_dict = launcher._run(sc, train_fn, run_id, args_dict, local_logdir) duration = experiment_utils._seconds_to_milliseconds(time.time() - start) metric = experiment_utils._get_metric(return_dict, metric_key) experiment_utils._finalize_experiment(experiment_json, metric, app_id, run_id, 'FINISHED', duration, logdir, None, None) return logdir, return_dict except: _exception_handler( experiment_utils._seconds_to_milliseconds(time.time() - start)) raise finally: _end_run(sc)
def random_search(train_fn, boundary_dict, direction=Direction.MAX, samples=10, name='no-name', local_logdir=False, description=None, optimization_key='metric'): """ *Parallel Experiment* Run an Experiment contained in *train_fn* for configured number of random samples controlled by the *samples* parameter. Each hyperparameter is contained in *boundary_dict* with the key corresponding to the name of the hyperparameter and a list containing two elements defining the lower and upper bound. The experiment must return a metric corresponding to how 'good' the given hyperparameter combination is. Example usage: >>> from hops import experiment >>> boundary_dict = {'learning_rate': [0.1, 0.3], 'layers': [2, 9], 'dropout': [0.1,0.9]} >>> def train_nn(learning_rate, layers, dropout): >>> # Do all imports in the function >>> import tensorflow >>> # Put all code inside the train_fn function >>> return network.evaluate(learning_rate, layers, dropout) >>> experiment.differential_evolution(train_nn, boundary_dict, direction='max') Returning multiple outputs, including images and logs: >>> from hops import experiment >>> boundary_dict = {'learning_rate': [0.1, 0.3], 'layers': [2, 9], 'dropout': [0.1,0.9]} >>> def train_nn(learning_rate, layers, dropout): >>> # Do all imports in the function >>> import tensorflow >>> # Put all code inside the train_fn function >>> from PIL import Image >>> f = open('logfile.txt', 'w') >>> f.write('Starting training...') >>> accuracy, loss = network.evaluate(learning_rate, layers, dropout) >>> img = Image.new(.....) >>> img.save('diagram.png') >>> return {'accuracy': accuracy, 'loss': loss, 'logfile': 'logfile.txt', 'diagram': 'diagram.png'} >>> # Important! Remember: optimization_key must be set when returning multiple outputs >>> experiment.differential_evolution(train_nn, boundary_dict, direction='max', optimization_key='accuracy') Args: :train_fn: The function to run :boundary_dict: dict containing hyperparameter name and corresponding boundaries, each experiment randomize a value in the boundary range. :direction: Direction.MAX to maximize the returned metric, Direction.MIN to minize the returned metric :samples: the number of random samples to evaluate for each hyperparameter given the boundaries, for example samples=3 would result in 3 hyperparameter combinations in total to evaluate :name: name of the experiment :local_logdir: True if *tensorboard.logdir()* should be in the local filesystem, otherwise it is in HDFS :description: A longer description for the experiment :optimization_key: When returning a dict, the key name of the metric to maximize or minimize in the dict should be set as this value Returns: HDFS path in your project where the experiment is stored, dict with best hyperparameters and return dict with best metrics """ num_ps = util.num_param_servers() assert num_ps == 0, "number of parameter servers should be 0" global running if running: raise RuntimeError("An experiment is currently running.") start = time.time() sc = util._find_spark().sparkContext try: global app_id global experiment_json global run_id app_id = str(sc.applicationId) _start_run() experiment_utils._create_experiment_dir(app_id, run_id) experiment_json = experiment_utils._populate_experiment( name, 'random_search', 'PARALLEL_EXPERIMENTS', json.dumps(boundary_dict), description, app_id, direction, optimization_key) experiment_json = experiment_utils._attach_experiment_xattr( app_id, run_id, experiment_json, 'CREATE') logdir, best_param, best_metric, return_dict = r_search_impl._run( sc, train_fn, run_id, boundary_dict, samples, direction=direction, local_logdir=local_logdir, optimization_key=optimization_key) duration = experiment_utils._seconds_to_milliseconds(time.time() - start) experiment_utils._finalize_experiment( experiment_json, best_metric, app_id, run_id, 'FINISHED', duration, experiment_utils._get_logdir(app_id, run_id), logdir, optimization_key) return logdir, best_param, return_dict except: _exception_handler( experiment_utils._seconds_to_milliseconds(time.time() - start)) raise finally: _end_run(sc)
def parameter_server(map_fun, name='no-name', local_logdir=False, description=None, evaluator=False): """ *Distributed Training* Sets up the cluster to run ParameterServerStrategy. TF_CONFIG is exported in the background and does not need to be set by the user themselves. Example usage: >>> from hops import experiment >>> def distributed_training(): >>> # Do all imports in the function >>> import tensorflow >>> # Put all code inside the wrapper function >>> from hops import tensorboard >>> from hops import devices >>> logdir = tensorboard.logdir() >>> ...ParameterServerStrategy(num_gpus_per_worker=devices.get_num_gpus())... >>> experiment.parameter_server(distributed_training, local_logdir=True) Args:f :map_fun: contains the code where you are using ParameterServerStrategy. :name: name of the experiment :local_logdir: True if *tensorboard.logdir()* should be in the local filesystem, otherwise it is in HDFS :description: a longer description for the experiment :evaluator: whether to run one of the workers as an evaluator Returns: HDFS path in your project where the experiment is stored and return value from the process running as chief """ num_ps = util.num_param_servers() num_executors = util.num_executors() assert num_ps > 0, "number of parameter servers should be greater than 0" assert num_ps < num_executors, "num_ps cannot be greater than num_executors (i.e. num_executors == num_ps + num_workers)" if evaluator: assert num_executors - num_ps > 2, "number of workers must be atleast 3 if evaluator is set to True" global running if running: raise RuntimeError("An experiment is currently running.") start = time.time() sc = util._find_spark().sparkContext try: global app_id global experiment_json global run_id app_id = str(sc.applicationId) _start_run() hdfs.mkdir(experiment_utils._get_logdir(app_id, run_id)) experiment_json = experiment_utils._populate_experiment( name, 'parameter_server', 'DISTRIBUTED_TRAINING', None, description, app_id, None, None) experiment_json = experiment_json = experiment_utils._attach_experiment_xattr( app_id, run_id, experiment_json, 'CREATE') logdir, return_dict = ps_impl._run(sc, map_fun, run_id, local_logdir=local_logdir, name=name, evaluator=evaluator) duration = experiment_utils._seconds_to_milliseconds(time.time() - start) experiment_utils._finalize_experiment(experiment_json, None, app_id, run_id, 'FINISHED', duration, logdir, None, None) return logdir, return_dict except: _exception_handler( experiment_utils._seconds_to_milliseconds(time.time() - start)) raise finally: _end_run(sc)
def differential_evolution(objective_function, boundary_dict, direction=Direction.MAX, generations=4, population=6, mutation=0.5, crossover=0.7, name='no-name', local_logdir=False, description=None, optimization_key='metric'): """ *Parallel Experiment* Run differential evolution to explore a given search space for each hyperparameter and figure out the best hyperparameter combination. The function is treated as a blackbox that returns a metric for some given hyperparameter combination. The returned metric is used to evaluate how 'good' the hyperparameter combination was. Example usage: >>> from hops import experiment >>> boundary_dict = {'learning_rate': [0.1, 0.3], 'layers': [2, 9], 'dropout': [0.1,0.9]} >>> def train_nn(learning_rate, layers, dropout): >>> import tensorflow >>> return network.evaluate(learning_rate, layers, dropout) >>> experiment.differential_evolution(train_nn, boundary_dict, direction=Direction.MAX) Returning multiple outputs, including images and logs: >>> from hops import experiment >>> boundary_dict = {'learning_rate': [0.1, 0.3], 'layers': [2, 9], 'dropout': [0.1,0.9]} >>> def train_nn(learning_rate, layers, dropout): >>> # Do all imports in the function >>> import tensorflow >>> # Put all code inside the wrapper function >>> from PIL import Image >>> f = open('logfile.txt', 'w') >>> f.write('Starting training...') >>> accuracy, loss = network.evaluate(learning_rate, layers, dropout) >>> img = Image.new(.....) >>> img.save('diagram.png') >>> return {'accuracy': accuracy, 'loss': loss, 'logfile': 'logfile.txt', 'diagram': 'diagram.png'} >>> # Important! Remember: optimization_key must be set when returning multiple outputs >>> experiment.differential_evolution(train_nn, boundary_dict, direction=Direction.MAX, optimization_key='accuracy') Args: :objective_function: the function to run, must return a metric :boundary_dict: a dict where each key corresponds to an argument of *objective_function* and the correspond value should be a list of two elements. The first element being the lower bound for the parameter and the the second element the upper bound. :direction: Direction.MAX to maximize the returned metric, Direction.MIN to minize the returned metric :generations: number of generations :population: size of population :mutation: mutation rate to explore more different hyperparameters :crossover: how fast to adapt the population to the best in each generation :name: name of the experiment :local_logdir: True if *tensorboard.logdir()* should be in the local filesystem, otherwise it is in HDFS :description: a longer description for the experiment :optimization_key: When returning a dict, the key name of the metric to maximize or minimize in the dict should be set as this value Returns: HDFS path in your project where the experiment is stored, dict with best hyperparameters and return dict with best metrics """ num_ps = util.num_param_servers() assert num_ps == 0, "number of parameter servers should be 0" global running if running: raise RuntimeError("An experiment is currently running.") start = time.time() sc = util._find_spark().sparkContext try: global app_id global experiment_json global run_id app_id = str(sc.applicationId) _start_run() diff_evo_impl.run_id = run_id hdfs.mkdir(experiment_utils._get_logdir(app_id, run_id)) experiment_json = experiment_utils._populate_experiment( name, 'differential_evolution', 'PARALLEL_EXPERIMENTS', json.dumps(boundary_dict), description, app_id, direction, optimization_key) experiment_json = experiment_utils._attach_experiment_xattr( app_id, run_id, experiment_json, 'CREATE') logdir, best_param, best_metric, return_dict = diff_evo_impl._run( objective_function, boundary_dict, direction=direction, generations=generations, population=population, mutation=mutation, crossover=crossover, cleanup_generations=False, local_logdir=local_logdir, name=name, optimization_key=optimization_key) duration = experiment_utils._seconds_to_milliseconds(time.time() - start) experiment_utils._finalize_experiment( experiment_json, best_metric, app_id, run_id, 'FINISHED', duration, experiment_utils._get_logdir(app_id, run_id), logdir, optimization_key) return logdir, best_param, return_dict except: _exception_handler( experiment_utils._seconds_to_milliseconds(time.time() - start)) raise finally: _end_run(sc)
def attach_experiment_xattr(self, ml_id, json_data, op_type): return experiment_utils._attach_experiment_xattr(ml_id, json_data, op_type)
def lagom( map_fun, name="no-name", experiment_type="optimization", searchspace=None, optimizer=None, direction="max", num_trials=1, ablation_study=None, ablator=None, optimization_key="metric", hb_interval=1, es_policy="median", es_interval=300, es_min=10, description="", ): """Launches a maggy experiment, which depending on `experiment_type` can either be a hyperparameter optimization or an ablation study experiment. Given a search space, objective and a model training procedure `map_fun` (black-box function), an experiment is the whole process of finding the best hyperparameter combination in the search space, optimizing the black-box function. Currently maggy supports random search and a median stopping rule. **lagom** is a Swedish word meaning "just the right amount". :param map_fun: User defined experiment containing the model training. :type map_fun: function :param name: A user defined experiment identifier. :type name: str :param experiment_type: Type of Maggy experiment, either 'optimization' (default) or 'ablation'. :type experiment_type: str :param searchspace: A maggy Searchspace object from which samples are drawn. :type searchspace: Searchspace :param optimizer: The optimizer is the part generating new trials. :type optimizer: str, AbstractOptimizer :param direction: If set to ‘max’ the highest value returned will correspond to the best solution, if set to ‘min’ the opposite is true. :type direction: str :param num_trials: the number of trials to evaluate given the search space, each containing a different hyperparameter combination :type num_trials: int :param ablation_study: Ablation study object. Can be None for optimization experiment type. :type ablation_study: AblationStudy :param ablator: Ablator to use for experiment type 'ablation'. :type ablator: str, AbstractAblator :param optimization_key: Name of the metric to be optimized :type optimization_key: str, optional :param hb_interval: The heartbeat interval in seconds from trial executor to experiment driver, defaults to 1 :type hb_interval: int, optional :param es_policy: The earlystopping policy, defaults to 'median' :type es_policy: str, optional :param es_interval: Frequency interval in seconds to check currently running trials for early stopping, defaults to 300 :type es_interval: int, optional :param es_min: Minimum number of trials finalized before checking for early stopping, defaults to 10 :type es_min: int, optional :param description: A longer description of the experiment. :type description: str, optional :raises RuntimeError: An experiment is currently running. :return: A dictionary indicating the best trial and best hyperparameter combination with it's performance metric :rtype: dict """ global running if running: raise RuntimeError("An experiment is currently running.") job_start = time.time() sc = hopsutil._find_spark().sparkContext exp_driver = None try: global app_id global experiment_json global run_id app_id = str(sc.applicationId) app_id, run_id = util._validate_ml_id(app_id, run_id) # start run running = True experiment_utils._set_ml_id(app_id, run_id) # create experiment dir experiment_utils._create_experiment_dir(app_id, run_id) tensorboard._register(experiment_utils._get_logdir(app_id, run_id)) num_executors = util.num_executors(sc) # start experiment driver if experiment_type == "optimization": assert num_trials > 0, "number of trials should be greater " + "than zero" tensorboard._write_hparams_config( experiment_utils._get_logdir(app_id, run_id), searchspace ) if num_executors > num_trials: num_executors = num_trials exp_driver = experimentdriver.ExperimentDriver( "optimization", searchspace=searchspace, optimizer=optimizer, direction=direction, num_trials=num_trials, name=name, num_executors=num_executors, hb_interval=hb_interval, es_policy=es_policy, es_interval=es_interval, es_min=es_min, description=description, log_dir=experiment_utils._get_logdir(app_id, run_id), ) exp_function = exp_driver.optimizer.name() elif experiment_type == "ablation": exp_driver = experimentdriver.ExperimentDriver( "ablation", ablation_study=ablation_study, ablator=ablator, name=name, num_executors=num_executors, hb_interval=hb_interval, description=description, log_dir=experiment_utils._get_logdir(app_id, run_id), ) # using exp_driver.num_executor since # it has been set using ablator.get_number_of_trials() # in experiment.py if num_executors > exp_driver.num_executors: num_executors = exp_driver.num_executors exp_function = exp_driver.ablator.name() else: running = False raise RuntimeError( "Unknown experiment_type:" "should be either 'optimization' or 'ablation', " "But it is '{0}'".format(str(experiment_type)) ) nodeRDD = sc.parallelize(range(num_executors), num_executors) # Do provenance after initializing exp_driver, because exp_driver does # the type checks for optimizer and searchspace sc.setJobGroup(os.environ["ML_ID"], "{0} | {1}".format(name, exp_function)) experiment_json = experiment_utils._populate_experiment( name, exp_function, "MAGGY", exp_driver.searchspace.json(), description, app_id, direction, optimization_key, ) experiment_json = experiment_utils._attach_experiment_xattr( app_id, run_id, experiment_json, "CREATE" ) util._log( "Started Maggy Experiment: {0}, {1}, run {2}".format(name, app_id, run_id) ) exp_driver.init(job_start) server_addr = exp_driver.server_addr # Force execution on executor, since GPU is located on executor nodeRDD.foreachPartition( trialexecutor._prepare_func( app_id, run_id, experiment_type, map_fun, server_addr, hb_interval, exp_driver._secret, optimization_key, experiment_utils._get_logdir(app_id, run_id), ) ) job_end = time.time() result = exp_driver.finalize(job_end) best_logdir = ( experiment_utils._get_logdir(app_id, run_id) + "/" + result["best_id"] ) util._finalize_experiment( experiment_json, float(result["best_val"]), app_id, run_id, "FINISHED", exp_driver.duration, experiment_utils._get_logdir(app_id, run_id), best_logdir, optimization_key, ) util._log("Finished Experiment") return result except: # noqa: E722 _exception_handler( experiment_utils._seconds_to_milliseconds(time.time() - job_start) ) if exp_driver: if exp_driver.exception: raise exp_driver.exception raise finally: # grace period to send last logs to sparkmagic # sparkmagic hb poll intervall is 5 seconds, therefore wait 6 seconds time.sleep(6) # cleanup spark jobs if running and exp_driver is not None: exp_driver.stop() run_id += 1 running = False sc.setJobGroup("", "") return result
def export(model_path, model_name, model_version=None, overwrite=False, metrics=None, description=None, synchronous=True, synchronous_timeout=120, project=None): """ Copies a trained model to the Models directory in the project and creates the directory structure of: >>> Models >>> | >>> - model_name >>> | >>> - version_x >>> | >>> - version_y For example if you run this: >>> from hops import model >>> model.export("iris_knn.pkl", "irisFlowerClassifier", metrics={'accuracy': accuracy}) It will copy the local model file "iris_knn.pkl" to /Projects/projectname/Models/irisFlowerClassifier/1/iris.knn.pkl on HDFS, and overwrite in case there already exists a file with the same name in the directory. If "model" is a directory on the local path exported by TensorFlow, and you run: >>> model.export("/model", "mnist", metrics={'accuracy': accuracy, 'loss': loss}) It will copy the model directory contents to /Projects/projectname/Models/mnist/1/ , e.g the "model.pb" file and the "variables" directory. Args: :model_path: absolute path to the trained model (HDFS or local) :model_name: name of the model :model_version: version of the model :overwrite: boolean flag whether to overwrite in case a model already exists in the exported directory :metrics: dict of evaluation metrics to attach to model :description: description about the model :synchronous: whether to synchronously wait for the model to be indexed in the models rest endpoint :synchronous_timeout: max timeout in seconds for waiting for the model to be indexed :project: the name of the project where the model should be saved to (default: current project). Note, the project must share its 'Models' dataset and make it writeable for this client. Returns: The path to where the model was exported Raises: :ValueError: if there was an error with the model due to invalid user input :ModelNotFound: if the model was not found """ # Make sure model name is a string, users could supply numbers model_name = str(model_name) if not isinstance(model_path, string_types): model_path = model_path.decode() if not description: description = 'A collection of models for ' + model_name project_path = hdfs.project_path(project) assert hdfs.exists(project_path + "Models"), "Your project is missing a dataset named Models, please create it." if not hdfs.exists(model_path) and not os.path.exists(model_path): raise ValueError("the provided model_path: {} , does not exist in HDFS or on the local filesystem".format( model_path)) # make sure metrics are numbers if metrics: _validate_metadata(metrics) model_dir_hdfs = project_path + constants.MODEL_SERVING.MODELS_DATASET + \ constants.DELIMITERS.SLASH_DELIMITER + model_name + constants.DELIMITERS.SLASH_DELIMITER if not hdfs.exists(model_dir_hdfs): hdfs.mkdir(model_dir_hdfs) hdfs.chmod(model_dir_hdfs, "ug+rwx") # User did not specify model_version, pick the current highest version + 1, set to 1 if no model exists version_list = [] if not model_version and hdfs.exists(model_dir_hdfs): model_version_directories = hdfs.ls(model_dir_hdfs) for version_dir in model_version_directories: try: if hdfs.isdir(version_dir): version_list.append(int(version_dir[len(model_dir_hdfs):])) except: pass if len(version_list) > 0: model_version = max(version_list) + 1 if not model_version: model_version = 1 # Path to directory in HDFS to put the model files model_version_dir_hdfs = model_dir_hdfs + str(model_version) # If version directory already exists and we are not overwriting it then fail if not overwrite and hdfs.exists(model_version_dir_hdfs): raise ValueError("Could not create model directory: {}, the path already exists, " "set flag overwrite=True " "to remove the version directory and create the correct directory structure".format(model_version_dir_hdfs)) # Overwrite version directory by deleting all content (this is needed for Provenance to register Model as deleted) if overwrite and hdfs.exists(model_version_dir_hdfs): hdfs.delete(model_version_dir_hdfs, recursive=True) hdfs.mkdir(model_version_dir_hdfs) # At this point we can create the version directory if it does not exist if not hdfs.exists(model_version_dir_hdfs): hdfs.mkdir(model_version_dir_hdfs) # Export the model files if os.path.exists(model_path): export_dir=_export_local_model(model_path, model_version_dir_hdfs, overwrite) else: export_dir=_export_hdfs_model(model_path, model_version_dir_hdfs, overwrite) print("Exported model " + model_name + " as version " + str(model_version) + " successfully.") jobName=None if constants.ENV_VARIABLES.JOB_NAME_ENV_VAR in os.environ: jobName = os.environ[constants.ENV_VARIABLES.JOB_NAME_ENV_VAR] kernelId=None if constants.ENV_VARIABLES.KERNEL_ID_ENV_VAR in os.environ: kernelId = os.environ[constants.ENV_VARIABLES.KERNEL_ID_ENV_VAR] # Attach modelName_modelVersion to experiment directory if project is None: model_project_name = hdfs.project_name() else : model_project_name = project experiment_project_name = hdfs.project_name() model_summary = { 'name': model_name, 'projectName': model_project_name, 'version': model_version, 'metrics': metrics, 'experimentId': None, 'experimentProjectName': experiment_project_name, 'description': description, 'jobName': jobName, 'kernelId': kernelId } if 'ML_ID' in os.environ: model_summary['experimentId'] = os.environ['ML_ID'] # Attach link from experiment to model experiment_json = experiment_utils._populate_experiment_model(model_name + '_' + str(model_version), project=project) experiment_utils._attach_experiment_xattr(os.environ['ML_ID'], experiment_json, 'MODEL_UPDATE') # Attach model metadata to models version folder experiment_utils._attach_model_xattr(model_name + "_" + str(model_version), experiment_utils.dumps(model_summary)) # Model metadata is attached asynchronously by Epipe, therefore this necessary to ensure following steps in a pipeline will not fail if synchronous: start_time = time.time() sleep_seconds = 5 for i in range(int(synchronous_timeout/sleep_seconds)): try: time.sleep(sleep_seconds) print("Polling " + model_name + " version " + str(model_version) + " for model availability.") resp = get_model(model_name, model_version, project_name=project) if resp.ok: print("Model now available.") return print(model_name + " not ready yet, retrying in " + str(sleep_seconds) + " seconds.") except ModelNotFound: pass print("Model not available during polling, set a higher value for synchronous_timeout to wait longer.") return export_dir