def applySyncByNamespace(self, app_id, exp_uid, task_name, args, namespace=None, ignore_result=False,time_limit=0): """ Run a task (task_name) on a set of args with a given app_id, and exp_uid asynchronously. Waits for computation to finish and returns the answer unless ignore_result=True in which case its a non-blocking call. If this method is called a sequence of times with the same namespace (defaults to exp_uid if not set) it is guaranteed that they will execute in order, each job finishing before the next begins Inputs: ::\n (string) app_id, (string) exp_id, (string) task_name, (json) args """ if namespace==None: namespace=exp_uid domain = self.__get_domain_for_job(app_id+"_"+exp_uid) job_uid = utils.getNewUID() submit_timestamp = utils.datetimeNow('string') if time_limit == 0: soft_time_limit = None hard_time_limit = None else: soft_time_limit = time_limit hard_time_limit = time_limit + .01 result = tasks.apply_sync_by_namespace.apply_async(args=[app_id,exp_uid,task_name, args, namespace, job_uid, submit_timestamp, time_limit], exchange='sync@'+domain, routing_key='sync@'+domain,soft_time_limit=soft_time_limit,time_limit=hard_time_limit) if ignore_result: return True else: return result.get(interval=.001)
def applySyncByNamespace(self, app_id, exp_uid, task_name, args, namespace=None, ignore_result=False,time_limit=0): """ Run a task (task_name) on a set of args with a given app_id, and exp_uid asynchronously. Waits for computation to finish and returns the answer unless ignore_result=True in which case its a non-blocking call. If this method is called a sequence of times with the same namespace (defaults to exp_uid if not set) it is guaranteed that they will execute in order, each job finishing before the next begins Inputs: ::\n (string) app_id, (string) exp_id, (string) task_name, (json) args """ submit_timestamp = utils.datetimeNow('string') if namespace==None: namespace=exp_uid domain = self.__get_domain_for_job(app_id+"_"+exp_uid) num_queues = next.constants.CELERY_SYNC_WORKER_COUNT # assign namespaces to queues (with worker of concurrency 1) in round-robbin try: namespace_cnt = int(self.r.get(namespace+"_cnt")) except: pipe = self.r.pipeline(True) while 1: try: pipe.watch(namespace+"_cnt","namespace_counter") if not pipe.exists(namespace+"_cnt"): if not pipe.exists('namespace_counter'): namespace_counter = 0 else: namespace_counter = pipe.get('namespace_counter') pipe.multi() pipe.set(namespace+"_cnt",int(namespace_counter)+1) pipe.set('namespace_counter',int(namespace_counter)+1) pipe.execute() else: pipe.unwatch() break except redis.exceptions.WatchError: continue finally: pipe.reset() namespace_cnt = int(self.r.get(namespace+"_cnt")) queue_number = (namespace_cnt % num_queues) + 1 queue_name = 'sync_queue_'+str(queue_number)+'@'+domain job_uid = utils.getNewUID() if time_limit == 0: soft_time_limit = None hard_time_limit = None else: soft_time_limit = time_limit hard_time_limit = time_limit + .01 result = tasks.apply_sync_by_namespace.apply_async(args=[app_id,exp_uid,task_name, args, namespace, job_uid, submit_timestamp, time_limit], queue=queue_name,soft_time_limit=soft_time_limit,time_limit=hard_time_limit) if ignore_result: return True else: return result.get(interval=.001)
def getQuery(self, exp_uid, args_json): try: args_dict = self.helper.convert_json(args_json) args_dict = verifier.verify(args_dict, self.reference_dict['getQuery']['args']) experiment_dict = self.butler.experiment.get() alg_list = experiment_dict['args']['alg_list'] participant_to_algorithm_management = experiment_dict['args']['participant_to_algorithm_management'] algorithm_management_settings = experiment_dict['args']['algorithm_management_settings'] # Create the participant dictionary in participants bucket if needed. Also pull out label and id for this algorithm participant_uid = args_dict['args'].get('participant_uid', args_dict['exp_uid']) # Check to see if the first participant has come by and if not, save to db participant_doc = self.butler.participants.get(uid=participant_uid) first_participant_query = participant_doc==None if first_participant_query: participant_doc = {} self.butler.participants.set(uid=participant_uid, value={'exp_uid':exp_uid, 'participant_uid':participant_uid}) if (participant_uid == exp_uid) or (participant_to_algorithm_management == 'one_to_many') or (first_participant_query): if algorithm_management_settings['mode'] == 'fixed_proportions': prop = [prop_item['proportion'] for prop_item in algorithm_management_settings['params']] chosen_alg = numpy.random.choice(alg_list, p=prop) elif algorithm_management_settings['mode'] == 'custom' : chosen_alg = self.myApp.chooseAlg(self.butler, alg_list, args_dict['args']) else: chosen_alg = numpy.random.choice(alg_list) alg_id = chosen_alg['alg_id'] alg_label = chosen_alg['alg_label'] if (first_participant_query) and (participant_to_algorithm_management=='one_to_one'): self.butler.participants.set(uid=participant_uid, key='alg_id',value=alg_id) self.butler.participants.set(uid=participant_uid, key='alg_label',value=alg_label) elif (participant_to_algorithm_management=='one_to_one'): alg_id = participant_doc['alg_id'] alg_label = participant_doc['alg_label'] query_uid = utils.getNewUID() args_dict['args'].update(query_uid=query_uid) query_doc = self.call_app_fn(alg_label, alg_id, 'getQuery', args_dict) query_doc.update({'participant_uid':participant_uid, 'alg_id':alg_id, 'exp_uid':exp_uid, 'alg_label':alg_label, 'timestamp_query_generated':str(utils.datetimeNow()), 'query_uid':query_uid}) self.butler.queries.set(uid=query_uid, value=query_doc) return json.dumps({'args':query_doc,'meta':{'log_entry_durations':self.log_entry_durations}}), True,'' except Exception, error: exc_type, exc_value, exc_traceback = sys.exc_info() full_error = str(traceback.format_exc())+'\n'+str(error) utils.debug_print("getQuery Exception: " + full_error, color='red') log_entry = { 'exp_uid':exp_uid,'task':'getQuery','error':full_error,'timestamp':utils.datetimeNow(),'args_json':args_json } self.butler.ell.log( self.app_id+':APP-EXCEPTION', log_entry ) traceback.print_tb(exc_traceback) return '{}', False, str(error)
def applySyncByNamespace(self, app_id, exp_uid, task_name, args, namespace=None, ignore_result=False, time_limit=0): """ Run a task (task_name) on a set of args with a given app_id, and exp_uid asynchronously. Waits for computation to finish and returns the answer unless ignore_result=True in which case its a non-blocking call. If this method is called a sequence of times with the same namespace (defaults to exp_uid if not set) it is guaranteed that they will execute in order, each job finishing before the next begins Inputs: ::\n (string) app_id, (string) exp_id, (string) task_name, (json) args """ if namespace == None: namespace = exp_uid domain = self.__get_domain_for_job(app_id + "_" + exp_uid) job_uid = utils.getNewUID() submit_timestamp = utils.datetimeNow('string') if time_limit == 0: soft_time_limit = None hard_time_limit = None else: soft_time_limit = time_limit hard_time_limit = time_limit + .01 result = tasks.apply_sync_by_namespace.apply_async( args=[ app_id, exp_uid, task_name, args, namespace, job_uid, submit_timestamp, time_limit ], exchange='sync@' + domain, routing_key='sync@' + domain, soft_time_limit=soft_time_limit, time_limit=hard_time_limit) if ignore_result: return True else: return result.get(interval=.001)
def initExp(self,exp_uid,args_json,db,ell): """ initialize the project and necessary experiments Expected input (in json structure with string keys): (int) n: number of arms (int) k: number of objects to display (float) failure_probability : confidence [optional] (list of dicts) alg_list : with fields (Defaults given by Info.get_app_default_alg_list) (string) alg_id : valid alg_id for this app_id (string) alg_label : unique identifier for algorithm (e.g. may have experiment with repeated alg_id's, but alg_labels must be unqiue, will also be used for plot legends [optional] (string) test_alg_label : must be one of the alg_label's in alg_list (Default is self) [optional] (dict) algorithm_management_settings : dictionary with fields (string) 'mode' and (dict) 'params'. mode in {'pure_exploration','explore_exploit','fixed_proportions'}. Default is 'fixed_proportions' and allocates uniform probability to each algorithm. If mode=fixed_proportions then params is a dictionary that contains the field 'proportions' which is a list of dictionaries with fields 'alg_label' and 'proportion' for all algorithms in alg_list. All proportions must be positive and sum to 1 over all algs in alg_list [optional] (string) participant_to_algorithm_management : in {'one_to_one','one_to_many'}. Default is 'one_to_many'. [optional] (string) instructions [optional] (string) debrief Expected output: if error: return (JSON) '{}', (bool) False, (str) error_str else: return (JSON) '{}', (bool) True,'' Usage: initExp_response_json,didSucceed,message = app.initExp(exp_uid,initExp_args_json) Example input: initExp_args_json = {"participant_to_algorithm_management": "one_to_many", "alg_list": [{"alg_label": "BR_LilUCB", "alg_id": "BR_LilUCB", "params": {}}], "algorithm_management_settings": {"params": {"proportions": [{"alg_label": "BR_LilUCB", "proportion": 1.0}]}, "mode": "fixed_proportions"}, "failure_probability": 0.01, "n": 10} Example output: initExp_response_json = {} """ try: app_id = self.app_id # remove any reminants of an experiment if it exists didSucceed,message = db.delete_docs_with_filter('experiments_admin',{'exp_uid':exp_uid}) didSucceed,message = db.delete_docs_with_filter(app_id+':experiments',{'exp_uid':exp_uid}) didSucceed,message = db.delete_docs_with_filter(app_id+':queries',{'exp_uid':exp_uid}) didSucceed,message = db.delete_docs_with_filter(app_id+':participants',{'exp_uid':exp_uid}) didSucceed,message = db.delete_docs_with_filter(app_id+':algorithms',{'exp_uid':exp_uid}) didSucceed,message = ell.delete_logs_with_filter(app_id+':APP-CALL',{'exp_uid':exp_uid}) didSucceed,message = ell.delete_logs_with_filter(app_id+':APP-RESPONSE',{'exp_uid':exp_uid}) didSucceed,message = ell.delete_logs_with_filter(app_id+':APP-EXCEPTION',{'exp_uid':exp_uid}) didSucceed,message = ell.delete_logs_with_filter(app_id+':ALG-DURATION',{'exp_uid':exp_uid}) didSucceed,message = ell.delete_logs_with_filter(app_id+':ALG-EVALUATION',{'exp_uid':exp_uid}) # add indexes (only adds them if they do not already exist) didSucceed,message = db.ensure_index('experiments_admin',{'exp_uid':1}) didSucceed,message = db.ensure_index(app_id+':experiments',{'exp_uid':1}) didSucceed,message = db.ensure_index(app_id+':queries',{'query_uid':1}) didSucceed,message = db.ensure_index(app_id+':queries',{'exp_uid':1}) didSucceed,message = db.ensure_index(app_id+':queries',{'alg_uid':1}) didSucceed,message = db.ensure_index(app_id+':queries',{'participant_uid':1}) didSucceed,message = db.ensure_index(app_id+':participants',{'participant_uid':1}) didSucceed,message = db.ensure_index(app_id+':participants',{'exp_uid':1}) didSucceed,message = db.ensure_index(app_id+':algorithms',{'alg_uid':1}) didSucceed,message = db.ensure_index(app_id+':algorithms',{'exp_uid':1}) didSucceed,message = ell.ensure_index(app_id+':APP-CALL',{'exp_uid':1}) didSucceed,message = ell.ensure_index(app_id+':APP-CALL',{'timestamp':1}) didSucceed,message = ell.ensure_index(app_id+':APP-CALL',{'exp_uid':1,'timestamp':1}) didSucceed,message = ell.ensure_index(app_id+':APP-CALL',{'exp_uid':1,'task':1}) didSucceed,message = ell.ensure_index(app_id+':APP-RESPONSE',{'exp_uid':1}) didSucceed,message = ell.ensure_index(app_id+':APP-RESPONSE',{'timestamp':1}) didSucceed,message = ell.ensure_index(app_id+':APP-RESPONSE',{'exp_uid':1,'timestamp':1}) didSucceed,message = ell.ensure_index(app_id+':APP-RESPONSE',{'exp_uid':1,'task':1}) didSucceed,message = ell.ensure_index(app_id+':APP-EXCEPTION',{'exp_uid':1}) didSucceed,message = ell.ensure_index(app_id+':APP-EXCEPTION',{'timestamp':1}) didSucceed,message = ell.ensure_index(app_id+':APP-EXCEPTION',{'exp_uid':1,'timestamp':1}) didSucceed,message = ell.ensure_index(app_id+':APP-EXCEPTION',{'exp_uid':1,'task':1}) didSucceed,message = ell.ensure_index(app_id+':ALG-DURATION',{'exp_uid':1}) didSucceed,message = ell.ensure_index(app_id+':ALG-DURATION',{'alg_uid':1}) didSucceed,message = ell.ensure_index(app_id+':ALG-DURATION',{'timestamp':1}) didSucceed,message = ell.ensure_index(app_id+':ALG-DURATION',{'exp_uid':1,'timestamp':1}) didSucceed,message = ell.ensure_index(app_id+':ALG-DURATION',{'alg_uid':1,'task':1}) didSucceed,message = ell.ensure_index(app_id+':ALG-EVALUATION',{'exp_uid':1}) didSucceed,message = ell.ensure_index(app_id+':ALG-EVALUATION',{'alg_uid':1}) didSucceed,message = ell.ensure_index(app_id+':ALG-EVALUATION',{'timestamp':1}) didSucceed,message = ell.ensure_index(app_id+':ALG-EVALUATION',{'exp_uid':1,'timestamp':1}) db.set('experiments_admin',exp_uid,'exp_uid',exp_uid) db.set('experiments_admin',exp_uid,'app_id',app_id) db.set('experiments_admin',exp_uid,'start_date',utils.datetime2str(utils.datetimeNow())) log_entry = { 'exp_uid':exp_uid,'task':'initExp','json':args_json,'timestamp':utils.datetimeNow() } ell.log( app_id+':APP-CALL', log_entry ) # convert args_json to args_dict try: args_dict = json.loads(args_json) except: error = "%s.initExp input args_json is in improper format" % self.app_id return '{}',False,error # check for the fields that must be contained in args or error occurs necessary_fields = ['n','k','failure_probability'] for field in necessary_fields: try: args_dict[field] except KeyError: error = "%s.initExp input arguments missing field: %s" % (self.app_id,str(field)) return '{}',False,error n = args_dict['n'] k = args_dict['k'] delta = args_dict['failure_probability'] if 'alg_list' in args_dict: alg_list = args_dict['alg_list'] supportedAlgs = utils.get_app_supported_algs(self.app_id) for algorithm in alg_list: if algorithm['alg_id'] not in supportedAlgs: error = "%s.initExp unsupported algorithm '%s' in alg_list" % (self.app_id,alg_id) return '{}',False,error else: alg_list = utils.get_app_default_alg_list(self.app_id) if 'instructions' not in args_dict: instructions = utils.get_app_default_instructions(app_id) else: instructions = args_dict['instructions'] if 'debrief' not in args_dict: debrief = utils.get_app_default_instructions(app_id) else: debrief = args_dict['debrief'] if 'num_tries' not in args_dict: num_tries = utils.get_app_default_num_tries(app_id) else: num_tries = args_dict['num_tries'] if 'context_type' not in args_dict: context_type = 'none' else: context_type = args_dict['context_type'] if 'context' not in args_dict: context = '' else: context = args_dict['context'] # ALGORITHM_MANAGEMENT_MODE FORMATTING CHECK if 'algorithm_management_settings' not in args_dict: params = {} params['proportions'] = [] for algorithm in alg_list: params['proportions'].append( { 'alg_label': algorithm['alg_label'] , 'proportion':1./len(alg_list)} ) algorithm_management_settings = {} algorithm_management_settings['mode'] = 'fixed_proportions' algorithm_management_settings['params'] = params else: algorithm_management_settings = args_dict['algorithm_management_settings'] try: mode = algorithm_management_settings['mode'] params = algorithm_management_settings['params'] except: error = "%s.initExp algorithm_management_settings must be a dictionary with fields 'mode' and 'params'" % (self.app_id) return '{}',False,error if mode == 'fixed_proportions': try: algorithm_proportions_list = params['proportions'] except: error = "%s.initExp algorithm_management_settings['params'] must be a dictionary with field 'proportions'" % (self.app_id) return '{}',False,error # check if alg_labels are properly labeled for proportion_item in algorithm_proportions_list: proportion = proportion_item['proportion'] target_alg_label = proportion_item['alg_label'] target_alg_label_in_alg_list = False for algorithm in alg_list: if algorithm['alg_label']==target_alg_label: target_alg_label_in_alg_list = True if not target_alg_label_in_alg_list: error = "%s.initExp algorithm_management_settings['params']['proportions'] must be a list of dictionaries, each dictionary containing the fields 'alg_label' and 'proportion'. The 'alg_label' value must be one of the alg_labels in a provided alg_list and 'proportion' must be nonnegative and sum to 1 : '%s' not in provided alg_list" % (self.app_id,target_alg_label) return '{}',False,error elif mode == 'pure_exploration': error = "%s.initExp Sorry, '%s' is not yet supported." % (self.app_id,mode) return '{}',False,error elif mode == 'explore_exploit': error = "%s.initExp Sorry, '%s' is not yet supported." % (self.app_id,mode) return '{}',False,error else: error = "%s.initExp unsupported algorithm_management_mode: '%s'. Must be in {'pure_exploration','explore_exploit','fixed_proportions'}" % (self.app_id,algorithm_management_mode) return '{}',False,error # ALGORITHM_MANAGEMENT_MODE FORMATTING CHECK if 'participant_to_algorithm_management' not in args_dict: participant_to_algorithm_management = 'one_to_many' else: participant_to_algorithm_management = args_dict['participant_to_algorithm_management'] if participant_to_algorithm_management not in ['one_to_many','one_to_one']: error = "%s.initExp unsupported participant_to_algorithm_management: '%s'. Must be in {'one_to_many','one_to_one'}" % (self.app_id,participant_to_algorithm_management) return '{}',False,error # assign uid to each algorithm and save it for algorithm in alg_list: alg_uid = utils.getNewUID() algorithm['alg_uid'] = alg_uid db.set(app_id+':algorithms',alg_uid,'alg_uid',alg_uid) db.set(app_id+':algorithms',alg_uid,'exp_uid',exp_uid) db.set(app_id+':experiments',exp_uid,'exp_uid',exp_uid) db.set(app_id+':experiments',exp_uid,'app_id',app_id) db.set(app_id+':experiments',exp_uid,'n',n) db.set(app_id+':experiments',exp_uid,'k',k) db.set(app_id+':experiments',exp_uid,'failure_probability',delta) db.set(app_id+':experiments',exp_uid,'alg_list',alg_list) db.set(app_id+':experiments',exp_uid,'algorithm_management_settings',algorithm_management_settings) db.set(app_id+':experiments',exp_uid,'participant_to_algorithm_management',participant_to_algorithm_management) db.set(app_id+':experiments',exp_uid,'instructions',instructions) db.set(app_id+':experiments',exp_uid,'debrief',debrief) db.set(app_id+':experiments',exp_uid,'context_type',context_type) db.set(app_id+':experiments',exp_uid,'context',context) db.set(app_id+':experiments',exp_uid,'num_tries',num_tries) # now create intitialize each algorithm for algorithm in alg_list: alg_id = algorithm['alg_id'] alg_uid = algorithm['alg_uid'] db.set(app_id+':algorithms',alg_uid,'alg_id',alg_id) db.set(app_id+':algorithms',alg_uid,'alg_uid',alg_uid) db.set(app_id+':algorithms',alg_uid,'exp_uid',exp_uid) # get sandboxed database for the specific app_id,alg_id,exp_uid - closing off the rest of the database to the algorithm rc = ResourceClient(app_id,exp_uid,alg_uid,db) # get specific algorithm to make calls to alg = utils.get_app_alg(self.app_id,alg_id) # call initExp didSucceed,dt = utils.timeit(alg.initExp)(resource=rc,n=n,k=k,failure_probability=delta) log_entry = { 'exp_uid':exp_uid,'alg_uid':alg_uid,'task':'initExp','duration':dt,'timestamp':utils.datetimeNow() } ell.log( app_id+':ALG-DURATION', log_entry ) response_json = '{}' log_entry = { 'exp_uid':exp_uid,'task':'initExp','json':response_json,'timestamp':utils.datetimeNow() } ell.log( app_id+':APP-RESPONSE', log_entry ) return response_json,True,'' except Exception, err: error = traceback.format_exc() log_entry = { 'exp_uid':exp_uid,'task':'initExp','error':error,'timestamp':utils.datetimeNow(),'args_json':args_json } ell.log( app_id+':APP-EXCEPTION', log_entry ) return '{}',False,error
def getQuery(self,exp_uid,args_json,db,ell): """ A request to ask which k arms to duel next Expected input (in jsonstructure with string keys): (int) k : number of objects to display [optional] (string) participant_uid : unique identifier of session for a participant answering questions (that is, an email address is not good enough as the participant could participate in multiple exp_uids so it would not be unique against all experiments), if key non-existant particpant_uid is assigned as exp_uid. Expected output (in json structure with string keys): (list) target_indices : list of k target indexes e.g. [ (int) target_index_1, ... , (int) target_index_k ] (str) query_uid : unique identifier of query (used to look up for reportAnswer) Usage: getQuery_response_json,didSucceed,message = app.getQuery(exp_uid,getQuery_args_json) Example input: getQuery_args_json = {"k": 3, "participant_uid": "0077110d03cf06b8f77d11acc399e8a7"} Example output: getQuery_response_json = {"query_uid": "4d02a9924f92138287edd17ca5feb6e1", "target_indices": [ 3, 6, 9 ] """ try: app_id = self.app_id log_entry = { 'exp_uid':exp_uid,'task':'getQuery','json':args_json,'timestamp':utils.datetimeNow() } ell.log( app_id+':APP-CALL', log_entry ) # convert args_json to args_dict try: args_dict = json.loads(args_json) except: error = "%s.initExp input args_json is in improper format" % self.app_id return '{}',False,error # get list of algorithms associated with project alg_list,didSucceed,message = db.get(app_id+':experiments',exp_uid,'alg_list') alg_label_to_alg_id = {} alg_label_to_alg_uid = {} for algorithm in alg_list: alg_label_to_alg_id[ algorithm['alg_label'] ] = algorithm['alg_id'] alg_label_to_alg_uid[ algorithm['alg_label'] ] = algorithm['alg_uid'] algorithm_management_settings,didSucceed,message = db.get(app_id+':experiments',exp_uid,'algorithm_management_settings') # ASSIGN ALGORITHM TO PARTICIPANT if 'participant_uid' in args_dict: participant_uid = args_dict['participant_uid'] else: participant_uid = exp_uid participant_doc_exists,didSucceed,message = db.exists(app_id+':participants',participant_uid,'participant_uid') first_participant_query = not participant_doc_exists if first_participant_query: db.set(app_id+':participants',participant_uid,'participant_uid',participant_uid) db.set(app_id+':participants',participant_uid,'exp_uid',exp_uid) participant_to_algorithm_management,didSucceed,message = db.get(app_id+':experiments',exp_uid,'participant_to_algorithm_management') if (participant_uid==exp_uid) or (participant_to_algorithm_management=='one_to_many') or (first_participant_query): if algorithm_management_settings['mode']=='fixed_proportions': proportions_list = algorithm_management_settings['params']['proportions'] prop = [ prop_item['proportion'] for prop_item in proportions_list ] prop_item = numpy.random.choice(alg_list,p=prop) else: raise Exception('algorithm_management_mode : '+algorithm_management_settings['mode']+' not implemented') alg_id = alg_label_to_alg_id[ prop_item['alg_label'] ] alg_uid = alg_label_to_alg_uid[ prop_item['alg_label'] ] alg_label = prop_item['alg_label'] if (first_participant_query) and (participant_to_algorithm_management=='one_to_one'): db.set(app_id+':participants',participant_uid,'alg_id',alg_id) db.set(app_id+':participants',participant_uid,'alg_uid',alg_uid) elif (participant_to_algorithm_management=='one_to_one'): # If here, then alg_uid should already be assigned in participant doc alg_id,didSucceed,message = db.get(app_id+':participants',participant_uid,'alg_id') alg_uid,didSucceed,message = db.get(app_id+':participants',participant_uid,'alg_uid') else: raise Exception('participant_to_algorithm_management : '+participant_to_algorithm_management+' not implemented') # get sandboxed database for the specific app_id,alg_id,exp_uid - closing off the rest of the database to the algorithm rc = ResourceClient(app_id,exp_uid,alg_uid,db) # get specific algorithm to make calls to alg = utils.get_app_alg(self.app_id,alg_id) # call getQuery targets,dt = utils.timeit(alg.getQuery)(resource=rc) # update targets target_list = [] for target in targets: target_list.append({'index':target}) # check for context context_type,didSucceed,message = db.get(app_id+':experiments',exp_uid,'context_type') context,didSucceed,message = db.get(app_id+':experiments',exp_uid,'context') # log log_entry_durations = { 'exp_uid':exp_uid,'alg_uid':alg_uid,'task':'getQuery','duration':dt } log_entry_durations.update( rc.getDurations() ) meta = {'log_entry_durations':log_entry_durations} # create JSON query payload timestamp = str(utils.datetimeNow()) query_uid = utils.getNewUID() query = {} query['query_uid'] = query_uid query['target_indices'] = target_list # save query data to database query_doc = {} query_doc.update(query) query_doc['participant_uid'] = participant_uid query_doc['alg_uid'] = alg_uid query_doc['exp_uid'] = exp_uid query_doc['alg_label'] = alg_label query_doc['timestamp_query_generated'] = timestamp for field in query_doc: db.set(app_id+':queries',query_uid,field,query_doc[field]) # add context after updating query doc to avoid redundant information query['context_type'] = context_type query['context'] = context args_out = {'args':query,'meta':meta} response_json = json.dumps(args_out) log_entry = { 'exp_uid':exp_uid,'task':'getQuery','json':response_json,'timestamp':utils.datetimeNow() } ell.log( app_id+':APP-RESPONSE', log_entry ) return response_json,True,'' except Exception, err: error = traceback.format_exc() log_entry = { 'exp_uid':exp_uid,'task':'getQuery','error':error,'timestamp':utils.datetimeNow(),'args_json':args_json } ell.log( app_id+':APP-EXCEPTION', log_entry ) return '{}',False,error
def applySyncByNamespace(self, app_id, exp_uid, alg_id, alg_label, task_name, args, namespace=None, ignore_result=False, time_limit=0): """ Run a task (task_name) on a set of args with a given app_id, and exp_uid asynchronously. Waits for computation to finish and returns the answer unless ignore_result=True in which case its a non-blocking call. If this method is called a sequence of times with the same namespace (defaults to exp_uid if not set) it is guaranteed that they will execute in order, each job finishing before the next begins Inputs: ::\n (string) app_id, (string) exp_id, (string) task_name, (json) args """ submit_timestamp = utils.datetimeNow('string') if namespace == None: namespace = exp_uid domain = self.__get_domain_for_job(app_id + "_" + exp_uid) num_queues = next.constants.CELERY_SYNC_WORKER_COUNT # assign namespaces to queues (with worker of concurrency 1) in round-robbin try: namespace_cnt = int(self.r.get(namespace + "_cnt")) except: pipe = self.r.pipeline(True) while 1: try: pipe.watch(namespace + "_cnt", "namespace_counter") if not pipe.exists(namespace + "_cnt"): if not pipe.exists('namespace_counter'): namespace_counter = 0 else: namespace_counter = pipe.get('namespace_counter') pipe.multi() pipe.set(namespace + "_cnt", int(namespace_counter) + 1) pipe.set('namespace_counter', int(namespace_counter) + 1) pipe.execute() else: pipe.unwatch() break except redis.exceptions.WatchError: continue finally: pipe.reset() namespace_cnt = int(self.r.get(namespace + "_cnt")) queue_number = (namespace_cnt % num_queues) + 1 queue_name = 'sync_queue_' + str(queue_number) + '@' + domain job_uid = utils.getNewUID() if time_limit == 0: soft_time_limit = None hard_time_limit = None else: soft_time_limit = time_limit hard_time_limit = time_limit + .01 if next.constants.CELERY_ON: result = tasks.apply_sync_by_namespace.apply_async( args=[ app_id, exp_uid, alg_id, alg_label, task_name, args, namespace, job_uid, submit_timestamp, time_limit ], queue=queue_name, soft_time_limit=soft_time_limit, time_limit=hard_time_limit) if ignore_result: return True else: return result.get(interval=.001) else: result = tasks.apply_sync_by_namespace(app_id, exp_uid, alg_id, alg_label, task_name, args, namespace, job_uid, submit_timestamp, time_limit) if ignore_result: return True else: return result
def getQuery(self,exp_uid,args_json,db,ell): """ A request to ask which two arms to duel next Expected input (in jsonstructure with string keys): [optional] (string) participant_uid : unique identifier of session for a participant answering questions (that is, an email address is not good enough as the participant could participate in multiple exp_uids so it would not be unique against all experiments), if key non-existant particpant_uid is assigned as exp_uid. Expected output (in json structure with string keys): (list) target_indices : list that stores dictionary of targets with fields: { (int) index : the index of the target of relevance (str) label : in {'left','right'} for display (int) flag : integer for algorithm's use } (str) query_uid : unique identifier of query (used to look up for processAnswer) """ try: app_id = self.app_id log_entry = { 'exp_uid':exp_uid,'task':'getQuery','json':args_json,'timestamp':utils.datetimeNow() } ell.log( app_id+':APP-CALL', log_entry ) # convert args_json to args_dict try: args_dict = json.loads(args_json) except: error = "%s.initExp input args_json is in improper format" % self.app_id return '{}',False,error # get list of algorithms associated with project alg_list,didSucceed,message = db.get(app_id+':experiments',exp_uid,'alg_list') alg_label_to_alg_id = {} alg_label_to_alg_uid = {} for algorithm in alg_list: alg_label_to_alg_id[ algorithm['alg_label'] ] = algorithm['alg_id'] alg_label_to_alg_uid[ algorithm['alg_label'] ] = algorithm['alg_uid'] algorithm_management_settings,didSucceed,message = db.get(app_id+':experiments',exp_uid,'algorithm_management_settings') # ASSIGN ALGORITHM TO PARTICIPANT if 'participant_uid' in args_dict: participant_uid = args_dict['participant_uid'] else: participant_uid = exp_uid participant_doc_exists,didSucceed,message = db.exists(app_id+':participants',participant_uid,'participant_uid') first_participant_query = not participant_doc_exists if first_participant_query: db.set(app_id+':participants',participant_uid,'participant_uid',participant_uid) db.set(app_id+':participants',participant_uid,'exp_uid',exp_uid) participant_to_algorithm_management,didSucceed,message = db.get(app_id+':experiments',exp_uid,'participant_to_algorithm_management') if (participant_uid==exp_uid) or (participant_to_algorithm_management=='one_to_many') or (first_participant_query): if algorithm_management_settings['mode']=='fixed_proportions': proportions_list = algorithm_management_settings['params']['proportions'] prop = [ prop_item['proportion'] for prop_item in proportions_list ] prop_item = numpy.random.choice(alg_list,p=prop) else: raise Exception('algorithm_management_mode : '+algorithm_management_settings['mode']+' not implemented') alg_id = alg_label_to_alg_id[ prop_item['alg_label'] ] alg_uid = alg_label_to_alg_uid[ prop_item['alg_label'] ] alg_label = prop_item['alg_label'] if (first_participant_query) and (participant_to_algorithm_management=='one_to_one'): db.set(app_id+':participants',participant_uid,'alg_id',alg_id) db.set(app_id+':participants',participant_uid,'alg_uid',alg_uid) elif (participant_to_algorithm_management=='one_to_one'): # If here, then alg_uid should already be assigned in participant doc alg_id,didSucceed,message = db.get(app_id+':participants',participant_uid,'alg_id') alg_uid,didSucceed,message = db.get(app_id+':participants',participant_uid,'alg_uid') else: raise Exception('participant_to_algorithm_management : '+participant_to_algorithm_management+' not implemented') # get sandboxed database for the specific app_id,alg_id,exp_uid - closing off the rest of the database to the algorithm rc = ResourceClient(app_id,exp_uid,alg_uid,db) # get specific algorithm to make calls to alg = utils.get_app_alg(self.app_id,alg_id) # call getQuery index_left,index_right,index_painted,dt = utils.timeit(alg.getQuery)(resource=rc) # check for context context_type,didSucceed,message = db.get(app_id+':experiments',exp_uid,'context_type') context,didSucceed,message = db.get(app_id+':experiments',exp_uid,'context') # log log_entry_durations = { 'exp_uid':exp_uid,'alg_uid':alg_uid,'task':'getQuery','duration':dt } log_entry_durations.update( rc.getDurations() ) meta = {'log_entry_durations':log_entry_durations} # create JSON query payload if index_left==index_painted: targets = [ {'index':index_left,'label':'left','flag':1}, {'index':index_right,'label':'right','flag':0} ] else: targets = [ {'index':index_left,'label':'left','flag':0}, {'index':index_right,'label':'right','flag':1} ] timestamp = str(utils.datetimeNow()) query_uid = utils.getNewUID() query = {} query['query_uid'] = query_uid query['target_indices'] = targets # save query data to database query_doc = {} query_doc.update(query) query_doc['participant_uid'] = participant_uid query_doc['alg_uid'] = alg_uid query_doc['exp_uid'] = exp_uid query_doc['alg_label'] = alg_label query_doc['timestamp_query_generated'] = timestamp db.set_doc(app_id+':queries',query_uid,query_doc) # add context after updating query doc to avoid redundant information query['context_type'] = context_type query['context'] = context args_out = {'args':query,'meta':meta} response_json = json.dumps(args_out) log_entry = { 'exp_uid':exp_uid,'task':'getQuery','json':response_json,'timestamp':utils.datetimeNow() } ell.log( app_id+':APP-RESPONSE', log_entry ) return response_json,True,'' except Exception, err: error = traceback.format_exc() log_entry = { 'exp_uid':exp_uid,'task':'getQuery','error':error,'timestamp':utils.datetimeNow(),'args_json':args_json } ell.log( app_id+':APP-EXCEPTION', log_entry ) return '{}',False,error
def getQuery(self, exp_uid, args_json): try: args_dict = self.helper.convert_json(args_json) args_dict = verifier.verify(args_dict, self.reference_dict['getQuery']['args']) experiment_dict = self.butler.experiment.get() alg_list = experiment_dict['args']['alg_list'] participant_to_algorithm_management = experiment_dict['args']['participant_to_algorithm_management'] algorithm_management_settings = experiment_dict['args']['algorithm_management_settings'] # Create the participant dictionary in participants bucket if needed. Also pull out label and id for this algorithm participant_uid = args_dict['args'].get('participant_uid', args_dict['exp_uid']) # Check to see if the first participant has come by and if not, save to db participant_doc = self.butler.participants.get(uid=participant_uid) first_participant_query = participant_doc==None if first_participant_query: participant_doc = {} self.butler.participants.set(uid=participant_uid, value={'exp_uid':exp_uid, 'participant_uid':participant_uid}) if (participant_uid == exp_uid) or (participant_to_algorithm_management == 'one_to_many') or (first_participant_query): if algorithm_management_settings['mode'] == 'fixed_proportions': labels = [alg['alg_label'] for alg in algorithm_management_settings['params']] prop = [prop_item['proportion'] for prop_item in algorithm_management_settings['params']] # reorder prop and alg_list to have same order new_alg_list = [] broken = False for label in labels: broken = False for alg in alg_list: if label == alg['alg_label']: new_alg_list += [alg] broken = True break if not broken: raise Exception('alg_label not present for both porportions and labels') chosen_alg = numpy.random.choice(new_alg_list, p=prop) elif algorithm_management_settings['mode'] == 'custom' : chosen_alg = self.myApp.chooseAlg(self.butler, alg_list, args_dict['args']) else: chosen_alg = numpy.random.choice(alg_list) alg_id = chosen_alg['alg_id'] alg_label = chosen_alg['alg_label'] if (first_participant_query) and (participant_to_algorithm_management=='one_to_one'): self.butler.participants.set(uid=participant_uid, key='alg_id',value=alg_id) self.butler.participants.set(uid=participant_uid, key='alg_label',value=alg_label) elif (participant_to_algorithm_management=='one_to_one'): alg_id = participant_doc['alg_id'] alg_label = participant_doc['alg_label'] query_uid = utils.getNewUID() args_dict['args'].update(query_uid=query_uid) query_doc = self.call_app_fn(alg_label, alg_id, 'getQuery', args_dict) query_doc.update({'participant_uid':participant_uid, 'alg_id':alg_id, 'exp_uid':exp_uid, 'alg_label':alg_label, 'timestamp_query_generated':str(utils.datetimeNow()), 'query_uid':query_uid}) self.butler.queries.set(uid=query_uid, value=query_doc) return json.dumps({'args':query_doc,'meta':{'log_entry_durations':self.log_entry_durations}}), True,'' except Exception, error: exc_type, exc_value, exc_traceback = sys.exc_info() full_error = str(traceback.format_exc())+'\n'+str(error) utils.debug_print("getQuery Exception: " + full_error, color='red') log_entry = { 'exp_uid':exp_uid,'task':'getQuery','error':full_error,'timestamp':utils.datetimeNow(),'args_json':args_json } self.butler.ell.log( self.app_id+':APP-EXCEPTION', log_entry ) traceback.print_tb(exc_traceback) return '{}', False, str(error)
def getQuery(self, exp_uid, args_json, db, ell): """ A request to ask which two arms to duel next Expected input (in jsonstructure with string keys): [optional] (string) participant_uid : unique identifier of session for a participant answering questions (that is, an email address is not good enough as the participant could participate in multiple exp_uids so it would not be unique against all experiments), if key non-existant particpant_uid is assigned as exp_uid. Expected output (in json structure with string keys): (list) target_indices : list that stores dictionary of targets with fields: { (int) index : the index of the target of relevance (str) label : in {'left','right'} for display (int) flag : integer for algorithm's use } (str) query_uid : unique identifier of query (used to look up for processAnswer) Usage: getQuery_response_json,didSucceed,message = app.getQuery(db_API,exp_uid,getQuery_args_json) Example input: getQuery_args_json = {"participant_uid": "0077110d03cf06b8f77d11acc399e8a7"} Example output: getQuery_response_json = {"query_uid": "4d02a9924f92138287edd17ca5feb6e1", "target_indices": [{"index": 9, "flag": 0, "label": "left"}, {"index": 8, "flag": 1, "label": "right"}]} """ try: app_id = self.app_id log_entry = { 'exp_uid': exp_uid, 'task': 'getQuery', 'json': args_json, 'timestamp': utils.datetimeNow() } ell.log(app_id + ':APP-CALL', log_entry) # convert args_json to args_dict try: args_dict = json.loads(args_json) except: error = "%s.initExp input args_json is in improper format" % self.app_id return '{}', False, error # get list of algorithms associated with project alg_list, didSucceed, message = db.get(app_id + ':experiments', exp_uid, 'alg_list') alg_label_to_alg_id = {} alg_label_to_alg_uid = {} for algorithm in alg_list: alg_label_to_alg_id[ algorithm['alg_label']] = algorithm['alg_id'] alg_label_to_alg_uid[ algorithm['alg_label']] = algorithm['alg_uid'] algorithm_management_settings, didSucceed, message = db.get( app_id + ':experiments', exp_uid, 'algorithm_management_settings') # ASSIGN ALGORITHM TO PARTICIPANT if 'participant_uid' in args_dict: participant_uid = args_dict['participant_uid'] else: participant_uid = exp_uid participant_doc_exists, didSucceed, message = db.exists( app_id + ':participants', participant_uid, 'participant_uid') first_participant_query = not participant_doc_exists if first_participant_query: db.set(app_id + ':participants', participant_uid, 'participant_uid', participant_uid) db.set(app_id + ':participants', participant_uid, 'exp_uid', exp_uid) participant_to_algorithm_management, didSucceed, message = db.get( app_id + ':experiments', exp_uid, 'participant_to_algorithm_management') if (participant_uid == exp_uid) or ( participant_to_algorithm_management == 'one_to_many') or (first_participant_query): if algorithm_management_settings[ 'mode'] == 'fixed_proportions': proportions_list = algorithm_management_settings['params'][ 'proportions'] prop = [ prop_item['proportion'] for prop_item in proportions_list ] prop_item = numpy.random.choice(alg_list, p=prop) else: raise Exception('algorithm_management_mode : ' + algorithm_management_settings['mode'] + ' not implemented') alg_id = alg_label_to_alg_id[prop_item['alg_label']] alg_uid = alg_label_to_alg_uid[prop_item['alg_label']] alg_label = prop_item['alg_label'] if (first_participant_query) and ( participant_to_algorithm_management == 'one_to_one'): db.set(app_id + ':participants', participant_uid, 'alg_id', alg_id) db.set(app_id + ':participants', participant_uid, 'alg_uid', alg_uid) elif (participant_to_algorithm_management == 'one_to_one'): # If here, then alg_uid should already be assigned in participant doc alg_id, didSucceed, message = db.get(app_id + ':participants', participant_uid, 'alg_id') alg_uid, didSucceed, message = db.get(app_id + ':participants', participant_uid, 'alg_uid') else: raise Exception('participant_to_algorithm_management : ' + participant_to_algorithm_management + ' not implemented') # get sandboxed database for the specific app_id,alg_id,exp_uid - closing off the rest of the database to the algorithm rc = ResourceClient(app_id, exp_uid, alg_uid, db) # get specific algorithm to make calls to alg = utils.get_app_alg(self.app_id, alg_id) # call getQuery index_left, index_right, index_painted, dt = utils.timeit( alg.getQuery)(resource=rc) # check for context context_type, didSucceed, message = db.get(app_id + ':experiments', exp_uid, 'context_type') context, didSucceed, message = db.get(app_id + ':experiments', exp_uid, 'context') # log log_entry_durations = { 'exp_uid': exp_uid, 'alg_uid': alg_uid, 'task': 'getQuery', 'duration': dt } log_entry_durations.update(rc.getDurations()) meta = {'log_entry_durations': log_entry_durations} # create JSON query payload if index_left == index_painted: targets = [{ 'index': index_left, 'label': 'left', 'flag': 1 }, { 'index': index_right, 'label': 'right', 'flag': 0 }] else: targets = [{ 'index': index_left, 'label': 'left', 'flag': 0 }, { 'index': index_right, 'label': 'right', 'flag': 1 }] timestamp = str(utils.datetimeNow()) query_uid = utils.getNewUID() query = {} query['query_uid'] = query_uid query['target_indices'] = targets # save query data to database query_doc = {} query_doc.update(query) query_doc['participant_uid'] = participant_uid query_doc['alg_uid'] = alg_uid query_doc['exp_uid'] = exp_uid query_doc['alg_label'] = alg_label query_doc['timestamp_query_generated'] = timestamp db.set_doc(app_id + ':queries', query_uid, query_doc) # add context after updating query doc to avoid redundant information query['context_type'] = context_type query['context'] = context args_out = {'args': query, 'meta': meta} response_json = json.dumps(args_out) log_entry = { 'exp_uid': exp_uid, 'task': 'getQuery', 'json': response_json, 'timestamp': utils.datetimeNow() } ell.log(app_id + ':APP-RESPONSE', log_entry) return response_json, True, '' except Exception, err: error = traceback.format_exc() log_entry = { 'exp_uid': exp_uid, 'task': 'getQuery', 'error': error, 'timestamp': utils.datetimeNow(), 'args_json': args_json } ell.log(app_id + ':APP-EXCEPTION', log_entry) return '{}', False, error
def initExp(self, exp_uid, args_json, db, ell): """ initialize the project and necessary experiments Expected input (in json structure with string keys): (int) n: number of arms (float) failure_probability : confidence [optional] (list of dicts) alg_list : with fields (Defaults given by Info.get_app_default_alg_list) (string) alg_id : valid alg_id for this app_id (string) alg_label : unique identifier for algorithm (e.g. may have experiment with repeated alg_id's, but alg_labels must be unqiue, will also be used for plot legends [optional] (string) test_alg_label : must be one of the alg_label's in alg_list (Default is self) [optional] (dict) algorithm_management_settings : dictionary with fields (string) 'mode' and (dict) 'params'. mode in {'pure_exploration','explore_exploit','fixed_proportions'}. Default is 'fixed_proportions' and allocates uniform probability to each algorithm. If mode=fixed_proportions then params is a dictionary that contains the field 'proportions' which is a list of dictionaries with fields 'alg_label' and 'proportion' for all algorithms in alg_list. All proportions must be positive and sum to 1 over all algs in alg_list [optional] (string) participant_to_algorithm_management : in {'one_to_one','one_to_many'}. Default is 'one_to_many'. [optional] (string) instructions [optional] (string) debrief [optional] (int) num_tries Expected output: if error: return (JSON) '{}', (bool) False, (str) error_str else: return (JSON) '{}', (bool) True,'' Usage: initExp_response_json,didSucceed,message = app.initExp(db_API,exp_uid,initExp_args_json) Example input: initExp_args_json = {"participant_to_algorithm_management": "one_to_many", "alg_list": [{"alg_label": "BR_LilUCB", "alg_id": "BR_LilUCB", "params": {}}], "algorithm_management_settings": {"params": {"proportions": [{"alg_label": "BR_LilUCB", "proportion": 1.0}]}, "mode": "fixed_proportions"}, "failure_probability": 0.01, "n": 10} Example output: initExp_response_json = {} """ try: app_id = self.app_id # remove any reminants of an experiment if it exists didSucceed, message = db.delete_docs_with_filter( 'experiments_admin', {'exp_uid': exp_uid}) didSucceed, message = db.delete_docs_with_filter( app_id + ':experiments', {'exp_uid': exp_uid}) didSucceed, message = db.delete_docs_with_filter( app_id + ':queries', {'exp_uid': exp_uid}) didSucceed, message = db.delete_docs_with_filter( app_id + ':participants', {'exp_uid': exp_uid}) didSucceed, message = db.delete_docs_with_filter( app_id + ':algorithms', {'exp_uid': exp_uid}) didSucceed, message = ell.delete_logs_with_filter( app_id + ':APP-CALL', {'exp_uid': exp_uid}) didSucceed, message = ell.delete_logs_with_filter( app_id + ':APP-RESPONSE', {'exp_uid': exp_uid}) didSucceed, message = ell.delete_logs_with_filter( app_id + ':APP-EXCEPTION', {'exp_uid': exp_uid}) didSucceed, message = ell.delete_logs_with_filter( app_id + ':ALG-DURATION', {'exp_uid': exp_uid}) didSucceed, message = ell.delete_logs_with_filter( app_id + ':ALG-EVALUATION', {'exp_uid': exp_uid}) # add indexes (only adds them if they do not already exist) didSucceed, message = db.ensure_index('experiments_admin', {'exp_uid': 1}) didSucceed, message = db.ensure_index(app_id + ':experiments', {'exp_uid': 1}) didSucceed, message = db.ensure_index(app_id + ':queries', {'query_uid': 1}) didSucceed, message = db.ensure_index(app_id + ':queries', {'exp_uid': 1}) didSucceed, message = db.ensure_index(app_id + ':queries', {'alg_uid': 1}) didSucceed, message = db.ensure_index(app_id + ':queries', {'participant_uid': 1}) didSucceed, message = db.ensure_index(app_id + ':participants', {'participant_uid': 1}) didSucceed, message = db.ensure_index(app_id + ':participants', {'exp_uid': 1}) didSucceed, message = db.ensure_index(app_id + ':algorithms', {'alg_uid': 1}) didSucceed, message = db.ensure_index(app_id + ':algorithms', {'exp_uid': 1}) didSucceed, message = ell.ensure_index(app_id + ':APP-CALL', {'exp_uid': 1}) didSucceed, message = ell.ensure_index(app_id + ':APP-CALL', {'timestamp': 1}) didSucceed, message = ell.ensure_index(app_id + ':APP-CALL', { 'exp_uid': 1, 'timestamp': 1 }) didSucceed, message = ell.ensure_index(app_id + ':APP-CALL', { 'exp_uid': 1, 'task': 1 }) didSucceed, message = ell.ensure_index(app_id + ':APP-RESPONSE', {'exp_uid': 1}) didSucceed, message = ell.ensure_index(app_id + ':APP-RESPONSE', {'timestamp': 1}) didSucceed, message = ell.ensure_index(app_id + ':APP-RESPONSE', { 'exp_uid': 1, 'timestamp': 1 }) didSucceed, message = ell.ensure_index(app_id + ':APP-RESPONSE', { 'exp_uid': 1, 'task': 1 }) didSucceed, message = ell.ensure_index(app_id + ':APP-EXCEPTION', {'exp_uid': 1}) didSucceed, message = ell.ensure_index(app_id + ':APP-EXCEPTION', {'timestamp': 1}) didSucceed, message = ell.ensure_index(app_id + ':APP-EXCEPTION', { 'exp_uid': 1, 'timestamp': 1 }) didSucceed, message = ell.ensure_index(app_id + ':APP-EXCEPTION', { 'exp_uid': 1, 'task': 1 }) didSucceed, message = ell.ensure_index(app_id + ':ALG-DURATION', {'exp_uid': 1}) didSucceed, message = ell.ensure_index(app_id + ':ALG-DURATION', {'alg_uid': 1}) didSucceed, message = ell.ensure_index(app_id + ':ALG-DURATION', {'timestamp': 1}) didSucceed, message = ell.ensure_index(app_id + ':ALG-DURATION', { 'exp_uid': 1, 'timestamp': 1 }) didSucceed, message = ell.ensure_index(app_id + ':ALG-DURATION', { 'alg_uid': 1, 'task': 1 }) didSucceed, message = ell.ensure_index(app_id + ':ALG-EVALUATION', {'exp_uid': 1}) didSucceed, message = ell.ensure_index(app_id + ':ALG-EVALUATION', {'alg_uid': 1}) didSucceed, message = ell.ensure_index(app_id + ':ALG-EVALUATION', {'timestamp': 1}) didSucceed, message = ell.ensure_index(app_id + ':ALG-EVALUATION', { 'exp_uid': 1, 'timestamp': 1 }) db.set('experiments_admin', exp_uid, 'exp_uid', exp_uid) db.set('experiments_admin', exp_uid, 'app_id', app_id) db.set('experiments_admin', exp_uid, 'start_date', utils.datetime2str(utils.datetimeNow())) log_entry = { 'exp_uid': exp_uid, 'task': 'initExp', 'json': args_json, 'timestamp': utils.datetimeNow() } ell.log(app_id + ':APP-CALL', log_entry) # convert args_json to args_dict try: args_dict = json.loads(args_json) except: error = "%s.initExp input args_json is in improper format" % self.app_id return '{}', False, error # check for the fields that must be contained in args or error occurs necessary_fields = ['n', 'failure_probability'] for field in necessary_fields: try: args_dict[field] except KeyError: error = "%s.initExp input arguments missing field: %s" % ( self.app_id, str(field)) return '{}', False, error n = args_dict['n'] delta = args_dict['failure_probability'] if 'alg_list' in args_dict: alg_list = args_dict['alg_list'] supportedAlgs = utils.get_app_supported_algs(self.app_id) for algorithm in alg_list: if algorithm['alg_id'] not in supportedAlgs: error = "%s.initExp unsupported algorithm '%s' in alg_list" % ( self.app_id, alg_id) return '{}', False, error else: alg_list = utils.get_app_default_alg_list(self.app_id) if 'instructions' not in args_dict: instructions = utils.get_app_default_instructions(app_id) else: instructions = args_dict['instructions'] if 'debrief' not in args_dict: debrief = utils.get_app_default_instructions(app_id) else: debrief = args_dict['debrief'] if 'num_tries' not in args_dict: num_tries = utils.get_app_default_num_tries(app_id) else: num_tries = args_dict['num_tries'] if 'context_type' not in args_dict: context_type = 'none' else: context_type = args_dict['context_type'] if 'context' not in args_dict: context = '' else: context = args_dict['context'] # ALGORITHM_MANAGEMENT_MODE FORMATTING CHECK if 'algorithm_management_settings' not in args_dict: params = {} params['proportions'] = [] for algorithm in alg_list: params['proportions'].append({ 'alg_label': algorithm['alg_label'], 'proportion': 1. / len(alg_list) }) algorithm_management_settings = {} algorithm_management_settings['mode'] = 'fixed_proportions' algorithm_management_settings['params'] = params else: algorithm_management_settings = args_dict[ 'algorithm_management_settings'] try: mode = algorithm_management_settings['mode'] params = algorithm_management_settings['params'] except: error = "%s.initExp algorithm_management_settings must be a dictionary with fields 'mode' and 'params'" % ( self.app_id) return '{}', False, error if mode == 'fixed_proportions': try: algorithm_proportions_list = params['proportions'] except: error = "%s.initExp algorithm_management_settings['params'] must be a dictionary with field 'proportions'" % ( self.app_id) return '{}', False, error # check if alg_labels are properly labeled for proportion_item in algorithm_proportions_list: proportion = proportion_item['proportion'] target_alg_label = proportion_item['alg_label'] target_alg_label_in_alg_list = False for algorithm in alg_list: if algorithm['alg_label'] == target_alg_label: target_alg_label_in_alg_list = True if not target_alg_label_in_alg_list: error = "%s.initExp algorithm_management_settings['params']['proportions'] must be a list of dictionaries, each dictionary containing the fields 'alg_label' and 'proportion'. The 'alg_label' value must be one of the alg_labels in a provided alg_list and 'proportion' must be nonnegative and sum to 1 : '%s' not in provided alg_list" % ( self.app_id, target_alg_label) return '{}', False, error elif mode == 'pure_exploration': error = "%s.initExp Sorry, '%s' is not yet supported." % ( self.app_id, mode) return '{}', False, error elif mode == 'explore_exploit': error = "%s.initExp Sorry, '%s' is not yet supported." % ( self.app_id, mode) return '{}', False, error else: error = "%s.initExp unsupported algorithm_management_mode: '%s'. Must be in {'pure_exploration','explore_exploit','fixed_proportions'}" % ( self.app_id, algorithm_management_mode) return '{}', False, error # ALGORITHM_MANAGEMENT_MODE FORMATTING CHECK if 'participant_to_algorithm_management' not in args_dict: participant_to_algorithm_management = 'one_to_many' else: participant_to_algorithm_management = args_dict[ 'participant_to_algorithm_management'] if participant_to_algorithm_management not in [ 'one_to_many', 'one_to_one' ]: error = "%s.initExp unsupported participant_to_algorithm_management: '%s'. Must be in {'one_to_many','one_to_one'}" % ( self.app_id, participant_to_algorithm_management) return '{}', False, error # assign uid to each algorithm and save it for algorithm in alg_list: alg_uid = utils.getNewUID() algorithm['alg_uid'] = alg_uid db.set(app_id + ':algorithms', alg_uid, 'alg_uid', alg_uid) db.set(app_id + ':algorithms', alg_uid, 'exp_uid', exp_uid) db.set(app_id + ':experiments', exp_uid, 'exp_uid', exp_uid) db.set(app_id + ':experiments', exp_uid, 'app_id', app_id) db.set(app_id + ':experiments', exp_uid, 'n', n) db.set(app_id + ':experiments', exp_uid, 'failure_probability', delta) db.set(app_id + ':experiments', exp_uid, 'alg_list', alg_list) db.set(app_id + ':experiments', exp_uid, 'algorithm_management_settings', algorithm_management_settings) db.set(app_id + ':experiments', exp_uid, 'participant_to_algorithm_management', participant_to_algorithm_management) db.set(app_id + ':experiments', exp_uid, 'instructions', instructions) db.set(app_id + ':experiments', exp_uid, 'debrief', debrief) db.set(app_id + ':experiments', exp_uid, 'context_type', context_type) db.set(app_id + ':experiments', exp_uid, 'context', context) db.set(app_id + ':experiments', exp_uid, 'num_tries', num_tries) # now create intitialize each algorithm for algorithm in alg_list: alg_id = algorithm['alg_id'] alg_uid = algorithm['alg_uid'] db.set(app_id + ':algorithms', alg_uid, 'alg_id', alg_id) db.set(app_id + ':algorithms', alg_uid, 'alg_uid', alg_uid) db.set(app_id + ':algorithms', alg_uid, 'exp_uid', exp_uid) # get sandboxed database for the specific app_id,alg_id,exp_uid - closing off the rest of the database to the algorithm rc = ResourceClient(app_id, exp_uid, alg_uid, db) # get specific algorithm to make calls to alg = utils.get_app_alg(self.app_id, alg_id) # call initExp didSucceed, dt = utils.timeit(alg.initExp)( resource=rc, n=n, failure_probability=delta) log_entry = { 'exp_uid': exp_uid, 'alg_uid': alg_uid, 'task': 'initExp', 'duration': dt, 'timestamp': utils.datetimeNow() } ell.log(app_id + ':ALG-DURATION', log_entry) response_json = '{}' log_entry = { 'exp_uid': exp_uid, 'task': 'initExp', 'json': response_json, 'timestamp': utils.datetimeNow() } ell.log(app_id + ':APP-RESPONSE', log_entry) return response_json, True, '' except Exception, err: error = traceback.format_exc() log_entry = { 'exp_uid': exp_uid, 'task': 'initExp', 'error': error, 'timestamp': utils.datetimeNow(), 'args_json': args_json } ell.log(app_id + ':APP-EXCEPTION', log_entry) return '{}', False, error
def getQuery(self, exp_uid, args_json, db, ell): """ A request to ask which k arms to duel next Expected input (in jsonstructure with string keys): (int) k : number of objects to display [optional] (string) participant_uid : unique identifier of session for a participant answering questions (that is, an email address is not good enough as the participant could participate in multiple exp_uids so it would not be unique against all experiments), if key non-existant particpant_uid is assigned as exp_uid. Expected output (in json structure with string keys): (list) target_indices : list of k target indexes e.g. [ (int) target_index_1, ... , (int) target_index_k ] (str) query_uid : unique identifier of query (used to look up for processAnswer) """ try: app_id = self.app_id log_entry = { 'exp_uid': exp_uid, 'task': 'getQuery', 'json': args_json, 'timestamp': utils.datetimeNow() } ell.log(app_id + ':APP-CALL', log_entry) # convert args_json to args_dict try: args_dict = json.loads(args_json) except: error = "%s.initExp input args_json is in improper format" % self.app_id return '{}', False, error # get list of algorithms associated with project alg_list, didSucceed, message = db.get(app_id + ':experiments', exp_uid, 'alg_list') alg_label_to_alg_id = {} alg_label_to_alg_uid = {} for algorithm in alg_list: alg_label_to_alg_id[ algorithm['alg_label']] = algorithm['alg_id'] alg_label_to_alg_uid[ algorithm['alg_label']] = algorithm['alg_uid'] algorithm_management_settings, didSucceed, message = db.get( app_id + ':experiments', exp_uid, 'algorithm_management_settings') # ASSIGN ALGORITHM TO PARTICIPANT if 'participant_uid' in args_dict: participant_uid = args_dict['participant_uid'] else: participant_uid = exp_uid participant_doc_exists, didSucceed, message = db.exists( app_id + ':participants', participant_uid, 'participant_uid') first_participant_query = not participant_doc_exists if first_participant_query: db.set(app_id + ':participants', participant_uid, 'participant_uid', participant_uid) db.set(app_id + ':participants', participant_uid, 'exp_uid', exp_uid) participant_to_algorithm_management, didSucceed, message = db.get( app_id + ':experiments', exp_uid, 'participant_to_algorithm_management') if (participant_uid == exp_uid) or ( participant_to_algorithm_management == 'one_to_many') or (first_participant_query): if algorithm_management_settings[ 'mode'] == 'fixed_proportions': proportions_list = algorithm_management_settings['params'][ 'proportions'] prop = [ prop_item['proportion'] for prop_item in proportions_list ] prop_item = numpy.random.choice(alg_list, p=prop) else: raise Exception('algorithm_management_mode : ' + algorithm_management_settings['mode'] + ' not implemented') alg_id = alg_label_to_alg_id[prop_item['alg_label']] alg_uid = alg_label_to_alg_uid[prop_item['alg_label']] alg_label = prop_item['alg_label'] if (first_participant_query) and ( participant_to_algorithm_management == 'one_to_one'): db.set(app_id + ':participants', participant_uid, 'alg_id', alg_id) db.set(app_id + ':participants', participant_uid, 'alg_uid', alg_uid) elif (participant_to_algorithm_management == 'one_to_one'): # If here, then alg_uid should already be assigned in participant doc alg_id, didSucceed, message = db.get(app_id + ':participants', participant_uid, 'alg_id') alg_uid, didSucceed, message = db.get(app_id + ':participants', participant_uid, 'alg_uid') else: raise Exception('participant_to_algorithm_management : ' + participant_to_algorithm_management + ' not implemented') # get sandboxed database for the specific app_id,alg_id,exp_uid - closing off the rest of the database to the algorithm rc = ResourceClient(app_id, exp_uid, alg_uid, db) # get specific algorithm to make calls to alg = utils.get_app_alg(self.app_id, alg_id) # call getQuery targets, dt = utils.timeit(alg.getQuery)(resource=rc) # update targets target_list = [] for target in targets: target_list.append({'index': target}) # check for context context_type, didSucceed, message = db.get(app_id + ':experiments', exp_uid, 'context_type') context, didSucceed, message = db.get(app_id + ':experiments', exp_uid, 'context') # log log_entry_durations = { 'exp_uid': exp_uid, 'alg_uid': alg_uid, 'task': 'getQuery', 'duration': dt } log_entry_durations.update(rc.getDurations()) meta = {'log_entry_durations': log_entry_durations} # create JSON query payload timestamp = str(utils.datetimeNow()) query_uid = utils.getNewUID() query = {} query['query_uid'] = query_uid query['target_indices'] = target_list # save query data to database query_doc = {} query_doc.update(query) query_doc['participant_uid'] = participant_uid query_doc['alg_uid'] = alg_uid query_doc['exp_uid'] = exp_uid query_doc['alg_label'] = alg_label query_doc['timestamp_query_generated'] = timestamp for field in query_doc: db.set(app_id + ':queries', query_uid, field, query_doc[field]) # add context after updating query doc to avoid redundant information query['context_type'] = context_type query['context'] = context args_out = {'args': query, 'meta': meta} response_json = json.dumps(args_out) log_entry = { 'exp_uid': exp_uid, 'task': 'getQuery', 'json': response_json, 'timestamp': utils.datetimeNow() } ell.log(app_id + ':APP-RESPONSE', log_entry) return response_json, True, '' except Exception, err: error = traceback.format_exc() log_entry = { 'exp_uid': exp_uid, 'task': 'getQuery', 'error': error, 'timestamp': utils.datetimeNow(), 'args_json': args_json } ell.log(app_id + ':APP-EXCEPTION', log_entry) return '{}', False, error
def getQuery(self,exp_uid,args_json,db,ell): """ A request to ask the query: "is {center} more similar to {left} or {right}?" Expected input (in jsonstructure with string keys): [optional] (string) participant_uid : unique identifier of session for a participant answering questions (that is, an email address is not good enough as the participant could participate in multiple exp_uids so it would not be unique against all experiments), if key non-existant particpant_uid is assigned as exp_uid. Expected output (in json structure with string keys): (list) target_indices : list that stores dictionary of targets with fields: { (int) index : the index of the target of relevance (str) label : in {'left','right','center'} (int) flag : integer for algorithm's use } (str) query_uid : unique identifier of query (used to look up for reportAnswer) Usage: getQuery_response_json,didSucceed,message = app.getQuery(db_API,exp_uid,getQuery_args_json) Example input: getQuery_args_json = {"participant_uid": "ecaf1d60ab995b3c57afb3a1f3f288f0"} Example output: getQuery_response_json = {"query_uid": "a061ce00742603afc540d23e08ab77b3", "target_indices": [{"index": 19, "flag": 0, "label": "center"}, {"index": 8, "flag": 0, "label": "left"}, {"index": 15, "flag": 0, "label": "right"}]} """ try: app_id = self.app_id log_entry = { 'exp_uid':exp_uid,'task':'getQuery','json':args_json,'timestamp':utils.datetimeNow() } ell.log( app_id+':APP-CALL', log_entry ) # convert args_json to args_dict try: args_dict = json.loads(args_json) except: error = "%s.initExp input args_json is in improper format" % self.app_id return '{}',False,error # get list of algorithms associated with project alg_list,didSucceed,message = db.get(app_id+':experiments',exp_uid,'alg_list') alg_label_to_alg_id = {} alg_label_to_alg_uid = {} for algorithm in alg_list: alg_label_to_alg_id[ algorithm['alg_label'] ] = algorithm['alg_id'] alg_label_to_alg_uid[ algorithm['alg_label'] ] = algorithm['alg_uid'] algorithm_management_settings,didSucceed,message = db.get(app_id+':experiments',exp_uid,'algorithm_management_settings') # ASSIGN ALGORITHM TO PARTICIPANT if 'participant_uid' in args_dict: participant_uid = args_dict['participant_uid'] else: participant_uid = exp_uid participant_doc_exists,didSucceed,message = db.exists(app_id+':participants',participant_uid,'participant_uid') first_participant_query = not participant_doc_exists if first_participant_query: db.set(app_id+':participants',participant_uid,'participant_uid',participant_uid) db.set(app_id+':participants',participant_uid,'exp_uid',exp_uid) participant_to_algorithm_management,didSucceed,message = db.get(app_id+':experiments',exp_uid,'participant_to_algorithm_management') if (participant_uid==exp_uid) or (participant_to_algorithm_management=='one_to_many') or (first_participant_query): if algorithm_management_settings['mode']=='fixed_proportions': proportions_list = algorithm_management_settings['params']['proportions'] prop = [ prop_item['proportion'] for prop_item in proportions_list ] prop_item = numpy.random.choice(alg_list,p=prop) else: raise Exception('algorithm_management_mode : '+algorithm_management_settings['mode']+' not implemented') alg_id = alg_label_to_alg_id[ prop_item['alg_label'] ] alg_uid = alg_label_to_alg_uid[ prop_item['alg_label'] ] alg_label = prop_item['alg_label'] if (first_participant_query) and (participant_to_algorithm_management=='one_to_one'): db.set(app_id+':participants',participant_uid,'alg_id',alg_id) db.set(app_id+':participants',participant_uid,'alg_uid',alg_uid) elif (participant_to_algorithm_management=='one_to_one'): # If here, then alg_uid should already be assigned in participant doc alg_id,didSucceed,message = db.get(app_id+':participants',participant_uid,'alg_id') alg_uid,didSucceed,message = db.get(app_id+':participants',participant_uid,'alg_uid') else: raise Exception('participant_to_algorithm_management : '+participant_to_algorithm_management+' not implemented') # get sandboxed database for the specific app_id,alg_id,exp_uid - closing off the rest of the database to the algorithm rc = ResourceClient(app_id,exp_uid,alg_uid,db) # get specific algorithm to make calls to alg = utils.get_app_alg(self.app_id,alg_id) # call getQuery index_center,index_left,index_right,dt = utils.timeit(alg.getQuery)(resource=rc) log_entry_durations = { 'exp_uid':exp_uid,'alg_uid':alg_uid,'task':'getQuery','duration':dt } log_entry_durations.update( rc.getDurations() ) meta = {'log_entry_durations':log_entry_durations} # create JSON query payload timestamp = str(utils.datetimeNow()) query_uid = utils.getNewUID() query = {} query['query_uid'] = query_uid query['target_indices'] = [ {'index':index_center,'label':'center','flag':0},{'index':index_left,'label':'left','flag':0},{'index':index_right,'label':'right','flag':0} ] # save query data to database query_doc = {} query_doc.update(query) query_doc['participant_uid'] = participant_uid query_doc['alg_uid'] = alg_uid query_doc['exp_uid'] = exp_uid query_doc['alg_label'] = alg_label query_doc['timestamp_query_generated'] = timestamp for field in query_doc: db.set(app_id+':queries',query_uid,field,query_doc[field]) args_out = {'args':query,'meta':meta} response_json = json.dumps(args_out) log_entry = { 'exp_uid':exp_uid,'task':'getQuery','json':response_json,'timestamp':utils.datetimeNow() } ell.log( app_id+':APP-RESPONSE', log_entry ) return response_json,True,'' except Exception, err: error = traceback.format_exc() log_entry = { 'exp_uid':exp_uid,'task':'getQuery','error':error,'timestamp':utils.datetimeNow() } ell.log( app_id+':APP-EXCEPTION', log_entry ) return '{}',False,error