def save_model(self, folder_path, file_name=None):
if file_name is None:
file_name = self.RECOMMENDER_NAME
self._print("Saving model in file '{}'".format(folder_path +
file_name))
data_dict_to_save = {"item_pop": self.item_pop}
dataIO = DataIO(folder_path=folder_path)
dataIO.save_data(file_name=file_name,
data_dict_to_save=data_dict_to_save)
self._print("Saving complete")
def _save_split(self, save_folder_path):
if save_folder_path:
if self.allow_cold_users:
allow_cold_users_suffix = "allow_cold_users"
else:
allow_cold_users_suffix = "only_warm_users"
if self.use_validation_set:
validation_set_suffix = "use_validation_set"
else:
validation_set_suffix = "no_validation_set"
name_suffix = "_{}_{}".format(allow_cold_users_suffix,
validation_set_suffix)
split_parameters_dict = {
"k_out_value": self.k_out_value,
"allow_cold_users": self.allow_cold_users
}
dataIO = DataIO(folder_path=save_folder_path)
dataIO.save_data(data_dict_to_save=split_parameters_dict,
file_name="split_parameters" + name_suffix)
dataIO.save_data(data_dict_to_save=self.SPLIT_GLOBAL_MAPPER_DICT,
file_name="split_mappers" + name_suffix)
dataIO.save_data(data_dict_to_save=self.SPLIT_URM_DICT,
file_name="split_URM" + name_suffix)
if len(self.dataReader_object.get_loaded_ICM_names()) > 0:
dataIO.save_data(data_dict_to_save=self.SPLIT_ICM_DICT,
file_name="split_ICM" + name_suffix)
dataIO.save_data(data_dict_to_save=self.SPLIT_ICM_MAPPER_DICT,
file_name="split_ICM_mappers" + name_suffix)
def save_model(self, folder_path, file_name = None):
if file_name is None:
file_name = self.RECOMMENDER_NAME
self._print("Saving model in file '{}'".format(folder_path + file_name))
data_dict_to_save = {"USER_factors": self.USER_factors,
"ITEM_factors": self.ITEM_factors,
"use_bias": self.use_bias,
}
if self.use_bias:
data_dict_to_save["ITEM_bias"] = self.ITEM_bias
data_dict_to_save["USER_bias"] = self.USER_bias
data_dict_to_save["GLOBAL_bias"] = self.GLOBAL_bias
dataIO = DataIO(folder_path=folder_path)
dataIO.save_data(file_name=file_name, data_dict_to_save = data_dict_to_save)
self._print("Saving complete")
def _save_dataset(self, save_folder_path):
dataIO = DataIO(folder_path=save_folder_path)
dataIO.save_data(data_dict_to_save=self._LOADED_GLOBAL_MAPPER_DICT,
file_name="dataset_global_mappers")
dataIO.save_data(data_dict_to_save=self._LOADED_URM_DICT,
file_name="dataset_URM")
if len(self.get_loaded_ICM_names()) > 0:
dataIO.save_data(data_dict_to_save=self._LOADED_ICM_DICT,
file_name="dataset_ICM")
dataIO.save_data(data_dict_to_save=self._LOADED_ICM_MAPPER_DICT,
file_name="dataset_ICM_mappers")
class CrossSearchFixedObject(object):
ALGORITHM_NAME = "CrossSearchFixedObject"
# Value to be assigned to invalid configuration or if an Exception is raised
INVALID_CONFIG_VALUE = np.finfo(np.float16).max
def __init__(self, recommender_object_list,
evaluator_validation_list: List[Evaluator] = None,
verbose=True):
super(CrossSearchFixedObject, self).__init__()
self.recommender_object_list = recommender_object_list
self.verbose = verbose
self.log_file = None
self.results_test_best = {}
self.parameter_dictionary_best = {}
self.evaluator_validation_list: List[Evaluator] = evaluator_validation_list
def search(self, recommender_input_args_list,
parameter_search_space,
metric_to_optimize="MAP",
n_cases=None,
output_folder_path=None,
output_file_name_root=None,
parallelize=False,
save_metadata=True,
):
raise NotImplementedError("Function search not implemented for this class")
def _set_search_attributes(self, recommender_input_args,
metric_to_optimize,
output_folder_path,
output_file_name_root,
resume_from_saved,
save_metadata,
n_cases):
self.output_folder_path = output_folder_path
self.output_file_name_root = output_file_name_root
# If directory does not exist, create
if not os.path.exists(self.output_folder_path):
os.makedirs(self.output_folder_path)
self.log_file = open(
self.output_folder_path + self.output_file_name_root + "_{}.txt".format(self.ALGORITHM_NAME), "a")
self.recommender_input_args_list = recommender_input_args
self.metric_to_optimize = metric_to_optimize
self.resume_from_saved = resume_from_saved
self.save_metadata = save_metadata
self.model_counter = 0
self._init_metadata_dict(n_cases=n_cases)
if self.save_metadata:
self.dataIO = DataIO(folder_path=self.output_folder_path)
def _init_metadata_dict(self, n_cases):
self.metadata_dict = {"algorithm_name_search": self.ALGORITHM_NAME,
"algorithm_name_recommender": self.recommender_object_list[0].RECOMMENDER_NAME,
"exception_list": [None] * n_cases,
"hyperparameters_list": [None] * n_cases,
"hyperparameters_best": None,
"hyperparameters_best_index": None,
"result_on_validation_list": [None] * n_cases,
"result_on_validation_best": None,
"std_result_on_validation_list": [None] * n_cases,
"std_result_on_validation_best": None,
"time_on_train_list": [None] * n_cases,
"time_on_train_total": 0.0,
"time_on_train_avg": 0.0,
"time_on_validation_list": [None] * n_cases,
"time_on_validation_total": 0.0,
"time_on_validation_avg": 0.0,
"time_on_test_list": [None] * n_cases,
"time_on_test_total": 0.0,
"time_on_test_avg": 0.0,
"result_on_last": None,
"time_on_last_train": None,
"time_on_last_test": None,
}
def _print(self, string):
if self.verbose:
print(string)
def _write_log(self, string):
self._print(string)
if self.log_file is not None:
self.log_file.write(string)
self.log_file.flush()
def _fit_model(self, split_index, current_fit_parameters, recommender_input_args):
start_time = time.time()
# Construct a new recommender instance
recommender_instance = self.recommender_object_list[split_index].copy()
recommender_instance.fit(*recommender_input_args.FIT_POSITIONAL_ARGS,
**recommender_input_args.FIT_KEYWORD_ARGS,
**current_fit_parameters)
train_time = time.time() - start_time
return recommender_instance, train_time
def _evaluate_on_validation(self, current_fit_parameters):
result_dicts = []
evaluation_times = np.zeros(len(self.evaluator_validation_list))
train_times = np.zeros(len(self.evaluator_validation_list))
self._print("{}: Testing config: {}".format(self.ALGORITHM_NAME, current_fit_parameters))
for i, evaluator in enumerate(self.evaluator_validation_list):
recommender_instance, train_times[i] = self._fit_model(i, current_fit_parameters,
self.recommender_input_args_list[i])
start_time = time.time()
# Evaluate recommender and get results for the first cutoff
mean_result_dict, _ = self.evaluator_validation_list[i].evaluateRecommender(recommender_instance)
mean_result_dict = mean_result_dict[list(mean_result_dict.keys())[0]]
result_dicts.append(mean_result_dict)
evaluation_times[i] = time.time() - start_time
mean_result_dict, std_result_dict = compute_mean_std_result_dict(result_dicts)
evaluation_time = evaluation_times.sum() / evaluation_times.size
train_time = train_times.sum() / train_times.size
result_string = get_result_string(mean_result_dict, std_result_dict, n_mean_decimals=6, n_std_decimals=4)
return mean_result_dict, std_result_dict, result_string, train_time, evaluation_time
def _objective_function(self, current_fit_parameters_dict):
try:
self.metadata_dict["hyperparameters_list"][self.model_counter] = current_fit_parameters_dict.copy()
mean_result_dict, std_result_dict, result_string, train_time, evaluation_time = self._evaluate_on_validation(
current_fit_parameters_dict)
current_result = - mean_result_dict[self.metric_to_optimize]
if self.metadata_dict["result_on_validation_best"] is None:
new_best_config_found = True
else:
best_solution_val = self.metadata_dict["result_on_validation_best"][self.metric_to_optimize]
new_best_config_found = best_solution_val < mean_result_dict[self.metric_to_optimize]
if new_best_config_found:
self._write_log("{}: New best config found. Config {}: {} - results: {}\n".format(self.ALGORITHM_NAME,
self.model_counter,
current_fit_parameters_dict,
result_string))
else:
self._write_log("{}: Config {} is suboptimal. Config: {} - results: {}\n".format(self.ALGORITHM_NAME,
self.model_counter,
current_fit_parameters_dict,
result_string))
if current_result >= self.INVALID_CONFIG_VALUE:
self._write_log(
"{}: WARNING! Config {} returned a value equal or worse than the default value to be assigned to invalid configurations."
" If no better valid configuration is found, this parameter search may produce an invalid result.\n")
self.metadata_dict["result_on_validation_list"][self.model_counter] = mean_result_dict.copy()
self.metadata_dict["std_result_on_validation_list"][self.model_counter] = std_result_dict.copy()
self.metadata_dict["time_on_train_list"][self.model_counter] = train_time
self.metadata_dict["time_on_validation_list"][self.model_counter] = evaluation_time
self.metadata_dict["time_on_train_total"], self.metadata_dict["time_on_train_avg"] = \
_compute_avg_time_non_none_values(self.metadata_dict["time_on_train_list"])
self.metadata_dict["time_on_validation_total"], self.metadata_dict["time_on_validation_avg"] = \
_compute_avg_time_non_none_values(self.metadata_dict["time_on_validation_list"])
if new_best_config_found:
self.metadata_dict["hyperparameters_best"] = current_fit_parameters_dict.copy()
self.metadata_dict["hyperparameters_best_index"] = self.model_counter
self.metadata_dict["result_on_validation_best"] = mean_result_dict.copy()
self.metadata_dict["std_result_on_validation_best"] = std_result_dict.copy()
except (KeyboardInterrupt, SystemExit) as e:
# If getting a interrupt, terminate without saving the exception
raise e
except:
# Catch any error: Exception, Tensorflow errors etc...
traceback_string = traceback.format_exc()
self._write_log("{}: Config {} Exception. Config: {} - Exception: {}\n".format(self.ALGORITHM_NAME,
self.model_counter,
current_fit_parameters_dict,
traceback_string))
self.metadata_dict["exception_list"][self.model_counter] = traceback_string
# Assign to this configuration the worst possible score
# Being a minimization problem, set it to the max value of a float
current_result = + self.INVALID_CONFIG_VALUE
traceback.print_exc()
if self.save_metadata:
self.dataIO.save_data(data_dict_to_save=self.metadata_dict.copy(),
file_name=self.output_file_name_root + "_metadata")
self.model_counter += 1
return current_result
class SearchAbstractClass(object):
ALGORITHM_NAME = "SearchAbstractClass"
# Available values for the save_model attribute
_SAVE_MODEL_VALUES = ["all", "best", "last", "no"]
# Value to be assigned to invalid configuration or if an Exception is raised
INVALID_CONFIG_VALUE = np.finfo(np.float16).max
def __init__(self,
recommender_class,
evaluator_validation=None,
evaluator_test=None,
verbose=True):
super(SearchAbstractClass, self).__init__()
self.recommender_class = recommender_class
self.verbose = verbose
self.log_file = None
self.results_test_best = {}
self.parameter_dictionary_best = {}
self.evaluator_validation = evaluator_validation
if evaluator_test is None:
self.evaluator_test = None
else:
self.evaluator_test = evaluator_test
def search(
self,
recommender_input_args,
parameter_search_space,
metric_to_optimize="MAP",
n_cases=None,
output_folder_path=None,
output_file_name_root=None,
parallelize=False,
save_model="best",
evaluate_on_test_each_best_solution=True,
save_metadata=True,
):
raise NotImplementedError(
"Function search not implemented for this class")
def _set_search_attributes(self, recommender_input_args,
recommender_input_args_last_test,
metric_to_optimize, output_folder_path,
output_file_name_root, resume_from_saved,
save_metadata, save_model,
evaluate_on_test_each_best_solution, n_cases):
if save_model not in self._SAVE_MODEL_VALUES:
raise ValueError(
"{}: parameter save_model must be in '{}', provided was '{}'.".
format(self.ALGORITHM_NAME, self._SAVE_MODEL_VALUES,
save_model))
self.output_folder_path = output_folder_path
self.output_file_name_root = output_file_name_root
# If directory does not exist, create
if not os.path.exists(self.output_folder_path):
os.makedirs(self.output_folder_path)
self.log_file = open(
self.output_folder_path + self.output_file_name_root +
"_{}.txt".format(self.ALGORITHM_NAME), "a")
if save_model == "last" and recommender_input_args_last_test is None:
self._write_log(
"{}: parameter save_model is 'last' but no recommender_input_args_last_test provided, saving best model on train data alone."
.format(self.ALGORITHM_NAME))
save_model = "best"
self.recommender_input_args = recommender_input_args
self.recommender_input_args_last_test = recommender_input_args_last_test
self.metric_to_optimize = metric_to_optimize
self.save_model = save_model
self.resume_from_saved = resume_from_saved
self.save_metadata = save_metadata
self.evaluate_on_test_each_best_solution = evaluate_on_test_each_best_solution
self.model_counter = 0
self._init_metadata_dict(n_cases=n_cases)
if self.save_metadata:
self.dataIO = DataIO(folder_path=self.output_folder_path)
def _init_metadata_dict(self, n_cases):
self.metadata_dict = {
"algorithm_name_search": self.ALGORITHM_NAME,
"algorithm_name_recommender":
self.recommender_class.RECOMMENDER_NAME,
"exception_list": [None] * n_cases,
"hyperparameters_list": [None] * n_cases,
"hyperparameters_best": None,
"hyperparameters_best_index": None,
"result_on_validation_list": [None] * n_cases,
"result_on_validation_best": None,
"result_on_test_list": [None] * n_cases,
"result_on_test_best": None,
"time_on_train_list": [None] * n_cases,
"time_on_train_total": 0.0,
"time_on_train_avg": 0.0,
"time_on_validation_list": [None] * n_cases,
"time_on_validation_total": 0.0,
"time_on_validation_avg": 0.0,
"time_on_test_list": [None] * n_cases,
"time_on_test_total": 0.0,
"time_on_test_avg": 0.0,
"result_on_last": None,
"time_on_last_train": None,
"time_on_last_test": None,
}
def _print(self, string):
if self.verbose:
print(string)
def _write_log(self, string):
self._print(string)
if self.log_file is not None:
self.log_file.write(string)
self.log_file.flush()
def _fit_model(self, current_fit_parameters):
start_time = time.time()
# Construct a new recommender instance
recommender_instance = self.recommender_class(
*self.recommender_input_args.CONSTRUCTOR_POSITIONAL_ARGS,
**self.recommender_input_args.CONSTRUCTOR_KEYWORD_ARGS)
self._print("{}: Testing config: {}".format(self.ALGORITHM_NAME,
current_fit_parameters))
recommender_instance.fit(
*self.recommender_input_args.FIT_POSITIONAL_ARGS,
**self.recommender_input_args.FIT_KEYWORD_ARGS,
**current_fit_parameters)
train_time = time.time() - start_time
return recommender_instance, train_time
def _evaluate_on_validation(self, current_fit_parameters):
recommender_instance, train_time = self._fit_model(
current_fit_parameters)
start_time = time.time()
# Evaluate recommender and get results for the first cutoff
result_dict, _ = self.evaluator_validation.evaluateRecommender(
recommender_instance)
result_dict = result_dict[list(result_dict.keys())[0]]
evaluation_time = time.time() - start_time
result_string = get_result_string_evaluate_on_validation(result_dict,
n_decimals=7)
return result_dict, result_string, recommender_instance, train_time, evaluation_time
def _evaluate_on_test(self,
recommender_instance,
current_fit_parameters_dict,
print_log=True):
start_time = time.time()
# Evaluate recommender and get results for the first cutoff
result_dict, result_string = self.evaluator_test.evaluateRecommender(
recommender_instance)
evaluation_test_time = time.time() - start_time
if print_log:
self._write_log(
"{}: Best config evaluated with evaluator_test. Config: {} - results:\n{}\n"
.format(self.ALGORITHM_NAME, current_fit_parameters_dict,
result_string))
return result_dict, result_string, evaluation_test_time
def _evaluate_on_test_with_data_last(self):
start_time = time.time()
# Construct a new recommender instance
recommender_instance = self.recommender_class(
*self.recommender_input_args_last_test.CONSTRUCTOR_POSITIONAL_ARGS,
**self.recommender_input_args_last_test.CONSTRUCTOR_KEYWORD_ARGS)
# Check if last was already evaluated
if self.resume_from_saved:
result_on_last_saved_flag = self.metadata_dict["result_on_last"] is not None and \
self.metadata_dict["time_on_last_train"] is not None and \
self.metadata_dict["time_on_last_test"] is not None
if result_on_last_saved_flag:
self._print(
"{}: Resuming '{}'... Result on last already available.".
format(self.ALGORITHM_NAME, self.output_file_name_root))
return
self._print(
"{}: Evaluation with constructor data for final test. Using best config: {}"
.format(self.ALGORITHM_NAME,
self.metadata_dict["hyperparameters_best"]))
# Use the hyperparameters that have been saved
assert self.metadata_dict[
"hyperparameters_best"] is not None, "{}: Best hyperparameters not available, the search might have failed.".format(
self.ALGORITHM_NAME)
fit_keyword_args = self.metadata_dict["hyperparameters_best"].copy()
fit_keyword_args = {
**fit_keyword_args,
**self.recommender_input_args_last_test.FIT_KEYWORD_ARGS
}
recommender_instance.fit(
*self.recommender_input_args_last_test.FIT_POSITIONAL_ARGS,
**fit_keyword_args)
train_time = time.time() - start_time
result_dict_test, result_string, evaluation_test_time = self._evaluate_on_test(
recommender_instance, fit_keyword_args, print_log=False)
self._write_log(
"{}: Best config evaluated with evaluator_test with constructor data for final test. Config: {} - results:\n{}\n"
.format(self.ALGORITHM_NAME,
self.metadata_dict["hyperparameters_best"], result_string))
self.metadata_dict["result_on_last"] = result_dict_test
self.metadata_dict["time_on_last_train"] = train_time
self.metadata_dict["time_on_last_test"] = evaluation_test_time
if self.save_metadata:
self.dataIO.save_data(data_dict_to_save=self.metadata_dict.copy(),
file_name=self.output_file_name_root +
"_metadata")
if self.save_model in ["all", "best", "last"]:
self._print("{}: Saving model in {}\n".format(
self.ALGORITHM_NAME,
self.output_folder_path + self.output_file_name_root))
recommender_instance.save_model(
self.output_folder_path,
file_name=self.output_file_name_root + "_best_model_last")
def _objective_function(self, current_fit_parameters_dict):
try:
self.metadata_dict["hyperparameters_list"][
self.model_counter] = current_fit_parameters_dict.copy()
result_dict, result_string, recommender_instance, train_time, evaluation_time = self._evaluate_on_validation(
current_fit_parameters_dict)
current_result = -result_dict[self.metric_to_optimize]
# If the recommender uses Earlystopping, get the selected number of epochs
if isinstance(recommender_instance,
Incremental_Training_Early_Stopping):
n_epochs_early_stopping_dict = recommender_instance.get_early_stopping_final_epochs_dict(
)
current_fit_parameters_dict = current_fit_parameters_dict.copy(
)
for epoch_label in n_epochs_early_stopping_dict.keys():
epoch_value = n_epochs_early_stopping_dict[epoch_label]
current_fit_parameters_dict[epoch_label] = epoch_value
# Always save best model separately
if self.save_model in ["all"]:
self._print("{}: Saving model in {}\n".format(
self.ALGORITHM_NAME,
self.output_folder_path + self.output_file_name_root))
recommender_instance.save_model(
self.output_folder_path,
file_name=self.output_file_name_root +
"_model_{}".format(self.model_counter))
if self.metadata_dict["result_on_validation_best"] is None:
new_best_config_found = True
else:
best_solution_val = self.metadata_dict[
"result_on_validation_best"][self.metric_to_optimize]
new_best_config_found = best_solution_val < result_dict[
self.metric_to_optimize]
if new_best_config_found:
self._write_log(
"{}: New best config found. Config {}: {} - results: {}\n".
format(self.ALGORITHM_NAME, self.model_counter,
current_fit_parameters_dict, result_string))
if self.save_model in ["all", "best"]:
self._print("{}: Saving model in {}\n".format(
self.ALGORITHM_NAME,
self.output_folder_path + self.output_file_name_root))
recommender_instance.save_model(
self.output_folder_path,
file_name=self.output_file_name_root + "_best_model")
if self.evaluator_test is not None and self.evaluate_on_test_each_best_solution:
result_dict_test, _, evaluation_test_time = self._evaluate_on_test(
recommender_instance,
current_fit_parameters_dict,
print_log=True)
else:
self._write_log(
"{}: Config {} is suboptimal. Config: {} - results: {}\n".
format(self.ALGORITHM_NAME, self.model_counter,
current_fit_parameters_dict, result_string))
if current_result >= self.INVALID_CONFIG_VALUE:
self._write_log(
"{}: WARNING! Config {} returned a value equal or worse than the default value to be assigned to invalid configurations."
" If no better valid configuration is found, this parameter search may produce an invalid result.\n"
)
self.metadata_dict["result_on_validation_list"][
self.model_counter] = result_dict.copy()
self.metadata_dict["time_on_train_list"][
self.model_counter] = train_time
self.metadata_dict["time_on_validation_list"][
self.model_counter] = evaluation_time
self.metadata_dict["time_on_train_total"], self.metadata_dict["time_on_train_avg"] = \
_compute_avg_time_non_none_values(self.metadata_dict["time_on_train_list"])
self.metadata_dict["time_on_validation_total"], self.metadata_dict["time_on_validation_avg"] = \
_compute_avg_time_non_none_values(self.metadata_dict["time_on_validation_list"])
if new_best_config_found:
self.metadata_dict[
"hyperparameters_best"] = current_fit_parameters_dict.copy(
)
self.metadata_dict[
"hyperparameters_best_index"] = self.model_counter
self.metadata_dict[
"result_on_validation_best"] = result_dict.copy()
if self.evaluator_test is not None and self.evaluate_on_test_each_best_solution:
self.metadata_dict[
"result_on_test_best"] = result_dict_test.copy()
self.metadata_dict["result_on_test_list"][
self.model_counter] = result_dict_test.copy()
self.metadata_dict["time_on_test_list"][
self.model_counter] = evaluation_test_time
self.metadata_dict["time_on_test_total"], self.metadata_dict["time_on_test_avg"] = \
_compute_avg_time_non_none_values(self.metadata_dict["time_on_test_list"])
except (KeyboardInterrupt, SystemExit) as e:
# If getting a interrupt, terminate without saving the exception
raise e
except:
# Catch any error: Exception, Tensorflow errors etc...
traceback_string = traceback.format_exc()
self._write_log(
"{}: Config {} Exception. Config: {} - Exception: {}\n".format(
self.ALGORITHM_NAME, self.model_counter,
current_fit_parameters_dict, traceback_string))
self.metadata_dict["exception_list"][
self.model_counter] = traceback_string
# Assign to this configuration the worst possible score
# Being a minimization problem, set it to the max value of a float
current_result = +self.INVALID_CONFIG_VALUE
traceback.print_exc()
if self.save_metadata:
self.dataIO.save_data(data_dict_to_save=self.metadata_dict.copy(),
file_name=self.output_file_name_root +
"_metadata")
self.model_counter += 1
return current_result
def _split_data_from_original_dataset(self, save_folder_path):
self.dataReader_object.load_data()
URM = self.dataReader_object.get_URM_all()
# Managing data reader
self.SPLIT_GLOBAL_MAPPER_DICT = {}
for mapper_name, mapper_object in self.dataReader_object.get_loaded_Global_mappers().items():
self.SPLIT_GLOBAL_MAPPER_DICT[mapper_name] = mapper_object.copy()
URM = sps.csr_matrix(URM)
if not self.allow_cold_users:
user_interactions = np.ediff1d(URM.indptr)
user_to_preserve = user_interactions >= self.n_folds
user_to_remove = np.logical_not(user_to_preserve)
print(
"DataSplitter_Warm: Removing {} of {} users because they have less interactions than the number of folds".format(
URM.shape[0] - user_to_preserve.sum(), URM.shape[0]))
URM = URM[user_to_preserve, :]
self.SPLIT_GLOBAL_MAPPER_DICT["user_original_ID_to_index"] = reconcile_mapper_with_removed_tokens(
self.SPLIT_GLOBAL_MAPPER_DICT["user_original_ID_to_index"],
np.arange(0, len(user_to_remove), dtype=np.int)[user_to_remove])
self.n_users, self.n_items = URM.shape
URM = sps.csr_matrix(URM)
# Create empty URM for each fold
self.fold_split = {}
for fold_index in range(self.n_folds):
self.fold_split[fold_index] = {}
self.fold_split[fold_index]["URM"] = sps.coo_matrix(URM.shape)
URM_fold_object = self.fold_split[fold_index]["URM"]
# List.extend is waaaay faster than numpy.concatenate
URM_fold_object.row = []
URM_fold_object.col = []
URM_fold_object.data = []
for user_id in range(self.n_users):
start_user_position = URM.indptr[user_id]
end_user_position = URM.indptr[user_id + 1]
user_profile = URM.indices[start_user_position:end_user_position]
indices_to_suffle = np.arange(len(user_profile), dtype=np.int)
np.random.shuffle(indices_to_suffle)
user_profile = user_profile[indices_to_suffle]
user_interactions = URM.data[start_user_position:end_user_position][indices_to_suffle]
# interactions_per_fold is a float number, to auto-adjust fold size
interactions_per_fold = len(user_profile) / self.n_folds
for fold_index in range(self.n_folds):
start_pos = int(interactions_per_fold * fold_index)
end_pos = int(interactions_per_fold * (fold_index + 1))
if fold_index == self.n_folds - 1:
end_pos = len(user_profile)
current_fold_user_profile = user_profile[start_pos:end_pos]
current_fold_user_interactions = user_interactions[start_pos:end_pos]
URM_fold_object = self.fold_split[fold_index]["URM"]
URM_fold_object.row.extend([user_id] * len(current_fold_user_profile))
URM_fold_object.col.extend(current_fold_user_profile)
URM_fold_object.data.extend(current_fold_user_interactions)
for fold_index in range(self.n_folds):
URM_fold_object = self.fold_split[fold_index]["URM"]
URM_fold_object.row = np.array(URM_fold_object.row, dtype=np.int)
URM_fold_object.col = np.array(URM_fold_object.col, dtype=np.int)
URM_fold_object.data = np.array(URM_fold_object.data, dtype=np.float)
self.fold_split[fold_index]["URM"] = sps.csr_matrix(URM_fold_object)
self.fold_split[fold_index]["items_in_fold"] = np.arange(0, self.n_items, dtype=np.int)
fold_dict_to_save = {"fold_split": self.fold_split,
"n_folds": self.n_folds,
"n_items": self.n_items,
"n_users": self.n_users,
"allow_cold_users": self.allow_cold_users,
}
if self.allow_cold_users:
allow_user = "******"
else:
allow_user = "******"
pickle.dump(fold_dict_to_save,
open(save_folder_path + "URM_{}_fold_split_{}".format(self.n_folds, allow_user), "wb"),
protocol=pickle.HIGHEST_PROTOCOL)
for ICM_name in self.dataReader_object.get_loaded_ICM_names():
pickle.dump(self.dataReader_object.get_ICM_from_name(ICM_name),
open(save_folder_path + "{}".format(ICM_name), "wb"),
protocol=pickle.HIGHEST_PROTOCOL)
pickle.dump(self.dataReader_object.get_ICM_feature_to_index_mapper_from_name(ICM_name),
open(save_folder_path + "tokenToFeatureMapper_{}".format(ICM_name), "wb"),
protocol=pickle.HIGHEST_PROTOCOL)
# MAPPER MANAGING
if self.allow_cold_users:
allow_cold_users_suffix = "allow_cold_users"
else:
allow_cold_users_suffix = "only_warm_users"
name_suffix = "_{}".format(allow_cold_users_suffix)
dataIO = DataIO(folder_path=save_folder_path)
dataIO.save_data(data_dict_to_save=self.SPLIT_GLOBAL_MAPPER_DICT,
file_name="split_mappers" + name_suffix)
print("DataSplitter: Split complete")
def _save_dataset(self, save_folder_path):
"""
Saves all URM, ICM and UCM
:param save_folder_path:
:return:
"""
dataIO = DataIO(folder_path=save_folder_path)
dataIO.save_data(data_dict_to_save=self._LOADED_GLOBAL_MAPPER_DICT,
file_name="dataset_global_mappers")
dataIO.save_data(data_dict_to_save=self._LOADED_URM_DICT,
file_name="dataset_URM")
if len(self.get_loaded_ICM_names()) > 0:
dataIO.save_data(data_dict_to_save=self._LOADED_ICM_DICT,
file_name="dataset_ICM")
dataIO.save_data(data_dict_to_save=self._LOADED_ICM_MAPPER_DICT,
file_name="dataset_ICM_mappers")
if len(self.get_loaded_UCM_names()) > 0:
dataIO.save_data(data_dict_to_save=self._LOADED_UCM_DICT,
file_name="dataset_UCM")
dataIO.save_data(data_dict_to_save=self._LOADED_UCM_MAPPER_DICT,
file_name="dataset_UCM_mappers")