def POST(self):
super(UpdateLeadController, self).POST()
params = web.input()
lead = self.store.get_item(Lead, dict(lead_url=params.lead_url, date_created=int(params.date_created)))
source = self.store.get_item(Source, dict(source_name=lead.source_name))
parser = find_class(source.parser)()
lead = parser.populate_lead(lead)
self.store.update(lead)
return json.dumps(lead._asdict())
def __init__(self, task_name, parameters=None, other_options=None):
self.count = -1
task_class = utils.find_class(
task_name,
class_folder='tasks',
inherited_class=Task,
excluded_files=['taskTemplate', '__init__', 'taskGenerator'])
valid_task_args = utils.get_class_args(task_class)
self.task_class = task_class
if not parameters:
parameters = {}
parameter_keys = list(parameters.keys())
for p in parameter_keys:
if p not in valid_task_args:
raise KeyError(
'{} is not a valid property for model ``{}``. Use {}'.
format(p, task_name, valid_task_args))
parameter_combinations = []
for p in utils.listMergeGen(*list(parameters.values())):
pc = {k: copy.copy(v) for k, v in zip(parameter_keys, p)}
parameter_combinations.append(pc)
self.parameter_combinations = parameter_combinations
if other_options:
checked_options = {}
for k, v in other_options.items():
if k not in valid_task_args:
raise KeyError(
'{} is not a valid property for task ``{}``. Use {}'.
format(k, task_name, valid_task_args))
elif k in parameter_keys:
warnings.warn(
"task parameter {} has been defined twice".format(k))
else:
checked_options[k] = v
self.other_options = checked_options
else:
self.other_options = {}
if parameter_combinations:
self.count_max = len(parameter_combinations)
else:
self.count_max = 1
def create_model(session, exp_settings, data_set, forward_only):
"""Create model and initialize or load parameters in session.
Args:
session: (tf.Session) The session used to run tensorflow models
exp_settings: (dictionary) The dictionary containing the model settings.
data_set: (Raw_data) The dataset used to build the input layer.
forward_only: Set true to conduct prediction only, false to conduct training.
"""
model = utils.find_class(exp_settings['learning_algorithm'])(data_set,
exp_settings,
forward_only)
ckpt = tf.train.get_checkpoint_state(FLAGS.model_dir)
if ckpt:
print("Reading model parameters from %s" % ckpt.model_checkpoint_path)
model.saver.restore(session, ckpt.model_checkpoint_path)
else:
print("Created model with fresh parameters.")
session.run(tf.global_variables_initializer())
return model
def get_ranking_scores(self, input_id_list, is_training=False, scope=None):
"""Compute ranking scores with the given inputs.
Args:
input_id_list: (list<tf.Tensor>) A list of tensors containing document ids.
Each tensor must have a shape of [None].
is_training: (bool) A flag indicating whether the model is running in training mode.
scope: (string) The name of the variable scope.
Returns:
A tensor with the same shape of input_docids.
"""
with tf.variable_scope(scope or "ranking_model"):
# Build feature padding
PAD_embed = tf.zeros([1,self.feature_size],dtype=tf.float32)
letor_features = tf.concat(axis=0,values=[self.letor_features, PAD_embed])
input_feature_list = []
model = utils.find_class(self.exp_settings['ranking_model'])(self.exp_settings['ranking_model_hparams'])
for i in range(len(input_id_list)):
input_feature_list.append(tf.nn.embedding_lookup(letor_features, input_id_list[i]))
return model.build(input_feature_list, is_training)
def __init__(self, model_name, parameters=None, other_options=None):
self.count = -1
if parameters is None:
parameters = {}
if other_options is None:
other_options = {}
model_class = utils.find_class(
model_name,
class_folder='model',
inherited_class=Model,
excluded_files=['modelTemplate', '__init__', 'modelGenerator'])
valid_model_args = utils.get_class_args(model_class)
valid_args = copy.copy(valid_model_args)
if 'stimulus_shaper_name' in parameters:
raise NotImplementedError(
"This system has not been created for changing stimulus shapers. Please put it in the ``other_options``"
)
stimulus_shaper_name = other_options.pop('stimulus_shaper_name', None)
if stimulus_shaper_name:
stimFunc = utils.find_class(
stimulus_shaper_name,
class_folder='tasks',
inherited_class=Stimulus,
excluded_files=['taskTemplate', '__init__', 'taskGenerator'])
valid_stimulus_args = utils.get_class_attributes(
stimFunc, ignore=['processStimulus'])
valid_args.extend(valid_stimulus_args)
else:
stimFunc = None
if 'reward_shaper_name' in parameters:
raise NotImplementedError(
"This system has not been created for changing reward shapers. Please put it in the ``other_options``"
)
reward_shaper_name = other_options.pop('reward_shaper_name', None)
if reward_shaper_name:
rewardFunc = utils.find_class(
reward_shaper_name,
class_folder='tasks',
inherited_class=Rewards,
excluded_files=['taskTemplate', '__init__', 'taskGenerator'])
valid_reward_args = utils.get_class_attributes(
rewardFunc, ignore=['processFeedback'])
valid_args.extend(valid_reward_args)
else:
rewardFunc = None
if 'decision_function_name' in parameters:
raise NotImplementedError(
"This system has not been created for changing decision functions. Please put it in the ``other_options``"
)
decision_function_name = other_options.pop('decision_function_name',
None)
if decision_function_name:
decisionFunc = utils.find_function(decision_function_name,
'model/decision',
excluded_files=['__init__'])
valid_decision_args = utils.get_function_args(decisionFunc)
valid_args.extend(valid_decision_args)
else:
decisionFunc = None
self.model_class = model_class
if not parameters:
parameters = {}
parameter_keys = list(parameters.keys())
for p in parameter_keys:
if p not in valid_args and len(
model_class.pattern_parameters_match(p)) == 0:
raise KeyError(
'{} is not a valid property for model ``{}``. Those available are {}'
.format(p, model_name, valid_args))
parameter_combinations = []
for p in utils.listMergeGen(*list(parameters.values())):
pc = {k: copy.copy(v) for k, v in zip(parameter_keys, p)}
parameter_combinations.append(pc)
self.parameter_combinations = parameter_combinations
if other_options:
checked_options = {}
for k, v in other_options.items():
if k not in valid_args:
raise KeyError(
'{} is not a valid property for model ``{}``. Those available are {}'
.format(k, model_name, valid_args))
elif k in parameter_keys:
warnings.warn(
"model parameter {} has been defined twice".format(k))
else:
checked_options[k] = v
self.other_options = checked_options
self.other_options[
'stimulus_shaper_properties'] = valid_stimulus_args
self.other_options['reward_shaper_properties'] = valid_reward_args
self.other_options[
'decision_function_properties'] = valid_decision_args
else:
self.other_options = {}
if stimFunc:
self.other_options['stimulus_shaper'] = stimFunc
if rewardFunc:
self.other_options['reward_shaper'] = rewardFunc
if decisionFunc:
self.other_options['decision_function'] = decisionFunc
if parameter_combinations:
self.count_max = len(parameter_combinations)
else:
self.count_max = 1
def __init__(self, data_set, exp_settings, forward_only=False):
"""Create the model.
Args:
data_set: (Raw_data) The dataset used to build the input layer.
exp_settings: (dictionary) The dictionary containing the model settings.
forward_only: Set true to conduct prediction only, false to conduct training.
"""
self.hparams = tf.contrib.training.HParams(
propensity_estimator_type=
'utils.propensity_estimator.RandomizedPropensityEstimator',
propensity_estimator_json=
'./example/PropensityEstimator/randomized_pbm_0.1_1.0_4_1.0.json', # the setting file for the predefined click models.
learning_rate=0.05, # Learning rate.
max_gradient_norm=5.0, # Clip gradients to this norm.
loss_func=
'click_weighted_softmax_cross_entropy', # Select Loss function
l2_loss=0.0, # Set strength for L2 regularization.
grad_strategy='ada', # Select gradient strategy
)
print(exp_settings['learning_algorithm_hparams'])
self.hparams.parse(exp_settings['learning_algorithm_hparams'])
self.exp_settings = exp_settings
self.propensity_estimator = utils.find_class(
self.hparams.propensity_estimator_type)(
self.hparams.propensity_estimator_json)
self.max_candidate_num = exp_settings['max_candidate_num']
self.feature_size = data_set.feature_size
self.learning_rate = tf.Variable(float(self.hparams.learning_rate),
trainable=False)
self.global_step = tf.Variable(0, trainable=False)
# Feeds for inputs.
self.is_training = tf.placeholder(tf.bool, name="is_train")
self.docid_inputs = [] # a list of top documents
self.letor_features = tf.placeholder(
tf.float32, shape=[None, self.feature_size],
name="letor_features") # the letor features for the documents
self.labels = [] # the labels for the documents (e.g., clicks)
for i in range(self.max_candidate_num):
self.docid_inputs.append(
tf.placeholder(tf.int64,
shape=[None],
name="docid_input{0}".format(i)))
self.labels.append(
tf.placeholder(tf.float32,
shape=[None],
name="label{0}".format(i)))
self.PAD_embed = tf.zeros([1, self.feature_size], dtype=tf.float32)
# Build model
self.output = self.ranking_model(self.max_candidate_num,
scope='ranking_model')
reshaped_labels = tf.transpose(
tf.convert_to_tensor(self.labels)
) # reshape from [max_candidate_num, ?] to [?, max_candidate_num]
pad_removed_output = self.remove_padding_for_metric_eval(
self.docid_inputs, self.output)
for metric in self.exp_settings['metrics']:
for topn in self.exp_settings['metrics_topn']:
metric_value = utils.make_ranking_metric_fn(metric, topn)(
reshaped_labels, pad_removed_output, None)
tf.summary.scalar('%s_%d' % (metric, topn),
metric_value,
collections=['eval'])
# Gradients and SGD update operation for training the model.
if not forward_only:
# Training outputs and losses.
self.rank_list_size = exp_settings['train_list_cutoff']
train_output = self.ranking_model(self.rank_list_size,
scope='ranking_model')
train_labels = self.labels[:self.rank_list_size]
self.propensity_weights = []
for i in range(self.rank_list_size):
self.propensity_weights.append(
tf.placeholder(tf.float32,
shape=[None],
name="propensity_weights{0}".format(i)))
tf.summary.scalar('Propensity weights %d' % i,
tf.reduce_max(self.propensity_weights[i]),
collections=['train'])
print('Loss Function is ' + self.hparams.loss_func)
self.loss = None
reshaped_train_labels = tf.transpose(
tf.convert_to_tensor(train_labels)
) # reshape from [rank_list_size, ?] to [?, rank_list_size]
reshaped_propensity = tf.transpose(
tf.convert_to_tensor(self.propensity_weights)
) # reshape from [rank_list_size, ?] to [?, rank_list_size]
if self.hparams.loss_func == 'softmax':
self.loss = self.softmax_loss(train_output,
reshaped_train_labels,
reshaped_propensity)
elif self.hparams.loss_func == 'click_weighted_softmax_cross_entropy':
self.loss = self.click_weighted_softmax_loss(
train_output, reshaped_train_labels, reshaped_propensity)
elif self.hparams.loss_func == 'click_weighted_pairwise_loss':
self.loss = self.click_weighted_pairwise_loss(
train_output, reshaped_train_labels, reshaped_propensity)
else:
self.loss = self.sigmoid_loss(train_output,
reshaped_train_labels,
reshaped_propensity)
params = tf.trainable_variables()
if self.hparams.l2_loss > 0:
for p in params:
self.loss += self.hparams.l2_loss * tf.nn.l2_loss(p)
# Select optimizer
self.optimizer_func = tf.train.AdagradOptimizer
if self.hparams.grad_strategy == 'sgd':
self.optimizer_func = tf.train.GradientDescentOptimizer
opt = self.optimizer_func(self.hparams.learning_rate)
self.gradients = tf.gradients(self.loss, params)
if self.hparams.max_gradient_norm > 0:
self.clipped_gradients, self.norm = tf.clip_by_global_norm(
self.gradients, self.hparams.max_gradient_norm)
self.updates = opt.apply_gradients(
zip(self.clipped_gradients, params),
global_step=self.global_step)
else:
self.norm = None
self.updates = opt.apply_gradients(
zip(self.gradients, params), global_step=self.global_step)
tf.summary.scalar('Learning Rate',
self.learning_rate,
collections=['train'])
tf.summary.scalar('Loss',
tf.reduce_mean(self.loss),
collections=['train'])
clipped_labels = tf.clip_by_value(reshaped_train_labels,
clip_value_min=0,
clip_value_max=1)
pad_removed_train_output = self.remove_padding_for_metric_eval(
self.docid_inputs, train_output)
for metric in self.exp_settings['metrics']:
for topn in self.exp_settings['metrics_topn']:
list_weights = tf.reduce_mean(reshaped_propensity *
clipped_labels,
axis=1,
keep_dims=True)
metric_value = utils.make_ranking_metric_fn(metric, topn)(
reshaped_train_labels, pad_removed_train_output, None)
tf.summary.scalar('%s_%d' % (metric, topn),
metric_value,
collections=['train'])
weighted_metric_value = utils.make_ranking_metric_fn(
metric, topn)(reshaped_train_labels,
pad_removed_train_output, list_weights)
tf.summary.scalar('Weighted_%s_%d' % (metric, topn),
weighted_metric_value,
collections=['train'])
self.train_summary = tf.summary.merge_all(key='train')
self.eval_summary = tf.summary.merge_all(key='eval')
self.saver = tf.train.Saver(tf.global_variables())
def run(data_folder='./',
data_format='csv',
data_file_filter=None,
data_file_terminal_ID=True,
data_read_options=None,
data_split_by=None,
data_group_by=None,
data_extra_processing=None,
model_name='QLearn',
model_changing_properties=None,
model_constant_properties=None,
participantID="Name",
participant_choices='Actions',
participant_rewards='Rewards',
model_fit_value='ActionProb',
fit_subset=None,
task_stimuli=None,
participant_action_options=None,
fit_method='Evolutionary',
fit_method_args=None,
fit_measure='-loge',
fit_measure_args=None,
fit_extra_measures=None,
participant_varying_model_parameters=None,
label=None,
save_fitting_progress=False,
config_file=None,
output_path=None,
pickle=False,
boundary_excess_cost_function=None,
min_log_level='INFO',
numpy_error_level="log",
fit_float_error_response_value=1 / 1e100,
calculate_covariance=False
):
"""
A framework for fitting models to data for tasks, along with
recording the data associated with the fits.
Parameters
----------
data_folder : string or list of strings, optional
The folder where the data can be found. Default is the current folder.
data_format : string, optional
The file type of the data, from ``mat``, ``csv``, ``xlsx`` and ``pkl``. Default is ``csv``
data_file_filter : callable, string, list of strings or None, optional
A function to process the file names or a list of possible prefixes as strings or a single string.
Default ``None``, no file names removed
data_file_terminal_ID : bool, optional
Is there an ID number at the end of the filename? If not then a more general search will be performed.
Default ``True``
data_read_options : dict, optional
The keyword arguments for the data importing method chosen
data_split_by : string or list of strings, optional
If multiple participant datasets are in one file sheet, this specifies the column or columns that can
distinguish and identify the rows for each participant. Default ``None``
data_group_by : list of strings, optional
A list of parts of filenames that are repeated across participants, identifying all the files that should
be grouped together to form one participants data. The rest of the filename is assumed to identify the
participant. Default is ``None``
data_extra_processing : callable, optional
A function that modifies the dictionary of data read for each participant in such that it is appropriate
for fitting. Default is ``None``
model_name : string, optional
The name of the file where a model.modelTemplate.Model class can be found. Default ``QLearn``
model_changing_properties : dictionary with values of tuple of two floats, optional
Parameters are the options that you allow to vary across model fits. Each model parameter is specified as a
dict key. The value is a tuple containing the upper and lower search bounds, e.g. ``alpha`` has the bounds
(0, 1). Default ``None``
model_constant_properties : dictionary of float, string or binary valued elements, optional
These contain all the the model options that define the version
of the model being studied. Default ``None``
participantID : str, optional
The key (label) used to identify each participant. Default ``Name``
participant_choices : string, optional
The participant data key of their action choices. Default ``'Actions'``
participant_rewards : string, optional
The participant data key of the participant reward data. Default ``'Rewards'``
model_fit_value : string, optional
The key to be compared in the model data. Default ``'ActionProb'``
fit_subset : ``float('Nan')``, ``None``, ``"rewarded"``, ``"unrewarded"``, ``"all"`` or list of int, optional
Describes which, if any, subset of trials will be used to evaluate the performance of the model.
This can either be described as a list of trial numbers or, by passing
- ``"all"`` for fitting all trials
- ``float('Nan')`` or ``"unrewarded"`` for all those trials whose feedback was ``float('Nan')``
- ``"rewarded"`` for those who had feedback that was not ``float('Nan')``
Default ``None``, which means all trials will be used.
task_stimuli : list of strings or None, optional
The keys containing the observational parameters seen by the
participant before taking a decision on an action. Default ``None``
participant_action_options : string or list of strings or None or one element list with a list, optional
If a string or list of strings these are treated as dict keys where the valid actions for each trial can
be found. If None then all trials will use all available actions. If the list contains one list then it will
be treated as a list of valid actions for each trialstep. Default ``'None'``
fit_method : string, optional
The fitting method to be used. The names accepted are those of the modules in the folder fitAlgs containing a
FitAlg class. Default ``'evolutionary'``
fit_method_args : dict, optional
A dictionary of arguments specific to the fitting method. Default ``None``
fit_measure : string, optional
The name of the function used to calculate the quality of the fit.
The value it returns provides the fitter with its fitting guide. Default ``-loge``
fit_measure_args : dict, optional
The parameters used to initialise fitMeasure and extraFitMeasures. Default ``None``
fit_extra_measures : list of strings, optional
List of fit measures not used to fit the model, but to provide more information. Any arguments needed for these
measures should be placed in fitMeasureArgs. Default ``None``
participant_varying_model_parameters : dict of string, optional
A dictionary of model settings whose values should vary from participant to participant based on the
values found in the imported participant data files. The key is the label given in the participant data file,
as a string, and the value is the associated label in the model, also as a string. Default ``{}``
label : string, optional
The label for the data fitting. Default ``None`` will mean no data is saved to files.
save_fitting_progress : bool, optional
Specifies if the results from each iteration of the fitting process should be returned. Default ``False``
config_file : string, optional
The file name and path of a ``.yaml`` configuration file. Overrides all other parameters if found.
Default ``None``
output_path : string, optional
The path that will be used for the run output. Default ``None``
pickle : bool, optional
If true the data for each model, and participant is recorded.
Default is ``False``
boundary_excess_cost_function : str or callable returning a function, optional
The function is used to calculate the penalty for exceeding the boundaries.
Default is ``boundFunc.scalarBound()``
min_log_level : str, optional
Defines the level of the log from (``DEBUG``, ``INFO``, ``WARNING``, ``ERROR``, ``CRITICAL``). Default ``INFO``
numpy_error_level : {'log', 'raise'}
Defines the response to numpy errors. Default ``log``. See numpy.seterr
fit_float_error_response_value : float, optional
If a floating point error occurs when running a fit the fitter function
will return a value for each element of fpRespVal. Default is ``1/1e100`
calculate_covariance : bool, optional
Is the covariance calculated. Default ``False``
See Also
--------
modelGenerator : The model factory
outputting : The outputting functions
fitAlgs.fitAlg.FitAlg : General class for a method of fitting data
fitAlgs.fitSims.fitSim : General class for a method of simulating the fitting of data
data.Data : Data import class
"""
config = copy.deepcopy(locals())
if participant_varying_model_parameters is None:
model_changing_variables = {}
else:
model_changing_variables = participant_varying_model_parameters
# TODO : Validate model_changing_properties with the data and the model
participants = data.Data.load_data(file_type=data_format,
folders=data_folder,
file_name_filter=data_file_filter,
terminal_ID=data_file_terminal_ID,
split_by=data_split_by,
participantID=participantID,
choices=participant_choices,
feedbacks=participant_rewards,
stimuli=task_stimuli,
action_options=participant_action_options,
group_by=data_group_by,
extra_processing=data_extra_processing,
data_read_options=data_read_options)
if model_changing_properties:
model_parameters = {}
for key, value in model_changing_properties.items():
if len(value) == 2:
v1, v2 = value
if v2 < v1:
raise OrderError('The bounds specified for model parameter ``{}`` must have the lower bound first'.format(key))
else:
model_parameters[key] = (v1 + v2) / 2
else:
raise LengthError("The parameter values for the ``model_changing_properties`` must be presented as a list of the maximum and minimum values. Review those of ``{}``".format(key))
else:
model_parameters = model_changing_properties
models = ModelGen(model_name=model_name,
parameters=model_parameters,
other_options=model_constant_properties)
model_simulator = FitSim(participant_choice_property=participants.choices,
participant_reward_property=participants.feedbacks,
model_fitting_variable=model_fit_value,
fit_subset=fit_subset,
task_stimuli_property=participants.stimuli,
action_options_property=participants.action_options,
float_error_response_value=fit_float_error_response_value
)
fitting_method = utils.find_class(fit_method,
class_folder='fitAlgs',
inherited_class=FitAlg,
excluded_files=['boundFunc', 'qualityFunc', 'fitSims'])
if fit_method_args is None:
fit_method_args = {}
fitter = fitting_method(fit_sim=model_simulator,
fit_measure=fit_measure,
extra_fit_measures=fit_extra_measures,
fit_measure_args=fit_measure_args,
bounds=model_changing_properties,
boundary_excess_cost=boundary_excess_cost_function,
calculate_covariance=calculate_covariance,
**fit_method_args)
with outputting.Saving(config=config) as file_name_generator:
logger = logging.getLogger('Fitting')
log_fitting_parameters(fitter.info())
message = 'Beginning the data fitting'
logger.info(message)
model_ID = 0
# Initialise the stores of information
participant_fits = collections.defaultdict(list) # type: collections.defaultdict[Any, list]
for model, model_parameter_variables, model_static_args in models.iter_details():
for v in model_changing_variables.values():
model_static_args[v] = "<Varies for each participant>"
log_model_fitting_parameters(model, model_parameter_variables, model_static_args)
participantID = participants.participantID
for participant in participants:
participant_name = participant[participantID]
if isinstance(participant_name, (list, tuple)):
participant_name = participant_name[0]
for k, v in model_changing_variables.items():
model_static_args[v] = participant[k]
# Find the best model values from those proposed
message = "Beginning participant fit for participant {}".format(participant_name)
logger.info(message)
model_fitted, fit_quality, fitting_data = fitter.participant(model,
model_parameter_variables,
model_static_args,
participant)
message = "Participant fitted"
logger.debug(message)
log_model_fitted_parameters(model_parameter_variables, model_fitted.params(), fit_quality, participant_name)
participant_fits = record_participant_fit(participant, participant_name, model_fitted.returnTaskState(),
str(model_ID), fitting_data, model_changing_variables,
participant_fits, fileNameGen=file_name_generator,
pickleData=pickle, saveFittingProgress=save_fitting_progress)
model_ID += 1
fit_record(participant_fits, file_name_generator)
def __init__(self,
number_actions=2,
number_cues=1,
number_critics=None,
action_codes=None,
non_action='None',
prior=None,
stimulus_shaper=None,
stimulus_shaper_name=None,
stimulus_shaper_properties=None,
reward_shaper=None,
reward_shaper_name=None,
reward_shaper_properties=None,
decision_function=None,
decision_function_name=None,
decision_function_properties=None,
**kwargs):
""""""
self.Name = self.get_name()
self.pattern_parameters = self.kwarg_pattern_parameters(kwargs)
for k, v in self.pattern_parameters.items():
setattr(self, k, v)
self.pattern_parameters = self.kwarg_pattern_parameters(kwargs)
for k, v in self.pattern_parameters.items():
setattr(self, k, v)
self.number_actions = number_actions
self.number_cues = number_cues
if number_critics is None:
number_critics = self.number_actions * self.number_cues
self.number_critics = number_critics
if action_codes is None:
action_codes = {k: k for k in range(self.number_actions)}
self.actionCode = action_codes
self.defaultNonAction = non_action
if prior is None:
prior = np.ones(self.number_actions) / self.number_actions
self.prior = prior
self.stimuli = np.ones(self.number_cues)
self.stimuliFilter = np.ones(self.number_cues)
self.currAction = None
self.decision = None
self.validActions = None
self.lastObservation = None
self.probabilities = np.array(self.prior)
self.decProbabilities = np.array(self.prior)
self.expectedRewards = np.ones(self.number_actions)
self.expectedReward = np.array([1])
if stimulus_shaper is not None and issubclass(stimulus_shaper,
Stimulus):
if stimulus_shaper_properties is not None:
stimulus_shaper_kwargs = {
k: v
for k, v in kwargs.items()
if k in stimulus_shaper_properties
}
else:
stimulus_shaper_kwargs = kwargs.copy()
self.stimulus_shaper = stimulus_shaper(**stimulus_shaper_kwargs)
elif isinstance(stimulus_shaper_name, str):
stimulus_class = utils.find_class(
stimulus_shaper_name,
class_folder='tasks',
inherited_class=Stimulus,
excluded_files=['taskTemplate', '__init__', 'taskGenerator'])
stimulus_shaper_kwargs = {
k: v
for k, v in kwargs.items()
if k in utils.get_class_args(stimulus_class)
}
self.stimulus_shaper = stimulus_class(**stimulus_shaper_kwargs)
else:
self.stimulus_shaper = Stimulus()
if reward_shaper is not None and issubclass(reward_shaper, Rewards):
if reward_shaper_properties is not None:
reward_shaper_kwargs = {
k: v
for k, v in kwargs.items() if k in reward_shaper_properties
}
else:
reward_shaper_kwargs = kwargs.copy()
self.reward_shaper = reward_shaper(**reward_shaper_kwargs)
elif isinstance(reward_shaper_name, str):
reward_class = utils.find_class(
reward_shaper_name,
class_folder='tasks',
inherited_class=Rewards,
excluded_files=['taskTemplate', '__init__', 'taskGenerator'])
reward_shaper_kwargs = {
k: v
for k, v in kwargs.items()
if k in utils.get_class_args(reward_class)
}
self.reward_shaper = reward_class.processFeedback(
**reward_shaper_kwargs)
else:
self.reward_shaper = Rewards()
if callable(decision_function):
if decision_function_properties is not None:
decision_shaper_kwargs = {
k: v
for k, v in kwargs.items()
if k in decision_function_properties
}
else:
decision_shaper_kwargs = kwargs.copy()
self.decision_function = decision_function(
**decision_shaper_kwargs)
elif isinstance(decision_function_name, str):
decision_function = utils.find_function(decision_function_name,
'model/decision')
decision_function_kwargs = {
k: v
for k, v in kwargs.items()
if k in utils.get_function_args(decision_function)
}
self.decision_function = decision_function(
**decision_function_kwargs)
else:
self.decision_function = weightProb(
list(range(self.number_actions)))
self.parameters = {
"Name":
self.Name,
"number_actions":
self.number_actions,
"number_cues":
self.number_cues,
"number_critics":
self.number_critics,
"prior":
copy.copy(self.prior),
"non_action":
self.defaultNonAction,
"actionCode":
copy.copy(self.actionCode),
"stimulus_shaper":
self.stimulus_shaper.details(),
"reward_shaper":
self.reward_shaper.details(),
"decision_function":
utils.callableDetailsString(self.decision_function)
}
self.parameters.update(self.pattern_parameters)
# Recorded information
self.recAction = []
self.recActionSymbol = []
self.recStimuli = []
self.recReward = []
self.recExpectations = []
self.recExpectedReward = []
self.recExpectedRewards = []
self.recValidActions = []
self.recDecision = []
self.recProbabilities = []
self.recActionProbs = []
self.recActionProb = []
self.simID = None
def train(exp_settings):
# Prepare data.
print("Reading data in %s" % FLAGS.data_dir)
train_set = utils.read_data(FLAGS.data_dir, 'train', FLAGS.max_list_cutoff)
valid_set = utils.read_data(FLAGS.data_dir, 'valid', FLAGS.max_list_cutoff)
print("Train Rank list size %d" % train_set.rank_list_size)
print("Valid Rank list size %d" % valid_set.rank_list_size)
exp_settings['max_candidate_num'] = max(train_set.rank_list_size,
valid_set.rank_list_size)
test_set = None
if FLAGS.test_while_train:
test_set = utils.read_data(FLAGS.data_dir, 'test',
FLAGS.max_list_cutoff)
print("Test Rank list size %d" % test_set.rank_list_size)
exp_settings['max_candidate_num'] = max(
test_set.rank_list_size, exp_settings['max_candidate_num'])
exp_settings['train_list_cutoff'] = min(
FLAGS.train_list_cutoff, exp_settings['max_candidate_num']
) if FLAGS.train_list_cutoff > 0 else exp_settings['max_candidate_num']
# Pad data
train_set.pad(exp_settings['max_candidate_num'])
valid_set.pad(exp_settings['max_candidate_num'])
if FLAGS.test_while_train:
test_set.pad(exp_settings['max_candidate_num'])
config = tf.ConfigProto()
config.gpu_options.allow_growth = True
with tf.Session(config=config) as sess:
# Create model based on the input layer.
print("Creating model...")
model = create_model(sess, exp_settings, train_set, False)
#model.print_info()
# Create data feed
train_input_feed = utils.find_class(exp_settings['train_input_feed'])(
model, FLAGS.batch_size, exp_settings['train_input_hparams'], sess)
valid_input_feed = utils.find_class(exp_settings['valid_input_feed'])(
model, FLAGS.batch_size, exp_settings['valid_input_hparams'], sess)
test_input_feed = None
if FLAGS.test_while_train:
test_input_feed = utils.find_class(
exp_settings['test_input_feed'])(
model, FLAGS.batch_size,
exp_settings['test_input_hparams'], sess)
# Create tensorboard summarizations.
train_writer = tf.summary.FileWriter(FLAGS.model_dir + '/train_log',
sess.graph)
valid_writer = tf.summary.FileWriter(FLAGS.model_dir + '/valid_log')
test_writer = None
if FLAGS.test_while_train:
test_writer = tf.summary.FileWriter(FLAGS.model_dir + '/test_log')
# This is the training loop.
step_time, loss = 0.0, 0.0
current_step = 0
previous_losses = []
best_perf = None
while True:
# Get a batch and make a step.
start_time = time.time()
input_feed, _ = train_input_feed.get_batch(train_set,
check_validation=True)
step_loss, _, summary = model.step(sess, input_feed, False)
step_time += (time.time() -
start_time) / FLAGS.steps_per_checkpoint
loss += step_loss / FLAGS.steps_per_checkpoint
current_step += 1
train_writer.add_summary(summary, model.global_step.eval())
# Once in a while, we save checkpoint, print statistics, and run evals.
if current_step % FLAGS.steps_per_checkpoint == 0:
# Print statistics for the previous epoch.
print("global step %d learning rate %.4f step-time %.2f loss "
"%.4f" % (model.global_step.eval(),
model.learning_rate.eval(), step_time, loss))
previous_losses.append(loss)
# Validate model
def validate_model(data_set, data_input_feed):
it = 0
count_batch = 0.0
summary_list = []
batch_size_list = []
while it < len(data_set.initial_list):
input_feed, info_map = data_input_feed.get_next_batch(
it, data_set, check_validation=False)
_, _, summary = model.step(sess, input_feed, True)
summary_list.append(summary)
batch_size_list.append(len(info_map['input_list']))
it += batch_size_list[-1]
count_batch += 1.0
return utils.merge_TFSummary(summary_list, batch_size_list)
valid_summary = validate_model(valid_set, valid_input_feed)
valid_writer.add_summary(valid_summary,
model.global_step.eval())
print(" valid: %s" % (' '.join([
'%s:%.3f' % (x.tag, x.simple_value)
for x in valid_summary.value
])))
if FLAGS.test_while_train:
test_summary = validate_model(test_set, test_input_feed)
test_writer.add_summary(test_summary,
model.global_step.eval())
print(" test: %s" % (' '.join([
'%s:%.3f' % (x.tag, x.simple_value)
for x in test_summary.value
])))
# Save checkpoint if the objective metric on the validation set is better
if "objective_metric" in exp_settings:
for x in valid_summary.value:
if x.tag == exp_settings["objective_metric"]:
if current_step > FLAGS.start_saving_iteration:
if best_perf == None or best_perf < x.simple_value:
checkpoint_path = os.path.join(
FLAGS.model_dir, "%s.ckpt" %
exp_settings['learning_algorithm'])
model.saver.save(
sess,
checkpoint_path,
global_step=model.global_step)
best_perf = x.simple_value
print('Save model, valid %s:%.3f' %
(x.tag, best_perf))
break
# Save checkpoint if there is no objective metic
if best_perf == None and current_step > FLAGS.start_saving_iteration:
checkpoint_path = os.path.join(
FLAGS.model_dir,
"%s.ckpt" % exp_settings['learning_algorithm'])
model.saver.save(sess,
checkpoint_path,
global_step=model.global_step)
if loss == float('inf'):
break
step_time, loss = 0.0, 0.0
sys.stdout.flush()
if FLAGS.max_train_iteration > 0 and current_step > FLAGS.max_train_iteration:
break
def test(exp_settings):
config = tf.ConfigProto()
config.gpu_options.allow_growth = True
with tf.Session(config=config) as sess:
# Load test data.
print("Reading data in %s" % FLAGS.data_dir)
test_set = utils.read_data(FLAGS.data_dir, 'test',
FLAGS.max_list_cutoff)
exp_settings['max_candidate_num'] = test_set.rank_list_size
test_set.pad(exp_settings['max_candidate_num'])
# Create model and load parameters.
model = create_model(sess, exp_settings, test_set, True)
# Create input feed
#test_input_feed = utils.find_class(exp_settings['test_input_feed'])(model, 1, exp_settings['test_input_hparams'], sess)
test_input_feed = utils.find_class(exp_settings['test_input_feed'])(
model, FLAGS.batch_size, exp_settings['test_input_hparams'], sess)
test_writer = tf.summary.FileWriter(FLAGS.model_dir + '/test_log')
rerank_scores = []
summary_list = []
# Start testing.
it = 0
count_batch = 0.0
batch_size_list = []
while it < len(test_set.initial_list):
input_feed, info_map = test_input_feed.get_next_batch(
it, test_set, check_validation=False)
_, output_logits, summary = model.step(sess, input_feed, True)
summary_list.append(summary)
batch_size_list.append(len(info_map['input_list']))
for x in range(batch_size_list[-1]):
rerank_scores.append(output_logits[x])
it += batch_size_list[-1]
count_batch += 1.0
print("Testing {:.0%} finished".format(
float(it) / len(test_set.initial_list)),
end="\r",
flush=True)
print("\n[Done]")
test_summary = utils.merge_TFSummary(summary_list, batch_size_list)
test_writer.add_summary(test_summary, it)
print(" eval: %s" % (' '.join(
['%s:%.3f' % (x.tag, x.simple_value)
for x in test_summary.value])))
#get rerank indexes with new scores
rerank_lists = []
for i in range(len(rerank_scores)):
scores = rerank_scores[i]
rerank_lists.append(
sorted(range(len(scores)),
key=lambda k: scores[k],
reverse=True))
if not os.path.exists(FLAGS.output_dir):
os.makedirs(FLAGS.output_dir)
utils.output_ranklist(test_set, rerank_scores, FLAGS.output_dir,
'test')
return
def __init__(self, source):
self.parser = find_class(source.parser)()
self.params = json.loads(source.params)
self.source = source
super(SourceAggregatorThread, self).__init__()