def summary(self, extended=False):
subsection('Training parameters')
settings = {"idx": self.idx, "model size": self.model_size}
if extended:
settings.update({
"training size": self.training_size,
"validation size": self.validation_size,
"batch size": self.batch_size
})
display_settings(settings)
subsection("Hyper-parameters")
# group params
data_by_group = defaultdict(dict)
for data in self.hyper_parameters.values():
data_by_group[data['group']][data['name']] = data['value']
# Generate the table.
rows = [['Group', 'Hyper-parameter', 'Value']]
idx = 0
for grp in sorted(data_by_group.keys()):
for param, value in data_by_group[grp].items():
row = [grp, param, value]
if idx % 2:
row = colorize_row(row, 'cyan')
rows.append(row)
idx += 1
display_table(rows)
def summary(self, extended=False):
"""Display a summary of the tuner state
Args:
extended (bool, optional):Display an extended summay.
Defaults to False.
"""
if self.debug:
extended = True
subsection('Tuning parameters')
summary = {'tuner': self.name}
if not extended:
for attr in self.to_report:
summary[attr] = getattr(self, attr)
else:
for attr in self.user_parameters:
if attr in ['user_info']:
continue
summary[attr] = getattr(self, attr)
summary['log file'] = self.log_file
display_settings(summary)
if len(self.user_info) and extended:
subsection('User info')
display_settings(self.user_info)
self.host.summary(extended=extended)
def summary(self, extended=False):
"display cloud service status summary"
human_time = datetime.utcfromtimestamp(self.last_update)
human_time = human_time.strftime('%Y-%m-%dT%H:%M:%SZ')
section('Cloud service status')
info = {"status": self.status, "last update": human_time}
display_settings(info)
def summary(self, extended=False):
subsection('Directories')
settings = {
"results": self.results_dir,
"tmp": self.tmp_dir,
"export": self.export_dir
}
display_settings(settings)
if extended:
config._Host.summary(extended=extended)
def summary(self, extended=False):
"display statistics summary"
subsection("Tuning stats")
display_settings(self.to_config())
def __init__(self, model_fn, objective, **kwargs):
""" RandomSearch hypertuner
Args:
model_fn (function): Function that returns the Keras model to be
hypertuned. This function is supposed to return a different model
at every invocation via the use of distribution.* hyperparameters
range.
objective (str): Name of the metric to optimize for. The referenced
metric must be part of the the `compile()` metrics.
Attributes:
epoch_budget (int, optional): how many epochs to hypertune for.
Defaults to 100.
max_budget (int, optional): how many epochs to spend at most on
a given model. Defaults to 10.
min_budget (int, optional): how many epochs to spend at least on
a given model. Defaults to 3.
num_executions(int, optional): number of execution for each model.
Defaults to 1.
project (str, optional): project the tuning belong to.
Defaults to 'default'.
architecture (str, optional): Name of the architecture tuned.
Default to 'default'.
user_info(dict, optional): user supplied information that will be
recorded alongside training data. Defaults to {}.
label_names (list, optional): Label names for confusion matrix.
Defaults to None, in which case the numerical labels are used.
max_model_parameters (int, optional):maximum number of parameters
allowed for a model. Prevent OOO issue. Defaults to 2500000.
checkpoint (Bool, optional): Checkpoint model. Setting it to false
disable it. Defaults to True
dry_run(bool, optional): Run the tuner without training models.
Defaults to False.
debug(bool, optional): Display debug information if true.
Defaults to False.
display_model(bool, optional):Display model summary if true.
Defaults to False.
results_dir (str, optional): Tuning results dir.
Defaults to results/. Can specify a gs:// path.
tmp_dir (str, optional): Temporary dir. Wiped at tuning start.
Defaults to tmp/. Can specify a gs:// path.
export_dir (str, optional): Export model dir. Defaults to export/.
Can specify a gs:// path.
FIXME:
- Deal with early stop correctly
- allows different halving ratio for epochs and models
- allows differnet type of distribution
"""
super(UltraBand, self).__init__(model_fn, objective, "UltraBand",
RandomDistributions, **kwargs)
self.config = UltraBandConfig(kwargs.get('ratio',
3), self.state.min_epochs,
self.state.max_epochs,
self.state.epoch_budget)
self.epoch_budget_expensed = 0
settings = {
"Epoch Budget": self.state.epoch_budget,
"Num Models Sequence": self.config.model_sequence,
"Num Epochs Sequence": self.config.epoch_sequence,
"Num Brackets": self.config.num_brackets,
"Number of Iterations": self.config.num_batches,
"Total Cost per Band": self.config.total_epochs_per_band
}
section('UltraBand Tuning')
subsection('Settings')
display_settings(settings)