def _setup_learner_options(self, model, **params_learner):
""" save main characteristics, fetch default parameters depending on the algo
generate init and bounds of the parameters (most tricky part)
"""
self._backup_initparams = params_learner # store initial params just in case
default = self._ALGO_INFOS[params_learner['algo']][0] # default hyper parameters
opt_l = ut.merge_dico(default, params_learner, update_type = 4)
opt_l.update({'model': model,
'algo': params_learner['algo'],
'nb_params':model.n_params,
'bounds_obj':params_learner.get('bounds_obj'),
'init_obj':params_learner.get('init_obj'),
'rdm_gen':self.rdm_gen, 'mp_obj':self.mp})
opt_l.update({'bounds_params':self._gen_boundaries_params(**opt_l)})
opt_l.update({'init_params':self._gen_init_params(**opt_l)})
self.options_learner = opt_l
def parse_meta_config(cls, input_file = 'inputfile.txt', update_rules = False, debug = False):
""" Parsing an input file containing a meta-config into a list of configs.
It relies on generating the cartesian product of the elements
Example (txt file containing a meta config)
-------
###
_CONTEXT {'_T':5}
key1 Val1 Val2
key2 Val3
###
--> [{key1: eval(Val1), key2: eval(Val3)}, {key1: eval(Val2), key2: eval(Val3)}]
Syntax of the meta-config file:
---------------------------------
Keys starting with a _ and in self.METAPARAMS_INFO.keys() are metaparameters
of the parser. o.w. they will be keys of the final config dico.
If the first line key = _CONTEXT the Val associated should be of the form
{'_contextvar1':val,'_contextvar2':val1} i.e. a dico with keys starting with a '_'
Arguments
---------
input_file: str
path of the meta-config file
update_rules :
When building the configs should the first value seen (e.g. Val1) be used
as a reference. Only works if Vals are <dic>. If True exampel's output becomes:
[{key1: eval(Val1), key2: eval(Val3)}, {key1: eval(Val2).update(eval(Val2)), key2: eval(Val3)}]
debug : bool
debug mode
Output
------
list_configs_processing_meta_configs: <list<dic>>
"""
if(debug): pdb.set_trace()
use_context = False
with open(input_file, 'r') as csvfile:
reader = csv.reader(csvfile, delimiter = ' ')
list_values = list([])
list_keys = list([])
nbline = 0
dico_METAPARAMS = {}
for line in reader:
nbline += 1
if(line in cls.EMPTY_LINE) or (line[0] in cls.LEX_NA): pass
elif(line[0] == '_CONTEXT'):
context = literal_eval(line[1])
if(ut.is_dico(context) and np.product([c[0] == '_' for c in context])):
use_context = True
elif(line[0] in cls.METAPARAMS_NAME):
assert(len(line) == 2), 'batch input file: 1 arg expected in l.' + str(nbline)+ ' (' +str(line[0]) + ')'
dico_METAPARAMS[line[0]] = literal_eval(line[1])
else:
assert (len(line)>=2), 'batch input file: not enough args in l.' + str(nbline)
list_keys.append(line[0])
if(use_context):
line_with_context = [cls._apply_context(line[i], context) for i in range(1,len(line))]
ev_tmp = [eval(lwc) for lwc in line_with_context]
else:
ev_tmp = [literal_eval(line[i]) for i in range(1,len(line))]
if(update_rules):
ref_value = copy.copy(ev_tmp[0])
ev_tmp_new = []
for ev_i in ev_tmp:
in_progress = ut.merge_dico(ref_value, ev_i, update_type = 0, copy = True)
ev_tmp_new.append(copy.copy(in_progress))
ev_tmp = ev_tmp_new
list_values = ut.cartesianProduct(list_values, ev_tmp, ut.appendList)
list_configs = [ut.lists2dico(list_keys, l) for l in list_values]
list_configs = cls._apply_metaparams(list_configs, dico_METAPARAMS)
list_configs = [cls._processing_meta_configs(c) for c in list_configs]
return list_configs