def run_models(models,
orders,
GP_likelihoods=['gaussian', 'warped'],
WD_kernel_degrees=[3],
adaboost_learning_rates=[0.1],
adaboost_num_estimators=[100],
adaboost_max_depths=[3],
learn_options_set=None,
test=False,
CV=True,
setup_function=setup,
set_target_fn=set_target,
pam_audit=True,
length_audit=True,
return_data=False):
'''
CV is set to false if want to train a final model and not cross-validate, but it goes in to what
looks like cv code
'''
results = {}
assert learn_options_set is not None, "need to specify learn_options_set"
all_learn_options = {}
# shorten so easier to display on graphs
feat_models_short = {
'L1': "L1",
'L2': "L2",
'elasticnet': "EN",
'linreg': "LR",
'RandomForest': "RF",
'AdaBoost': "AB",
'AdaBoostClassifier': "ABClass",
'doench': 'doench',
"logregL1": "logregL1",
"sgrna_from_doench": "sgrna_from_doench",
'SVC': 'SVC',
'xu_et_al': 'xu_et_al'
}
if not CV:
print(
"Received option CV=False, so I'm training using all of the data")
assert len(list(learn_options_set.keys(
))) == 1, "when CV is False, only 1 set of learn options is allowed"
assert len(models) == 1, "when CV is False, only 1 model is allowed"
for learn_options_str in list(learn_options_set.keys()):
# these options get augmented in setup
partial_learn_opt = learn_options_set[learn_options_str]
# if the model requires encoded features
for model in models:
# models requiring explicit featurization
if model in list(feat_models_short.keys()):
for order in orders:
print("running %s, order %d for %s" %
(model, order, learn_options_str))
Y, feature_sets, target_genes, learn_options, num_proc = setup_function(
test=test,
order=order,
learn_options=partial_learn_opt,
pam_audit=pam_audit,
length_audit=length_audit
) # TODO precompute features for all orders, as this is repated for each model
if model == 'L1':
learn_options_model = L1_setup(
copy.deepcopy(learn_options),
set_target_fn=set_target_fn)
elif model == 'L2':
learn_options_model = L2_setup(
copy.deepcopy(learn_options),
set_target_fn=set_target_fn)
elif model == 'elasticnet':
learn_options_model = elasticnet_setup(
copy.deepcopy(learn_options),
set_target_fn=set_target_fn)
elif model == 'linreg':
learn_options_model = linreg_setup(
copy.deepcopy(learn_options),
set_target_fn=set_target_fn)
elif model == "logregL1":
learn_options_model = logregL1_setup(
copy.deepcopy(learn_options),
set_target_fn=set_target_fn)
elif model == 'RandomForest':
learn_options_model = RF_setup(
copy.deepcopy(learn_options),
set_target_fn=set_target_fn)
elif model == 'SVC':
learn_options_model = SVC_setup(
copy.deepcopy(learn_options),
set_target_fn=set_target_fn)
elif model == 'doench':
learn_options_model = doench_setup(
copy.deepcopy(learn_options),
set_target_fn=set_target_fn)
elif model == 'sgrna_from_doench':
learn_options_model = sgrna_from_doench_setup(
copy.deepcopy(learn_options),
set_target_fn=set_target_fn)
elif model == 'xu_et_al':
learn_options_model = xu_et_al_setup(
copy.deepcopy(learn_options),
set_target_fn=set_target_fn)
elif model == 'AdaBoost' or 'AdaBoostClassifier':
for learning_rate in adaboost_learning_rates:
for num_estimators in adaboost_num_estimators:
for max_depth in adaboost_max_depths:
learn_options_model = adaboost_setup(
copy.deepcopy(learn_options),
learning_rate=learning_rate,
num_estimators=num_estimators,
max_depth=max_depth,
set_target_fn=set_target_fn,
model=model)
model_string = feat_models_short[
model] + '_or%d_md%d_lr%.2f_n%d_%s' % (
learn_options_set[learn_options_str]["order"],
max_depth, learning_rate, num_estimators,
learn_options_str)
if model != 'AdaBoost':
model_string = feat_models_short[
model] + '_ord%d_%s' % (
learn_options_set[learn_options_str]["order"],
learn_options_str)
results[model_string] = pd.cross_validate(
Y,
feature_sets,
learn_options=learn_options_model,
TEST=test,
CV=CV)
all_learn_options[model_string] = learn_options_model
# if the model doesn't require explicit featurization
else:
assert setup_fn == setup, "not yet modified to handle this"
print("running %s for %s" % (model, learn_options_str))
Y, feature_sets, target_genes, learn_options, num_proc = setup(
test=test,
order=1,
learn_options=partial_learn_opt,
pam_audit=pam_audit,
length_audit=length_audit)
if model == 'mean':
learn_options_model = mean_setup(
copy.deepcopy(learn_options))
elif model == 'random':
learn_options_model = random_setup(
copy.deepcopy(learn_options))
elif model == 'DNN':
learn_options_model = DNN_setup(
copy.deepcopy(learn_options))
elif model == 'GP':
for likelihood in GP_likelihoods:
for degree in WD_kernel_degrees:
learn_options_model = GP_setup(
copy.deepcopy(learn_options),
likelihood=likelihood,
degree=degree)
model_string = '%s_%s_degree%d_%s' % (
model, likelihood, degree, learn_options_str)
results[model_string] = pd.cross_validate(
Y,
feature_sets,
learn_options=learn_options_model,
TEST=test,
CV=CV)
else:
raise NotImplementedError("model %s not supported" % model)
# "GP" already calls pd.cross_validate() and has its own model_string, so skip this.
if model != "GP":
model_string = model + '_%s' % learn_options_str
results[model_string] = pd.cross_validate(
Y,
feature_sets,
learn_options=learn_options_model,
TEST=test,
CV=CV)
all_learn_options[model_string] = learn_options_model
return results, all_learn_options
def run_models(models, orders, GP_likelihoods=['gaussian', 'warped'], WD_kernel_degrees=[3],
adaboost_learning_rates=[0.1], adaboost_num_estimators=[100], adaboost_max_depths=[3],
learn_options_set=None, test=False, CV=True, setup_function=setup, set_target_fn=set_target):
'''
CV is set to false if want to train a final model and not cross-validate, but it goes in to what
looks like cv code
'''
results = {}
assert learn_options_set is not None, "need to specify learn_options_set"
all_learn_options = {}
#shorten so easier to display on graphs
feat_models_short = {'L1':"L1", 'L2':"L2", 'elasticnet':"EN", 'linreg':"LR",
'RandomForest': "RF",
'AdaBoost':"AB", 'doench': 'doench',
"logregL1": "logregL1", "sgrna_from_doench":"sgrna_from_doench", 'SVC': 'SVC', 'xu_et_al': 'xu_et_al'}
if not CV:
print "Received option CV=False, so I'm training using all of the data"
assert len(learn_options_set.keys()) == 1, "when CV is False, only 1 set of learn options is allowed"
assert len(models) == 1, "when CV is False, only 1 model is allowed"
for learn_options_str in learn_options_set.keys():
# these options get augmented in setup
partial_learn_opt = learn_options_set[learn_options_str]
# if the model requires encoded features
for model in models:
# models requiring explicit featurization
if model in feat_models_short.keys():
for order in orders:
print "running %s, order %d for %s" % (model, order, learn_options_str)
Y, feature_sets, target_genes, learn_options, num_proc = setup_function(test=test, order=order, learn_options=partial_learn_opt) # TODO precompute features for all orders, as this is repated for each model
if model == 'L1':
learn_options_model = L1_setup(copy.deepcopy(learn_options), set_target_fn=set_target_fn)
elif model == 'L2':
learn_options_model = L2_setup(copy.deepcopy(learn_options), set_target_fn=set_target_fn)
elif model == 'elasticnet':
learn_options_model = elasticnet_setup(copy.deepcopy(learn_options), set_target_fn=set_target_fn)
elif model == 'linreg':
learn_options_model = linreg_setup(copy.deepcopy(learn_options), set_target_fn=set_target_fn)
elif model == "logregL1":
learn_options_model = logregL1_setup(copy.deepcopy(learn_options), set_target_fn=set_target_fn)
elif model == 'RandomForest':
learn_options_model = RF_setup(copy.deepcopy(learn_options), set_target_fn=set_target_fn)
elif model == 'SVC':
learn_options_model = SVC_setup(copy.deepcopy(learn_options), set_target_fn=set_target_fn)
elif model == 'doench':
learn_options_model = doench_setup(copy.deepcopy(learn_options), set_target_fn=set_target_fn)
elif model == 'sgrna_from_doench':
learn_options_model = sgrna_from_doench_setup(copy.deepcopy(learn_options), set_target_fn=set_target_fn)
elif model == 'xu_et_al':
learn_options_model = xu_et_al_setup(copy.deepcopy(learn_options), set_target_fn=set_target_fn)
elif model == 'AdaBoost':
for learning_rate in adaboost_learning_rates:
for num_estimators in adaboost_num_estimators:
for max_depth in adaboost_max_depths:
learn_options_model = adaboost_setup(copy.deepcopy(learn_options), learning_rate=learning_rate, num_estimators=num_estimators, max_depth=max_depth, set_target_fn=set_target_fn)
model_string = feat_models_short[model] + '_or%d_md%d_lr%.2f_n%d_%s' % (learn_options_set[learn_options_str]["order"], max_depth, learning_rate, num_estimators, learn_options_str)
if model != 'AdaBoost':
model_string = feat_models_short[model] + '_ord%d_%s' % (learn_options_set[learn_options_str]["order"], learn_options_str)
results[model_string] = pd.cross_validate(Y, feature_sets, learn_options=learn_options_model, TEST=test, CV=CV)
all_learn_options[model_string] = learn_options_model
# if the model doesn't require explicit featurization
else:
assert setup_fn==setup, "not yet modified to handle this"
print "running %s for %s" % (model, learn_options_str)
Y, feature_sets, target_genes, learn_options, num_proc = setup(test=test, order=1, learn_options=partial_learn_opt)
if model == 'mean':
learn_options_model = mean_setup(copy.deepcopy(learn_options))
elif model == 'random':
learn_options_model = random_setup(copy.deepcopy(learn_options))
elif model == 'DNN':
learn_options_model = DNN_setup(copy.deepcopy(learn_options))
elif model == 'GP':
for likelihood in GP_likelihoods:
for degree in WD_kernel_degrees:
learn_options_model = GP_setup(copy.deepcopy(learn_options), likelihood=likelihood, degree=degree)
model_string = '%s_%s_degree%d_%s' % (model, likelihood, degree, learn_options_str)
results[model_string] = pd.cross_validate(Y, feature_sets, learn_options=learn_options_model,TEST=test, CV=CV)
else:
raise NotImplementedError("model %s not supported" % model)
# "GP" already calls pd.cross_validate() and has its own model_string, so skip this.
if model != "GP":
model_string = model + '_%s' % learn_options_str
results[model_string] = pd.cross_validate(Y, feature_sets, learn_options=learn_options_model, TEST=test, CV=CV)
all_learn_options[model_string] = learn_options_model
return results, all_learn_options