if cycle==1 and max_cycle==1:
# Directly use up all time left in one iteration
n_estimators = M.model.n_estimators
new_n_estimators = int((np.floor(time_left_over / time_spent) - 1 ) * n_estimators)
if new_n_estimators<=n_estimators: break
M.model.n_estimators = new_n_estimators
else:
# Make a learning curve by exponentially increasing the number of estimators
M.model.n_estimators = int(np.exp2(cycle))
M.model.n_estimators = min(max_estimators, M.model.n_estimators)
vprint( verbose, "[+] Number of estimators: %d" % (M.model.n_estimators))
last_n_estimators = M.model.n_estimators
# Fit base estimators
# -------------------
M.fit(D.data['X_train'], D.data['Y_train'])
vprint( verbose, "[+] Fitting success, time spent so far %5.2f sec" % (time.time() - start))
vprint( verbose, "[+] Size of trained model %5.2f bytes" % data_io.total_size(M))
# Make predictions
# -----------------
Y_valid = M.predict(D.data['X_valid'])
Y_test = M.predict(D.data['X_test'])
vprint( verbose, "[+] Prediction success, time spent so far %5.2f sec" % (time.time() - start))
# Write results
# -------------
if overwrite_output:
filename_valid = basename + '_valid.predict'
filename_test = basename + '_test.predict'
else:
filename_valid = basename + '_valid_' + str(cycle).zfill(3) + '.predict'
arr_range = np.array([[elem_hyps[3], elem_hyps[4]]
for elem_hyps in space_hyps])
model_bo = bayesobo.BO(arr_range, str_acq='ucb')
cur_hyps = [elem_hyps[2] for elem_hyps in space_hyps]
list_hyps_all = []
list_measures_all = []
while time_spent <= time_budget - (
2 * time_one_loop) - TIME_RESIDUAL and is_bo and True:
vprint(
verbose, "=========== " + basename.capitalize() +
" Training cycle " + str(cycle) + " ================")
M = MyAutoML(D.info, verbose=False, debug_mode=debug_mode)
M.fit(X_train_train, Y_train_train)
vprint(
verbose, "[+] Fitting success, time spent so far %5.2f sec" %
(time.time() - start))
# Make predictions
# -----------------
pred_train_valid = M.predict(X_train_valid)
if 'classification' in D.info['task']:
performance = sklearn.metrics.roc_auc_score(
Y_train_valid, pred_train_valid)
performance = 2 * performance - 1
vprint(verbose,
"[+] AUC for X_train_valid, %5.4f" % (performance))
else:
preformance = 0.0