def compute_inner_error(overall_df_inner, learning_rate_cv, num_iterations_cv, num_season_factors_cv,num_home_factors_cv, lam_cv, A_source):
print num_iterations_cv, num_season_factors_cv,num_home_factors_cv,lam_cv
inner_kf = KFold(n_splits=2)
pred_inner = {}
for train_inner, test_inner in inner_kf.split(overall_df_inner):
train_ix_inner = overall_df_inner.index[train_inner]
test_ix_inner = overall_df_inner.index[test_inner]
train_test_ix_inner = np.concatenate([test_ix_inner, train_ix_inner])
df_t_inner, dfc_t_inner = target_df.loc[train_test_ix_inner], target_dfc.loc[train_test_ix_inner]
tensor_inner = get_tensor(df_t_inner, start, stop)
tensor_copy_inner = tensor_inner.copy()
# First n
tensor_copy_inner[:len(test_ix_inner), 1:, :] = np.NaN
L_inner = target_L[np.ix_(np.concatenate([test_inner, train_inner]), np.concatenate([test_inner, train_inner]))]
if setting=="transfer":
A_source = A_store[learning_rate_cv][num_season_factors_cv][num_home_factors_cv][lam_cv][num_iterations_cv]
else:
A_source = None
H, A, T, Hs, As, Ts, HATs, costs = learn_HAT_adagrad_graph(case, tensor_copy_inner,
L_inner,
num_home_factors_cv,
num_season_factors_cv,
num_iter=num_iterations_cv,
lr=learning_rate_cv, dis=False,
lam=lam_cv,
A_known=A_source,
T_known=T_constant)
HAT = multiply_case(H, A, T, case)
for appliance in APPLIANCES_ORDER:
if appliance not in pred_inner:
pred_inner[appliance] = []
pred_inner[appliance].append(pd.DataFrame(HAT[:len(test_ix_inner), appliance_index[appliance], :], index=test_ix_inner))
err = {}
appliance_to_weight = []
for appliance in APPLIANCES_ORDER[1:]:
pred_inner[appliance] = pd.DataFrame(pd.concat(pred_inner[appliance]))
try:
if appliance == "hvac":
err[appliance] = compute_rmse_fraction(appliance, pred_inner[appliance][range(5-start, 11-start)], target, start, stop)[2]
else:
err[appliance] = compute_rmse_fraction(appliance, pred_inner[appliance], target, start, stop)[2]
appliance_to_weight.append(appliance)
except Exception, e:
# This appliance does not have enough samples. Will not be
# weighed
print 'here'
print(e)
print(appliance)
directory = os.path.expanduser(
'~/git/scalable-nilm/aaai18/predictions/TF-all/{}/case-{}/{}/{}'.format(setting, case, static_use, constant_use))
if not os.path.exists(directory):
os.makedirs(directory)
filename = os.path.join(directory, name + '.pkl')
try:
out[case][constant_use][static_use][setting][train_percentage][random_seed]={}
params[case][constant_use][static_use][setting][train_percentage][random_seed] = {}
pr = pickle.load(open(filename, 'r'))
pred = pr['Predictions']
parameter_data = pr['Learning Params']
params[case][constant_use][static_use][setting][train_percentage][random_seed] = parameter_data
for appliance in APPLIANCES_ORDER[1:]:
# prediction = pred[appliance]
prediction = pred[appliance].ix[idx]
if appliance == 'hvac':
prediction = prediction[range(5-start, 11-start)]
out[case][constant_use][static_use][setting][train_percentage][random_seed][appliance] = \
compute_rmse_fraction(appliance, prediction, target, start, stop)[2]
print("Computed for: {}".format(name))
# print case, constant_use, static_use, setting, train_percentage, random_seed
# print out[case][constant_use][static_use][setting][train_percentage][random_seed]
except Exception, e:
print(e)
print("Exception")
out[case][constant_use][static_use][setting][train_percentage] = pd.DataFrame(out[case][constant_use][static_use][setting][train_percentage]).mean(axis=1)
pickle.dump(out, open('../predictions/lr-tf-{}-{}-part.pkl'.format(source, target), 'w'))
pickle.dump(params, open('../predictions/params-lr-tf-{}-{}-part.pkl'.format(source, target), 'w'))
pred_inner[appliance].append(
pd.DataFrame(
HAT[:len(test_ix_inner),
appliance_index[appliance], :],
index=test_ix_inner))
err = {}
appliance_to_weight = []
for appliance in APPLIANCES_ORDER[1:]:
pred_inner[appliance] = pd.DataFrame(
pd.concat(pred_inner[appliance]))
try:
if appliance == "hvac":
err[appliance] = compute_rmse_fraction(
appliance, pred_inner[appliance][range(
5 - start, 11 - start)], target,
start, stop)[2]
else:
err[appliance] = compute_rmse_fraction(
appliance, pred_inner[appliance],
target, start, stop)[2]
appliance_to_weight.append(appliance)
except Exception, e:
# This appliance does not have enough samples. Will not be
# weighed
print(e)
print(appliance)
sys.stdout.flush()
print("Error weighted on: {}".format(
appliance_to_weight))
features, random_seed,
train_percentage)
else:
name = "{}-{}-{}-{}-{}-{}".format(
source, target, appliance, features, random_seed,
train_percentage)
directory = os.path.expanduser(
'~/git/scalable-nilm/aaai18/predictions/MF/{}'.format(
setting))
if not os.path.exists(directory):
os.makedirs(directory)
filename = os.path.join(directory, name + '.pkl')
pred = pickle.load(open(filename, 'r'))['Prediction']
print(filename)
out[setting][train_percentage][random_seed][appliance] = \
compute_rmse_fraction(appliance, pred, target)[2]
except:
pass
out[setting][train_percentage] = pd.DataFrame(
out[setting][train_percentage]).mean(axis=1)
save_loc = os.path.expanduser(
"~/git/scalable-nilm/aaai18/predictions/mf-{}-{}.pkl".format(
source, target))
pickle.dump(out, open(save_loc, 'w'))
pickle.dump(
out, open('../../../predictions/mf-{}-{}.pkl'.format(source, target), 'w'))
#pickle.dump(params, open('../predictions/params-lr-tf-{}-{}.pkl'.format(source, target), 'w'))
sys.stdout.flush()
err = {}
for num_iterations in range(100, 1400, 200):
err[num_iterations] = {}
for appliance in APPLIANCES_ORDER:
pred[appliance] = []
for appliance in APPLIANCES_ORDER:
pred[appliance].append(pd.DataFrame(HATs[num_iterations][:, appliance_index[appliance], :], index=source_df.index))
for appliance in APPLIANCES_ORDER:
pred[appliance] = pd.DataFrame(pd.concat(pred[appliance]))
for appliance in APPLIANCES_ORDER[1:]:
prediction = pred[appliance]
if appliance == "hvac":
prediction = prediction[range(5-start, 11-start)]
err[num_iterations][appliance] = \
compute_rmse_fraction(appliance, prediction, source, start, stop)[2]
# print("Computed for: {}".format())
err_weight = {}
for appliance in APPLIANCES_ORDER[1:]:
err_weight[appliance] = err[num_iterations][appliance]*contri[source][appliance]
mean_err = pd.Series(err_weight).sum()
err_store[learning_rate_cv][num_season_factors_cv][num_home_factors_cv][lam_cv][num_iterations] = mean_err
pickle.dump(H_store, open('../predictions/case-{}-graph_{}_{}_{}_{}_Hs.pkl'.format(case, source, constant_use, start, stop), 'w'))
pickle.dump(A_store, open('../predictions/case-{}-graph_{}_{}_{}_{}_As.pkl'.format(case, source, constant_use, start, stop), 'w'))
pickle.dump(err_store, open('../predictions/case-{}-graph_{}_{}_{}_{}_errs.pkl'.format(case, source, constant_use, start, stop), 'w'))
pd.DataFrame(
HAT[:num_test,
appliance_index[appliance], :],
index=test_ix))
H_factors[learning_rate_cv][num_iterations_cv][
num_season_factors_cv][num_home_factors_cv][
lam_cv].append(
pd.DataFrame(H[:num_test, :],
index=test_ix))
# get the overall prediction error of all homes
s = pd.concat(pred[appliance]).ix[target_df.index]
err = {}
for appliance in APPLIANCES_ORDER:
if appliance == "hvac":
err[appliance] = compute_rmse_fraction(
appliance, s[range(4, 10)], target)[2]
else:
err[appliance] = compute_rmse_fraction(
appliance, s, target)[2]
err_weight = {}
for appliance in APPLIANCES_ORDER[1:]:
err_weight[appliance] = err[appliance] * contri[
target][appliance]
mean_err = pd.Series(err_weight).sum()
print learning_rate_cv, num_iterations_cv, num_season_factors_cv, num_home_factors_cv, lam_cv, mean_err
error.append(mean_err)
params[count] = []
params[count].extend(
(learning_rate_cv, num_iterations_cv,
pred_inner = []
pred_inner.append(
create_prediction(test_ix_inner, X, Y, X_normalised,
appliance, col_max, col_min,
appliance_cols))
err = {}
pred_inner = pd.DataFrame(pd.concat(pred_inner))
try:
if appliance == "hvac":
err[appliance] = \
compute_rmse_fraction(appliance, pred_inner[['hvac_{}'.format(month) for month in range(5, 11)]],
'SanDiego')[2]
else:
err[appliance] = compute_rmse_fraction(
appliance, pred_inner, 'SanDiego')[2]
except Exception, e:
# This appliance does not have enough samples. Will not be
# weighed
print(e)
print(appliance)
err_weight = {}
mean_err = pd.Series(err).sum()
if mean_err < least_error:
best_num_iterations = num_iterations_cv
best_num_latent_factors = num_latent_factors_cv
least_error = mean_err