def RunModel(train_filename, test_filename, label_filename, config, ratio):
negative_sample = True if "noise" in config.dataset else False
train_data, abnormal_data, abnormal_label = read_dataset(
train_filename,
test_filename,
label_filename,
normalize=True,
file_logger=file_logger,
negative_sample=negative_sample,
ratio=ratio)
if abnormal_data.shape[0] < config.rolling_size:
train_logger.warning(
"test data is less than rolling_size! Ignore the current data!")
TN, FP, FN, TP, precision, recall, f1 = {}, {}, {}, {}, {}, {}, {}
for threshold_method in ["SD", "MAD", "IQR"]:
TN[threshold_method] = -1
FP[threshold_method] = -1
FN[threshold_method] = -1
TP[threshold_method] = -1
precision[threshold_method] = -1
recall[threshold_method] = -1
f1[threshold_method] = -1
roc_auc = -1
pr_auc = -1
metrics_result = MetricsResult(TN=TN,
FP=FP,
FN=FN,
TP=TP,
precision=precision,
recall=recall,
fbeta=f1,
pr_auc=pr_auc,
roc_auc=roc_auc)
return metrics_result
original_x_dim = abnormal_data.shape[1]
rolling_train_data = None
rolling_valid_data = None
if config.preprocessing:
if config.use_overlapping:
if train_data is not None:
rolling_train_data, rolling_abnormal_data, rolling_abnormal_label = rolling_window_2D(
train_data, config.rolling_size), rolling_window_2D(
abnormal_data, config.rolling_size), rolling_window_2D(
abnormal_label, config.rolling_size)
train_split_idx = int(rolling_train_data.shape[0] * 0.7)
rolling_train_data, rolling_valid_data = rolling_train_data[:train_split_idx], rolling_train_data[
train_split_idx:]
else:
rolling_abnormal_data, rolling_abnormal_label = rolling_window_2D(
abnormal_data, config.rolling_size), rolling_window_2D(
abnormal_label, config.rolling_size)
else:
if train_data is not None:
rolling_train_data, rolling_abnormal_data, rolling_abnormal_label = cutting_window_2D(
train_data, config.rolling_size), cutting_window_2D(
abnormal_data, config.rolling_size), cutting_window_2D(
abnormal_label, config.rolling_size)
train_split_idx = int(rolling_train_data.shape[0] * 0.7)
rolling_train_data, rolling_valid_data = rolling_train_data[:train_split_idx], rolling_train_data[
train_split_idx:]
else:
rolling_abnormal_data, rolling_abnormal_label = cutting_window_2D(
abnormal_data, config.rolling_size), cutting_window_2D(
abnormal_label, config.rolling_size)
else:
if train_data is not None:
rolling_train_data, rolling_abnormal_data, rolling_abnormal_label = np.expand_dims(
train_data,
axis=0), np.expand_dims(abnormal_data,
axis=0), np.expand_dims(abnormal_label,
axis=0)
train_split_idx = int(rolling_train_data.shape[0] * 0.7)
rolling_train_data, rolling_valid_data = rolling_train_data[:train_split_idx], rolling_train_data[
train_split_idx:]
else:
rolling_abnormal_data, rolling_abnormal_label = np.expand_dims(
abnormal_data, axis=0), np.expand_dims(abnormal_label, axis=0)
config.x_dim = rolling_abnormal_data.shape[2]
model = DONUT(file_name=train_filename, config=config)
model = model.to(device)
donut_output = None
if train_data is not None and config.robustness == False:
donut_output = model.fit(train_input=rolling_train_data,
train_label=rolling_train_data,
valid_input=rolling_valid_data,
valid_label=rolling_valid_data,
test_input=rolling_abnormal_data,
test_label=rolling_abnormal_label,
abnormal_data=abnormal_data,
abnormal_label=abnormal_label,
original_x_dim=original_x_dim)
elif train_data is None or config.robustness == True:
donut_output = model.fit(train_input=rolling_abnormal_data,
train_label=rolling_abnormal_data,
valid_input=rolling_valid_data,
valid_label=rolling_valid_data,
test_input=rolling_abnormal_data,
test_label=rolling_abnormal_label,
abnormal_data=abnormal_data,
abnormal_label=abnormal_label,
original_x_dim=original_x_dim)
# %%
min_max_scaler = preprocessing.MinMaxScaler()
if config.preprocessing:
if config.use_overlapping:
if config.use_last_point:
dec_mean_unroll = np.reshape(
donut_output.dec_means.detach().cpu().numpy(),
(-1, config.rolling_size, original_x_dim))[:, -1]
latent_mean_unroll = donut_output.zs.detach().cpu().numpy()
dec_mean_unroll = min_max_scaler.fit_transform(dec_mean_unroll)
x_original_unroll = abnormal_data[config.rolling_size - 1:]
else:
dec_mean_unroll = unroll_window_3D(
np.reshape(
donut_output.dec_means.detach().cpu().numpy(),
(-1, config.rolling_size, original_x_dim)))[::-1]
latent_mean_unroll = donut_output.zs.detach().cpu().numpy()
dec_mean_unroll = min_max_scaler.fit_transform(dec_mean_unroll)
x_original_unroll = abnormal_data[:dec_mean_unroll.shape[0]]
else:
dec_mean_unroll = np.reshape(
donut_output.dec_means.detach().cpu().numpy(),
(-1, original_x_dim))
latent_mean_unroll = donut_output.zs.detach().cpu().numpy()
dec_mean_unroll = min_max_scaler.fit_transform(dec_mean_unroll)
x_original_unroll = abnormal_data[:dec_mean_unroll.shape[0]]
else:
dec_mean_unroll = donut_output.dec_means.detach().cpu().numpy()
latent_mean_unroll = donut_output.zs.detach().cpu().numpy()
dec_mean_unroll = min_max_scaler.fit_transform(dec_mean_unroll)
x_original_unroll = abnormal_data
if config.save_output:
if not os.path.exists('./outputs/NPY/{}/'.format(config.dataset)):
os.makedirs('./outputs/NPY/{}/'.format(config.dataset))
np.save(
'./outputs/NPY/{}/Dec_DONUT_hdim_{}_rollingsize_{}_{}_pid={}.npy'.
format(config.dataset, config.h_dim, config.rolling_size,
train_filename.stem, config.pid), dec_mean_unroll)
np.save(
'./outputs/NPY/{}/Latent_DONUT_hdim_{}_rollingsize_{}_{}_pid={}.npy'
.format(config.dataset, config.h_dim, config.rolling_size,
train_filename.stem, config.pid), latent_mean_unroll)
error = np.sum(x_original_unroll -
np.reshape(dec_mean_unroll, [-1, original_x_dim]),
axis=1)**2
# final_zscore = zscore(error)
# np_decision = create_label_based_on_zscore(final_zscore, 2.5, True)
#np_decision = create_label_based_on_quantile(error, quantile=99)
SD_Tmin, SD_Tmax = SD_autothreshold(error)
SD_y_hat = get_labels_by_threshold(error,
Tmax=SD_Tmax,
use_max=True,
use_min=False)
MAD_Tmin, MAD_Tmax = MAD_autothreshold(error)
MAD_y_hat = get_labels_by_threshold(error,
Tmax=MAD_Tmax,
use_max=True,
use_min=False)
IQR_Tmin, IQR_Tmax = IQR_autothreshold(error)
IQR_y_hat = get_labels_by_threshold(error,
Tmax=IQR_Tmax,
use_max=True,
use_min=False)
np_decision = {}
np_decision["SD"] = SD_y_hat
np_decision["MAD"] = MAD_y_hat
np_decision["IQR"] = IQR_y_hat
# TODO metrics computation.
# %%
if config.save_figure:
if original_x_dim == 1:
plt.figure(figsize=(9, 3))
plt.plot(x_original_unroll, color='blue', lw=1.5)
plt.title('Original Data')
plt.grid(True)
plt.tight_layout()
# plt.show()
plt.savefig(
'./figures/{}/Ori_DONUT_hdim_{}_rollingsize_{}_{}_pid={}.png'.
format(config.dataset, config.h_dim, config.rolling_size,
train_filename.stem, config.pid),
dpi=600)
plt.close()
# Plot decoder output
plt.figure(figsize=(9, 3))
plt.plot(dec_mean_unroll, color='blue', lw=1.5)
plt.title('Decoding Output')
plt.grid(True)
plt.tight_layout()
# plt.show()
plt.savefig(
'./figures/{}/Dec_DONUT_hdim_{}_rollingsize_{}_{}_pid={}.png'.
format(config.dataset, config.h_dim, config.rolling_size,
train_filename.stem, config.pid),
dpi=600)
plt.close()
t = np.arange(0, abnormal_data.shape[0])
# markercolors = ['blue' if i == 1 else 'red' for i in abnormal_label[: dec_mean_unroll.shape[0]]]
# markersize = [4 if i == 1 else 25 for i in abnormal_label[: dec_mean_unroll.shape[0]]]
# plt.figure(figsize=(9, 3))
# ax = plt.axes()
# plt.yticks([0, 0.25, 0.5, 0.75, 1])
# ax.set_xlim(t[0] - 10, t[-1] + 10)
# ax.set_ylim(-0.10, 1.10)
# plt.xlabel('$t$')
# plt.ylabel('$s$')
# plt.grid(True)
# plt.tight_layout()
# plt.margins(0.1)
# plt.plot(abnormal_data[: dec_mean_unroll.shape[0]], alpha=0.7)
# plt.scatter(t[: dec_mean_unroll.shape[0]], x_original_unroll[: np_decision.shape[0]], s=markersize, c=markercolors)
# # plt.show()
# plt.savefig('./figures/{}/VisInp_DONUT_{}_pid={}.png'.format(config.dataset, Path(file_name).stem, config.pid), dpi=600)
# plt.close()
markercolors = ['blue' for i in range(config.rolling_size - 1)] + [
'blue' if i == 1 else 'red' for i in np_decision["SD"]
]
markersize = [4 for i in range(config.rolling_size - 1)
] + [4 if i == 1 else 25 for i in np_decision["SD"]]
plt.figure(figsize=(9, 3))
ax = plt.axes()
plt.yticks([0, 0.25, 0.5, 0.75, 1])
ax.set_xlim(t[0] - 10, t[-1] + 10)
ax.set_ylim(-0.10, 1.10)
plt.xlabel('$t$')
plt.ylabel('$s$')
plt.grid(True)
plt.tight_layout()
plt.margins(0.1)
plt.plot(abnormal_data, alpha=0.7)
plt.scatter(t, abnormal_data, s=markersize, c=markercolors)
# plt.show()
plt.savefig(
'./figures/{}/VisOut_DONUT_hdim_{}_rollingsize_{}_SD_{}_pid={}.png'
.format(config.dataset, config.h_dim, config.rolling_size,
train_filename.stem, config.pid),
dpi=300)
plt.close()
markercolors = ['blue' for i in range(config.rolling_size - 1)] + [
'blue' if i == 1 else 'red' for i in np_decision["MAD"]
]
markersize = [4 for i in range(config.rolling_size - 1)] + [
4 if i == 1 else 25 for i in np_decision["MAD"]
]
plt.figure(figsize=(9, 3))
ax = plt.axes()
plt.yticks([0, 0.25, 0.5, 0.75, 1])
ax.set_xlim(t[0] - 10, t[-1] + 10)
ax.set_ylim(-0.10, 1.10)
plt.xlabel('$t$')
plt.ylabel('$s$')
plt.grid(True)
plt.tight_layout()
plt.margins(0.1)
plt.plot(abnormal_data, alpha=0.7)
plt.scatter(t, abnormal_data, s=markersize, c=markercolors)
# plt.show()
plt.savefig(
'./figures/{}/VisOut_DONUT_hdim_{}_rollingsize_{}_MAD_{}_pid={}.png'
.format(config.dataset, config.h_dim, config.rolling_size,
train_filename.stem, config.pid),
dpi=300)
plt.close()
markercolors = ['blue' for i in range(config.rolling_size - 1)] + [
'blue' if i == 1 else 'red' for i in np_decision["IQR"]
]
markersize = [4 for i in range(config.rolling_size - 1)] + [
4 if i == 1 else 25 for i in np_decision["IQR"]
]
plt.figure(figsize=(9, 3))
ax = plt.axes()
plt.yticks([0, 0.25, 0.5, 0.75, 1])
ax.set_xlim(t[0] - 10, t[-1] + 10)
ax.set_ylim(-0.10, 1.10)
plt.xlabel('$t$')
plt.ylabel('$s$')
plt.grid(True)
plt.tight_layout()
plt.margins(0.1)
plt.plot(abnormal_data, alpha=0.7)
plt.scatter(t, abnormal_data, s=markersize, c=markercolors)
# plt.show()
plt.savefig(
'./figures/{}/VisOut_DONUT_hdim_{}_rollingsize_{}_IQR_{}_pid={}.png'
.format(config.dataset, config.h_dim, config.rolling_size,
train_filename.stem, config.pid),
dpi=300)
plt.close()
else:
file_logger.info('cannot plot image with x_dim > 1')
if config.use_spot:
pass
else:
pos_label = -1
TN, FP, FN, TP, precision, recall, f1 = {}, {}, {}, {}, {}, {}, {}
for threshold_method in np_decision:
cm = confusion_matrix(y_true=abnormal_label,
y_pred=np_decision[threshold_method],
labels=[1, -1])
TN[threshold_method] = cm[0][0]
FP[threshold_method] = cm[0][1]
FN[threshold_method] = cm[1][0]
TP[threshold_method] = cm[1][1]
precision[threshold_method] = precision_score(
y_true=abnormal_label,
y_pred=np_decision[threshold_method],
pos_label=pos_label)
recall[threshold_method] = recall_score(
y_true=abnormal_label,
y_pred=np_decision[threshold_method],
pos_label=pos_label)
f1[threshold_method] = f1_score(
y_true=abnormal_label,
y_pred=np_decision[threshold_method],
pos_label=pos_label)
fpr, tpr, _ = roc_curve(y_true=abnormal_label,
y_score=np.nan_to_num(error),
pos_label=pos_label)
roc_auc = auc(fpr, tpr)
pre, re, _ = precision_recall_curve(y_true=abnormal_label,
probas_pred=np.nan_to_num(error),
pos_label=pos_label)
pr_auc = auc(re, pre)
metrics_result = MetricsResult(
TN=TN,
FP=FP,
FN=FN,
TP=TP,
precision=precision,
recall=recall,
fbeta=f1,
pr_auc=pr_auc,
roc_auc=roc_auc,
best_TN=donut_output.best_TN,
best_FP=donut_output.best_FP,
best_FN=donut_output.best_FN,
best_TP=donut_output.best_TP,
best_precision=donut_output.best_precision,
best_recall=donut_output.best_recall,
best_fbeta=donut_output.best_fbeta,
best_pr_auc=donut_output.best_pr_auc,
best_roc_auc=donut_output.best_roc_auc,
best_cks=donut_output.best_cks)
return metrics_result
def RunModel(train_filename, test_filename, label_filename, config, ratio):
negative_sample = True if "noise" in config.dataset else False
train_data, abnormal_data, abnormal_label = read_dataset(
train_filename,
test_filename,
label_filename,
normalize=True,
file_logger=file_logger,
negative_sample=negative_sample,
ratio=ratio)
original_x_dim = abnormal_data.shape[1]
config.x_dim = abnormal_data.shape[1]
model = LOF(train_filename, config)
lof_output = model.fit(train_input=abnormal_data,
train_label=abnormal_label,
test_input=abnormal_data,
test_label=abnormal_label)
SD_Tmin, SD_Tmax = SD_autothreshold(-lof_output.negative_factor)
SD_y_hat = get_labels_by_threshold(-lof_output.negative_factor,
Tmax=SD_Tmax,
use_max=True,
use_min=False)
MAD_Tmin, MAD_Tmax = MAD_autothreshold(-lof_output.negative_factor)
MAD_y_hat = get_labels_by_threshold(-lof_output.negative_factor,
Tmax=MAD_Tmax,
use_max=True,
use_min=False)
IQR_Tmin, IQR_Tmax = IQR_autothreshold(-lof_output.negative_factor)
IQR_y_hat = get_labels_by_threshold(-lof_output.negative_factor,
Tmax=IQR_Tmax,
use_max=True,
use_min=False)
lof_output.y_hat = {}
lof_output.y_hat["SD"] = SD_y_hat
lof_output.y_hat["MAD"] = MAD_y_hat
lof_output.y_hat["IQR"] = IQR_y_hat
if config.save_output == True:
if not os.path.exists('./outputs/NPY/{}/'.format(config.dataset)):
os.makedirs('./outputs/NPY/{}/'.format(config.dataset))
np.save(
'./outputs/NPY/{}/Score_LOF_hdim_{}_rollingsize_{}_{}_pid={}.npy'.
format(config.dataset, config.n_neighbors, 1, train_filename.stem,
config.pid), lof_output.negative_factor)
np.save(
'./outputs/NPY/{}/Pred_LOF_hdim_{}_rollingsize_{}_{}_pid={}.npy'.
format(config.dataset, config.n_neighbors, 1, train_filename.stem,
config.pid), lof_output.y_hat)
# %%
if config.save_figure:
if not os.path.exists('./figures/{}/'.format(config.dataset)):
os.makedirs('./figures/{}/'.format(config.dataset))
if original_x_dim == 1:
plt.figure(figsize=(9, 3))
plt.plot(abnormal_data, color='blue', lw=1.5)
plt.title('Original Data')
plt.grid(True)
plt.tight_layout()
# plt.show()
plt.savefig(
'./figures/{}/Ori_LOF_hdim_{}_rollingsize_{}_{}_pid={}.png'.
format(config.dataset, config.n_neighbors, 1,
train_filename.stem, config.pid),
dpi=300)
plt.close()
t = np.arange(0, abnormal_data.shape[0])
markercolors = [
'blue' if i == 1 else 'red' for i in abnormal_label
]
markersize = [4 if i == 1 else 25 for i in abnormal_label]
plt.figure(figsize=(9, 3))
ax = plt.axes()
plt.yticks([0, 0.25, 0.5, 0.75, 1])
ax.set_xlim(t[0] - 10, t[-1] + 10)
ax.set_ylim(-0.10, 1.10)
plt.xlabel('$t$')
plt.ylabel('$s$')
plt.grid(True)
plt.tight_layout()
plt.margins(0.1)
plt.plot(abnormal_data, alpha=0.7)
plt.scatter(t, abnormal_data, s=markersize, c=markercolors)
# plt.show()
plt.savefig(
'./figures/{}/VisInp_LOF_hdim_{}_rollingsize_{}_{}_pid={}.png'.
format(config.dataset, config.n_neighbors, 1,
train_filename.stem, config.pid),
dpi=300)
plt.close()
markercolors = [
'blue' if i == 1 else 'red' for i in lof_output.y_hat["SD"]
]
markersize = [4 if i == 1 else 25 for i in lof_output.y_hat["SD"]]
plt.figure(figsize=(9, 3))
ax = plt.axes()
plt.yticks([0, 0.25, 0.5, 0.75, 1])
ax.set_xlim(t[0] - 10, t[-1] + 10)
ax.set_ylim(-0.10, 1.10)
plt.xlabel('$t$')
plt.ylabel('$s$')
plt.grid(True)
plt.tight_layout()
plt.margins(0.1)
plt.plot(abnormal_data, alpha=0.7)
plt.scatter(t, abnormal_data, s=markersize, c=markercolors)
# plt.show()
plt.savefig(
'./figures/{}/VisOut_LOF_hdim_{}_rollingsize_{}_SD_{}_pid={}.png'
.format(config.dataset, config.n_neighbors, 1,
train_filename.stem, config.pid),
dpi=300)
plt.close()
markercolors = [
'blue' if i == 1 else 'red' for i in lof_output.y_hat["MAD"]
]
markersize = [4 if i == 1 else 25 for i in lof_output.y_hat["MAD"]]
plt.figure(figsize=(9, 3))
ax = plt.axes()
plt.yticks([0, 0.25, 0.5, 0.75, 1])
ax.set_xlim(t[0] - 10, t[-1] + 10)
ax.set_ylim(-0.10, 1.10)
plt.xlabel('$t$')
plt.ylabel('$s$')
plt.grid(True)
plt.tight_layout()
plt.margins(0.1)
plt.plot(abnormal_data, alpha=0.7)
plt.scatter(t, abnormal_data, s=markersize, c=markercolors)
# plt.show()
plt.savefig(
'./figures/{}/VisOut_LOF_hdim_{}_rollingsize_{}_MAD_{}_pid={}.png'
.format(config.dataset, config.n_neighbors, 1,
train_filename.stem, config.pid),
dpi=300)
plt.close()
markercolors = [
'blue' if i == 1 else 'red' for i in lof_output.y_hat["IQR"]
]
markersize = [4 if i == 1 else 25 for i in lof_output.y_hat["IQR"]]
plt.figure(figsize=(9, 3))
ax = plt.axes()
plt.yticks([0, 0.25, 0.5, 0.75, 1])
ax.set_xlim(t[0] - 10, t[-1] + 10)
ax.set_ylim(-0.10, 1.10)
plt.xlabel('$t$')
plt.ylabel('$s$')
plt.grid(True)
plt.tight_layout()
plt.margins(0.1)
plt.plot(abnormal_data, alpha=0.7)
plt.scatter(t, abnormal_data, s=markersize, c=markercolors)
plt.savefig(
'./figures/{}/VisOut_LOF_hdim_{}_rollingsize_{}_IQR_{}_pid={}.png'
.format(config.dataset, config.n_neighbors, 1,
train_filename.stem, config.pid),
dpi=300)
plt.close()
else:
file_logger.info('cannot plot image with x_dim > 1')
if config.use_spot:
pass
else:
pos_label = -1
TN, FP, FN, TP, precision, recall, f1 = {}, {}, {}, {}, {}, {}, {}
for threshold_method in lof_output.y_hat:
cm = confusion_matrix(y_true=abnormal_label,
y_pred=lof_output.y_hat[threshold_method],
labels=[1, -1])
TN[threshold_method] = cm[0][0]
FP[threshold_method] = cm[0][1]
FN[threshold_method] = cm[1][0]
TP[threshold_method] = cm[1][1]
precision[threshold_method] = precision_score(
y_true=abnormal_label,
y_pred=lof_output.y_hat[threshold_method],
pos_label=pos_label)
recall[threshold_method] = recall_score(
y_true=abnormal_label,
y_pred=lof_output.y_hat[threshold_method],
pos_label=pos_label)
f1[threshold_method] = f1_score(
y_true=abnormal_label,
y_pred=lof_output.y_hat[threshold_method],
pos_label=pos_label)
fpr, tpr, _ = roc_curve(y_true=abnormal_label,
y_score=-lof_output.negative_factor,
pos_label=pos_label)
roc_auc = auc(fpr, tpr)
pre, re, _ = precision_recall_curve(
y_true=abnormal_label,
probas_pred=-lof_output.negative_factor,
pos_label=pos_label)
pr_auc = auc(re, pre)
metrics_result = MetricsResult(TN=TN,
FP=FP,
FN=FN,
TP=TP,
precision=precision,
recall=recall,
fbeta=f1,
pr_auc=pr_auc,
roc_auc=roc_auc)
return metrics_result
def RunModel(train_filename, test_filename, label_filename, config, ratio):
negative_sample = True if "noise" in config.dataset else False
train_data, abnormal_data, abnormal_label = read_dataset(
train_filename,
test_filename,
label_filename,
normalize=True,
file_logger=file_logger,
negative_sample=negative_sample,
ratio=ratio)
original_x_dim = abnormal_data.shape[1]
config.x_dim = abnormal_data.shape[1]
Pab = []
for i in range(abnormal_data.shape[1]):
ts = abnormal_data[:, i]
Pab_i, _ = stomp(ts, config.pattern_size)
Pab.append(np.nan_to_num(Pab_i))
Pab = np.sum(Pab, axis=0)
# final_zscore = zscore(Pab)
# np_decision = create_label_based_on_zscore(final_zscore, 2.5, True)
#np_decision = create_label_based_on_quantile(-Pab, quantile=99)
# higher -Pab is more likely to be anomalies.
SD_Tmin, SD_Tmax = SD_autothreshold(-Pab)
SD_y_hat = get_labels_by_threshold(-Pab,
Tmax=SD_Tmax,
use_max=True,
use_min=False)
MAD_Tmin, MAD_Tmax = MAD_autothreshold(-Pab)
MAD_y_hat = get_labels_by_threshold(-Pab,
Tmax=MAD_Tmax,
use_max=True,
use_min=False)
IQR_Tmin, IQR_Tmax = IQR_autothreshold(-Pab)
IQR_y_hat = get_labels_by_threshold(-Pab,
Tmax=IQR_Tmax,
use_max=True,
use_min=False)
np_decision = {}
np_decision["SD"] = SD_y_hat
np_decision["MAD"] = MAD_y_hat
np_decision["IQR"] = IQR_y_hat
if config.save_output:
if not os.path.exists('./outputs/NPY/{}/'.format(config.dataset)):
os.makedirs('./outputs/NPY/{}/'.format(config.dataset))
np.save(
'./outputs/NPY/{}/MP_hdim_None_rollingsize_{}_{}_pid={}.npy'.
format(config.dataset, config.pattern_size, train_filename.stem,
config.pid), Pab)
# TODO metrics computation.
# %%
if config.save_figure:
if original_x_dim == 1:
plt.figure(figsize=(9, 3))
plt.plot(ts, color='blue', lw=1.5)
plt.title('Original Data')
plt.grid(True)
plt.tight_layout()
# plt.show()
plt.savefig(
'./figures/{}/Ori_MP_hdim_None_rollingsize_{}_{}_pid={}.png'.
format(config.dataset, config.pattern_size,
train_filename.stem, config.pid),
dpi=300)
plt.close()
# Plot decoder output
plt.figure(figsize=(9, 3))
plt.plot(Pab, color='blue', lw=1.5)
plt.title('Profile Output')
plt.grid(True)
plt.tight_layout()
# plt.show()
plt.savefig(
'./figures/{}/Profile_MP_hdim_None_rollingsize_{}_{}_pid={}.png'
.format(config.dataset, config.pattern_size,
train_filename.stem, config.pid),
dpi=300)
plt.close()
t = np.arange(0, abnormal_data.shape[0])
markercolors = ['blue' for i in range(config.pattern_size - 1)] + [
'blue' if i == 1 else 'red'
for i in abnormal_label[config.pattern_size - 1:]
]
markersize = [4 for i in range(config.pattern_size - 1)] + [
4 if i == 1 else 25
for i in abnormal_label[config.pattern_size - 1:]
]
plt.figure(figsize=(9, 3))
ax = plt.axes()
plt.yticks([0, 0.25, 0.5, 0.75, 1])
ax.set_xlim(t[0] - 10, t[-1] + 10)
ax.set_ylim(-0.10, 1.10)
plt.xlabel('$t$')
plt.ylabel('$s$')
plt.grid(True)
plt.tight_layout()
plt.margins(0.1)
plt.plot(np.squeeze(abnormal_data), alpha=0.7)
plt.scatter(t, abnormal_data, s=markersize, c=markercolors)
# plt.show()
plt.savefig(
'./figures/{}/VisInp_MP_hdim_None_rollingsize_{}_{}_pid={}.png'
.format(config.dataset, config.pattern_size,
train_filename.stem, config.pid),
dpi=600)
plt.close()
markercolors = [
'blue' if i == 1 else 'red' for i in np_decision["SD"]
] + ['blue' for i in range(config.pattern_size - 1)]
markersize = [4 if i == 1 else 25 for i in np_decision["SD"]
] + [4 for i in range(config.pattern_size - 1)]
plt.figure(figsize=(9, 3))
ax = plt.axes()
plt.yticks([0, 0.25, 0.5, 0.75, 1])
ax.set_xlim(t[0] - 10, t[-1] + 10)
ax.set_ylim(-0.10, 1.10)
plt.xlabel('$t$')
plt.ylabel('$s$')
plt.grid(True)
plt.tight_layout()
plt.margins(0.1)
plt.plot(np.squeeze(abnormal_data), alpha=0.7)
plt.scatter(t, abnormal_data, s=markersize, c=markercolors)
# plt.show()
plt.savefig(
'./figures/{}/VisOut_MP_hdim_None_rollingsize_{}_SD_{}_pid={}.png'
.format(config.dataset, config.pattern_size,
train_filename.stem, config.pid),
dpi=300)
plt.close()
markercolors = [
'blue' if i == 1 else 'red' for i in np_decision["MAD"]
] + ['blue' for i in range(config.pattern_size - 1)]
markersize = [4 if i == 1 else 25 for i in np_decision["MAD"]
] + [4 for i in range(config.pattern_size - 1)]
plt.figure(figsize=(9, 3))
ax = plt.axes()
plt.yticks([0, 0.25, 0.5, 0.75, 1])
ax.set_xlim(t[0] - 10, t[-1] + 10)
ax.set_ylim(-0.10, 1.10)
plt.xlabel('$t$')
plt.ylabel('$s$')
plt.grid(True)
plt.tight_layout()
plt.margins(0.1)
plt.plot(np.squeeze(abnormal_data), alpha=0.7)
plt.scatter(t, abnormal_data, s=markersize, c=markercolors)
# plt.show()
plt.savefig(
'./figures/{}/VisOut_MP_hdim_None_rollingsize_{}_MAD_{}_pid={}.png'
.format(config.dataset, config.pattern_size,
train_filename.stem, config.pid),
dpi=300)
plt.close()
markercolors = [
'blue' if i == 1 else 'red' for i in np_decision["IQR"]
] + ['blue' for i in range(config.pattern_size - 1)]
markersize = [4 if i == 1 else 25 for i in np_decision["IQR"]
] + [4 for i in range(config.pattern_size - 1)]
plt.figure(figsize=(9, 3))
ax = plt.axes()
plt.yticks([0, 0.25, 0.5, 0.75, 1])
ax.set_xlim(t[0] - 10, t[-1] + 10)
ax.set_ylim(-0.10, 1.10)
plt.xlabel('$t$')
plt.ylabel('$s$')
plt.grid(True)
plt.tight_layout()
plt.margins(0.1)
plt.plot(np.squeeze(abnormal_data), alpha=0.7)
plt.scatter(t, abnormal_data, s=markersize, c=markercolors)
# plt.show()
plt.savefig(
'./figures/{}/VisOut_MP_hdim_None_rollingsize_{}_IQR_{}_pid={}.png'
.format(config.dataset, config.pattern_size,
train_filename.stem, config.pid),
dpi=300)
plt.close()
else:
file_logger.info('cannot plot image with x_dim > 1')
if config.use_spot:
pass
else:
pos_label = -1
TN, FP, FN, TP, precision, recall, f1 = {}, {}, {}, {}, {}, {}, {}
for threshold_method in np_decision:
cm = confusion_matrix(y_true=abnormal_label[config.pattern_size -
1:],
y_pred=np_decision[threshold_method],
labels=[1, -1])
TN[threshold_method] = cm[0][0]
FP[threshold_method] = cm[0][1]
FN[threshold_method] = cm[1][0]
TP[threshold_method] = cm[1][1]
precision[threshold_method] = precision_score(
y_true=abnormal_label[config.pattern_size - 1:],
y_pred=np_decision[threshold_method],
pos_label=pos_label)
recall[threshold_method] = recall_score(
y_true=abnormal_label[config.pattern_size - 1:],
y_pred=np_decision[threshold_method],
pos_label=pos_label)
f1[threshold_method] = f1_score(
y_true=abnormal_label[config.pattern_size - 1:],
y_pred=np_decision[threshold_method],
pos_label=pos_label)
fpr, tpr, _ = roc_curve(y_true=abnormal_label[config.pattern_size -
1:],
y_score=-Pab,
pos_label=pos_label)
roc_auc = auc(fpr, tpr)
pre, re, _ = precision_recall_curve(
y_true=abnormal_label[config.pattern_size - 1:],
probas_pred=-Pab,
pos_label=pos_label)
pr_auc = auc(re, pre)
metrics_result = MetricsResult(TN=TN,
FP=FP,
FN=FN,
TP=TP,
precision=precision,
recall=recall,
fbeta=f1,
pr_auc=pr_auc,
roc_auc=roc_auc)
return metrics_result
def RunModel(train_filename, test_filename, label_filename, config, ratio,
gpu_id):
negative_sample = True if "noise" in config.dataset else False
train_data, abnormal_data, abnormal_label = read_dataset(
train_filename,
test_filename,
label_filename,
normalize=True,
file_logger=file_logger,
negative_sample=negative_sample,
ratio=ratio)
original_x_dim = abnormal_data.shape[1]
if abnormal_data.shape[0] < config.rolling_size:
train_logger.warning(
"test data is less than rolling_size! Ignore the current data!")
TN, FP, FN, TP, precision, recall, f1 = {}, {}, {}, {}, {}, {}, {}
for threshold_method in ["SD", "MAD", "IQR"]:
TN[threshold_method] = -1
FP[threshold_method] = -1
FN[threshold_method] = -1
TP[threshold_method] = -1
precision[threshold_method] = -1
recall[threshold_method] = -1
f1[threshold_method] = -1
roc_auc = -1
pr_auc = -1
metrics_result = MetricsResult(TN=TN,
FP=FP,
FN=FN,
TP=TP,
precision=precision,
recall=recall,
fbeta=f1,
pr_auc=pr_auc,
roc_auc=roc_auc)
return metrics_result
rolling_train_data = None
rolling_valid_data = None
if config.preprocessing:
if config.use_overlapping:
if train_data is not None:
rolling_train_data, rolling_abnormal_data, rolling_abnormal_label = rolling_window_2D(
train_data, config.rolling_size), rolling_window_2D(
abnormal_data, config.rolling_size), rolling_window_2D(
abnormal_label, config.rolling_size)
train_split_idx = int(rolling_train_data.shape[0] * 0.7)
rolling_train_data, rolling_valid_data = rolling_train_data[:train_split_idx], rolling_train_data[
train_split_idx:]
else:
rolling_abnormal_data, rolling_abnormal_label = rolling_window_2D(
abnormal_data, config.rolling_size), rolling_window_2D(
abnormal_label, config.rolling_size)
else:
if train_data is not None:
rolling_train_data, rolling_abnormal_data, rolling_abnormal_label = cutting_window_2D(
train_data, config.rolling_size), cutting_window_2D(
abnormal_data, config.rolling_size), cutting_window_2D(
abnormal_label, config.rolling_size)
train_split_idx = int(rolling_train_data.shape[0] * 0.7)
rolling_train_data, rolling_valid_data = rolling_train_data[:train_split_idx], rolling_train_data[
train_split_idx:]
else:
rolling_abnormal_data, rolling_abnormal_label = cutting_window_2D(
abnormal_data, config.rolling_size), cutting_window_2D(
abnormal_label, config.rolling_size)
if train_data is not None:
rolling_train_data, rolling_valid_data, rolling_abnormal_data, rolling_abnormal_label = np.reshape(
rolling_train_data, [rolling_train_data.shape[0], rolling_train_data.shape[1] * rolling_train_data.shape[2]]), \
np.reshape(
rolling_train_data, [rolling_train_data.shape[0], rolling_train_data.shape[1] * rolling_train_data.shape[2]]), \
np.reshape(
rolling_abnormal_data, [rolling_abnormal_data.shape[0], rolling_abnormal_data.shape[1] * rolling_abnormal_data.shape[2]]), \
np.reshape(
rolling_abnormal_label, [rolling_abnormal_label.shape[0], rolling_abnormal_label.shape[1] * rolling_abnormal_label.shape[2]])
else:
rolling_abnormal_data, rolling_abnormal_label = np.reshape(
rolling_abnormal_data, [
rolling_abnormal_data.shape[0],
rolling_abnormal_data.shape[1] *
rolling_abnormal_data.shape[2]
]), np.reshape(rolling_abnormal_label, [
rolling_abnormal_label.shape[0],
rolling_abnormal_label.shape[1] *
rolling_abnormal_label.shape[2]
])
else:
if train_data is not None:
rolling_train_data, rolling_abnormal_data, rolling_abnormal_label = train_data, abnormal_data, abnormal_label
else:
rolling_abnormal_data, rolling_abnormal_label = abnormal_data, abnormal_label
config.x_dim = rolling_abnormal_data.shape[1]
ensemble_error = []
ensemble_output = []
ensemble_TN = []
ensemble_TP = []
ensemble_FN = []
ensemble_FP = []
ensemble_PRECISION = []
ensemble_RECALL = []
ensemble_FBETA = []
ensemble_PR_AUC = []
ensemble_ROC_AUC = []
ensemble_CKS = []
# only to show the results
for i in range(config.ensemble_space):
train_logger.info('component #{}'.format(i))
model = RN(file_name=train_filename, config=config, gpu_id=gpu_id)
model = model.to(device)
if train_data is not None and config.robustness == False:
rn_output = model.fit(train_input=rolling_train_data,
train_label=rolling_train_data,
valid_input=rolling_valid_data,
valid_label=rolling_valid_data,
test_input=rolling_abnormal_data,
test_label=rolling_abnormal_label,
abnormal_data=abnormal_data,
abnormal_label=abnormal_label,
original_x_dim=original_x_dim)
elif train_data is None or config.robustness == True:
rn_output = model.fit(train_input=rolling_abnormal_data,
train_label=rolling_abnormal_data,
valid_input=rolling_valid_data,
valid_label=rolling_valid_data,
test_input=rolling_abnormal_data,
test_label=rolling_abnormal_label,
abnormal_data=abnormal_data,
abnormal_label=abnormal_label,
original_x_dim=original_x_dim)
# %%
min_max_scaler = preprocessing.MinMaxScaler()
if config.preprocessing:
if config.use_overlapping:
if config.use_last_point:
dec_mean_unroll = np.reshape(
rn_output.dec_means.detach().cpu().numpy(),
(-1, config.rolling_size, original_x_dim))[:, -1]
dec_mean_unroll = min_max_scaler.fit_transform(
dec_mean_unroll)
x_original_unroll = abnormal_data[config.rolling_size - 1:]
else:
dec_mean_unroll = unroll_window_3D(
np.reshape(
rn_output.dec_means.detach().cpu().numpy(),
(-1, config.rolling_size, original_x_dim)))[::-1]
dec_mean_unroll = min_max_scaler.fit_transform(
dec_mean_unroll)
x_original_unroll = abnormal_data[:dec_mean_unroll.
shape[0]]
else:
dec_mean_unroll = np.reshape(
rn_output.dec_means.detach().cpu().numpy(),
(-1, original_x_dim))
dec_mean_unroll = min_max_scaler.fit_transform(dec_mean_unroll)
x_original_unroll = abnormal_data[:dec_mean_unroll.shape[0]]
else:
dec_mean_unroll = rn_output.dec_means.detach().cpu().numpy()
dec_mean_unroll = min_max_scaler.fit_transform(dec_mean_unroll)
x_original_unroll = abnormal_data
if config.save_output:
np.save(
'./save_outputs/NPY/{}/Dec_RN_{}_{}_pid={}.npy'.format(
config.dataset,
Path(train_filename).stem, i, config.pid), dec_mean_unroll)
error = np.sum(abnormal_data[:dec_mean_unroll.shape[0]] -
np.reshape(dec_mean_unroll, [-1, original_x_dim]),
axis=1)**2
ensemble_error.append(error)
ensemble_output.append(dec_mean_unroll)
ensemble_FN.append(rn_output.best_FN)
ensemble_TN.append(rn_output.best_TN)
ensemble_FP.append(rn_output.best_FP)
ensemble_TP.append(rn_output.best_TP)
ensemble_PRECISION.append(rn_output.best_precision)
ensemble_RECALL.append(rn_output.best_recall)
ensemble_FBETA.append(rn_output.best_fbeta)
ensemble_PR_AUC.append(rn_output.best_pr_auc)
ensemble_ROC_AUC.append(rn_output.best_roc_auc)
ensemble_CKS.append(rn_output.best_cks)
error = np.stack(ensemble_error, axis=0)
error = np.median(error, axis=0)
dec_mean_unroll = np.stack(ensemble_output, axis=0)
dec_mean_unroll = np.median(dec_mean_unroll, axis=0)
SD_Tmin, SD_Tmax = SD_autothreshold(error)
SD_y_hat = get_labels_by_threshold(error,
Tmax=SD_Tmax,
use_max=True,
use_min=False)
MAD_Tmin, MAD_Tmax = MAD_autothreshold(error)
MAD_y_hat = get_labels_by_threshold(error,
Tmax=MAD_Tmax,
use_max=True,
use_min=False)
IQR_Tmin, IQR_Tmax = IQR_autothreshold(error)
IQR_y_hat = get_labels_by_threshold(error,
Tmax=IQR_Tmax,
use_max=True,
use_min=False)
np_decision = {}
np_decision["SD"] = SD_y_hat
np_decision["MAD"] = MAD_y_hat
np_decision["IQR"] = IQR_y_hat
# TODO metrics computation.
# TODO save output
if config.save_output:
np.save(
'./save_outputs/NPY/{}/Dec_RN_{}_pid={}.npy'.format(
config.dataset,
Path(train_filename).stem, config.pid), dec_mean_unroll)
np.save(
'./save_outputs/NPY/{}/Error_RN_{}_pid={}.npy'.format(
config.dataset,
Path(train_filename).stem, config.pid), error)
# %%
if config.save_figure:
if original_x_dim == 1:
plt.figure(figsize=(9, 3))
plt.plot(x_original_unroll, color='blue', lw=1.5)
plt.title('Original Data')
plt.grid(True)
plt.tight_layout()
# plt.show()
plt.savefig('./save_figures/{}/Ori_RN_{}_pid={}.png'.format(
config.dataset,
Path(train_filename).stem, config.pid),
dpi=300)
plt.close()
# Plot decoder output
plt.figure(figsize=(9, 3))
plt.plot(dec_mean_unroll, color='blue', lw=1.5)
plt.title('Decoding Output')
plt.grid(True)
plt.tight_layout()
# plt.show()
plt.savefig('./save_figures/{}/Dec_RN_{}_pid={}.png'.format(
config.dataset,
Path(train_filename).stem, config.pid),
dpi=300)
plt.close()
t = np.arange(0, abnormal_data.shape[0])
markercolors = [
'blue' if i == 1 else 'red'
for i in abnormal_label[:dec_mean_unroll.shape[0]]
]
markersize = [
4 if i == 1 else 25
for i in abnormal_label[:dec_mean_unroll.shape[0]]
]
plt.figure(figsize=(9, 3))
ax = plt.axes()
plt.yticks([0, 0.25, 0.5, 0.75, 1])
ax.set_xlim(t[0] - 10, t[-1] + 10)
ax.set_ylim(-0.10, 1.10)
plt.xlabel('$t$')
plt.ylabel('$s$')
plt.grid(True)
plt.tight_layout()
plt.margins(0.1)
plt.plot(abnormal_data[:dec_mean_unroll.shape[0]], alpha=0.7)
plt.scatter(t[:dec_mean_unroll.shape[0]],
x_original_unroll[:np_decision.shape[0]],
s=markersize,
c=markercolors)
# plt.show()
plt.savefig('./save_figures/{}/VisInp_RN_{}_pid={}.png'.format(
config.dataset,
Path(train_filename).stem, config.pid),
dpi=300)
plt.close()
markercolors = ['blue' for i in range(config.rolling_size - 1)] + [
'blue' if i == 1 else 'red' for i in np_decision
]
markersize = [4 for i in range(config.rolling_size - 1)
] + [4 if i == 1 else 25 for i in np_decision]
plt.figure(figsize=(9, 3))
ax = plt.axes()
plt.yticks([0, 0.25, 0.5, 0.75, 1])
ax.set_xlim(t[0] - 10, t[-1] + 10)
ax.set_ylim(-0.10, 1.10)
plt.xlabel('$t$')
plt.ylabel('$s$')
plt.grid(True)
plt.tight_layout()
plt.margins(0.1)
plt.plot(abnormal_data, alpha=0.7)
plt.scatter(t, abnormal_data, s=markersize, c=markercolors)
# plt.show()
plt.savefig('./save_figures/{}/VisOut_RN_{}_pid={}.png'.format(
config.dataset,
Path(train_filename).stem, config.pid),
dpi=300)
plt.close()
else:
file_logger.info('cannot plot image with x_dim > 1')
if config.use_spot:
pass
else:
try:
pos_label = -1
TN, FP, FN, TP, precision, recall, f1, cks = {}, {}, {}, {}, {}, {}, {}, {}
for threshold_method in np_decision:
cm = confusion_matrix(
y_true=abnormal_label[config.rolling_size - 1:],
y_pred=np_decision[threshold_method],
labels=[1, -1])
TN[threshold_method] = cm[0][0]
FP[threshold_method] = cm[0][1]
FN[threshold_method] = cm[1][0]
TP[threshold_method] = cm[1][1]
precision[threshold_method] = precision_score(
y_true=abnormal_label[config.rolling_size - 1:],
y_pred=np_decision[threshold_method],
pos_label=pos_label)
recall[threshold_method] = recall_score(
y_true=abnormal_label[config.rolling_size - 1:],
y_pred=np_decision[threshold_method],
pos_label=pos_label)
f1[threshold_method] = f1_score(
y_true=abnormal_label[config.rolling_size - 1:],
y_pred=np_decision[threshold_method],
pos_label=pos_label)
cks[threshold_method] = cohen_kappa_score(
y1=abnormal_label[config.rolling_size - 1:],
y2=np_decision[threshold_method])
fpr, tpr, _ = roc_curve(y_true=abnormal_label[config.rolling_size -
1:],
y_score=np.nan_to_num(error),
pos_label=pos_label)
roc_auc = auc(fpr, tpr)
pre, re, _ = precision_recall_curve(
y_true=abnormal_label[config.rolling_size - 1:],
probas_pred=np.nan_to_num(error),
pos_label=pos_label)
pr_auc = auc(re, pre)
metrics_result = MetricsResult(
TN=TN,
FP=FP,
FN=FN,
TP=TP,
precision=precision,
recall=recall,
fbeta=f1,
pr_auc=pr_auc,
roc_auc=roc_auc,
best_TN=rn_output.best_TN,
best_FP=rn_output.best_FP,
best_FN=rn_output.best_FN,
best_TP=rn_output.best_TP,
best_precision=rn_output.best_precision,
best_recall=rn_output.best_recall,
best_fbeta=rn_output.best_fbeta,
best_pr_auc=rn_output.best_pr_auc,
best_roc_auc=rn_output.best_roc_auc,
best_cks=rn_output.best_cks,
min_valid_loss=rn_output.min_valid_loss,
testing_time=rn_output.testing_time,
training_time=rn_output.training_time,
memory_usage_nvidia=rn_output.memory_usage_nvidia)
return metrics_result
except:
pass
def RunModel(train_filename, test_filename, label_filename, config, ratio):
negative_sample = True if "noise" in config.dataset else False
train_data, abnormal_data, abnormal_label = read_dataset(
train_filename,
test_filename,
label_filename,
normalize=True,
file_logger=file_logger,
negative_sample=negative_sample,
ratio=ratio)
if abnormal_data.shape[0] < config.rolling_size:
train_logger.warning(
"test data is less than rolling_size! Ignore the current data!")
TN, FP, FN, TP, precision, recall, f1 = {}, {}, {}, {}, {}, {}, {}
for threshold_method in ["SD", "MAD", "IQR"]:
TN[threshold_method] = -1
FP[threshold_method] = -1
FN[threshold_method] = -1
TP[threshold_method] = -1
precision[threshold_method] = -1
recall[threshold_method] = -1
f1[threshold_method] = -1
roc_auc = -1
pr_auc = -1
metrics_result = MetricsResult(TN=TN,
FP=FP,
FN=FN,
TP=TP,
precision=precision,
recall=recall,
fbeta=f1,
pr_auc=pr_auc,
roc_auc=roc_auc)
return metrics_result
original_x_dim = abnormal_data.shape[1]
rolling_train_data = None
rolling_valid_data = None
if config.preprocessing:
if config.use_overlapping:
if train_data is not None:
rolling_train_data, rolling_abnormal_data, rolling_abnormal_label = rolling_window_2D(
train_data, config.rolling_size), rolling_window_2D(
abnormal_data, config.rolling_size), rolling_window_2D(
abnormal_label, config.rolling_size)
train_split_idx = int(rolling_train_data.shape[0] * 0.7)
rolling_train_data, rolling_valid_data = rolling_train_data[:train_split_idx], rolling_train_data[
train_split_idx:]
else:
rolling_abnormal_data, rolling_abnormal_label = rolling_window_2D(
abnormal_data, config.rolling_size), rolling_window_2D(
abnormal_label, config.rolling_size)
else:
if train_data is not None:
rolling_train_data, rolling_abnormal_data, rolling_abnormal_label = cutting_window_2D(
train_data, config.rolling_size), cutting_window_2D(
abnormal_data, config.rolling_size), cutting_window_2D(
abnormal_label, config.rolling_size)
train_split_idx = int(rolling_train_data.shape[0] * 0.7)
rolling_train_data, rolling_valid_data = rolling_train_data[:train_split_idx], rolling_train_data[
train_split_idx:]
else:
rolling_abnormal_data, rolling_abnormal_label = cutting_window_2D(
abnormal_data, config.rolling_size), cutting_window_2D(
abnormal_label, config.rolling_size)
else:
if train_data is not None:
rolling_train_data, rolling_abnormal_data, rolling_abnormal_label = np.expand_dims(
train_data,
axis=0), np.expand_dims(abnormal_data,
axis=0), np.expand_dims(abnormal_label,
axis=0)
train_split_idx = int(rolling_train_data.shape[0] * 0.7)
rolling_train_data, rolling_valid_data = rolling_train_data[:train_split_idx], rolling_train_data[
train_split_idx:]
else:
rolling_abnormal_data, rolling_abnormal_label = np.expand_dims(
abnormal_data, axis=0), np.expand_dims(abnormal_label, axis=0)
config.x_dim = rolling_abnormal_data.shape[1]
model = CAE(file_name=train_filename, config=config)
model = model.to(device)
cae_output = None
if train_data is not None and config.robustness == False:
cae_output = model.fit(train_input=rolling_train_data,
train_label=rolling_train_data,
valid_input=rolling_valid_data,
valid_label=rolling_valid_data,
test_input=rolling_abnormal_data,
test_label=rolling_abnormal_label,
abnormal_data=abnormal_data,
abnormal_label=abnormal_label,
original_x_dim=original_x_dim)
elif train_data is None or config.robustness == True:
cae_output = model.fit(train_input=rolling_abnormal_data,
train_label=rolling_abnormal_data,
valid_input=rolling_valid_data,
valid_label=rolling_valid_data,
test_input=rolling_abnormal_data,
test_label=rolling_abnormal_label,
abnormal_data=abnormal_data,
abnormal_label=abnormal_label,
original_x_dim=original_x_dim)
# %%
min_max_scaler = preprocessing.MinMaxScaler()
if config.preprocessing:
if config.use_overlapping:
if config.use_last_point:
dec_mean_unroll = cae_output.dec_means.detach().cpu().numpy(
)[:, -1]
dec_mean_unroll = min_max_scaler.fit_transform(dec_mean_unroll)
x_original_unroll = abnormal_data[config.rolling_size - 1:]
else:
dec_mean_unroll = unroll_window_3D(
np.reshape(
cae_output.dec_means.detach().cpu().numpy(),
(-1, config.rolling_size, original_x_dim)))[::-1]
dec_mean_unroll = min_max_scaler.fit_transform(dec_mean_unroll)
x_original_unroll = abnormal_data[:dec_mean_unroll.shape[0]]
else:
dec_mean_unroll = np.reshape(
cae_output.dec_means.detach().cpu().numpy(),
(-1, original_x_dim))
dec_mean_unroll = min_max_scaler.fit_transform(dec_mean_unroll)
x_original_unroll = abnormal_data[:dec_mean_unroll.shape[0]]
else:
dec_mean_unroll = cae_output.dec_means.detach().cpu().numpy()
dec_mean_unroll = np.transpose(np.squeeze(dec_mean_unroll, axis=0))
dec_mean_unroll = min_max_scaler.fit_transform(dec_mean_unroll)
x_original_unroll = abnormal_data
if config.save_output:
if not os.path.exists('./outputs/NPY/{}/'.format(config.dataset)):
os.makedirs('./outputs/NPY/{}/'.format(config.dataset))
np.save(
'./outputs/NPY/{}/Dec_CAE_hdim_{}_rollingsize_{}_{}_pid={}.npy'.
format(config.dataset, config.h_dim, config.rolling_size,
Path(train_filename).stem, config.pid), dec_mean_unroll)
error = np.sum(x_original_unroll -
np.reshape(dec_mean_unroll, [-1, original_x_dim]),
axis=1)**2
# final_zscore = zscore(error)
# np_decision = create_label_based_on_zscore(final_zscore, 2.5, True)
# np_decision = create_label_based_on_quantile(error, quantile=99)
SD_Tmin, SD_Tmax = SD_autothreshold(error)
SD_y_hat = get_labels_by_threshold(error,
Tmax=SD_Tmax,
use_max=True,
use_min=False)
MAD_Tmin, MAD_Tmax = MAD_autothreshold(error)
MAD_y_hat = get_labels_by_threshold(error,
Tmax=MAD_Tmax,
use_max=True,
use_min=False)
IQR_Tmin, IQR_Tmax = IQR_autothreshold(error)
IQR_y_hat = get_labels_by_threshold(error,
Tmax=IQR_Tmax,
use_max=True,
use_min=False)
np_decision = {}
np_decision["SD"] = SD_y_hat
np_decision["MAD"] = MAD_y_hat
np_decision["IQR"] = IQR_y_hat
# TODO metrics computation.
# %%
if config.save_figure:
file_logger.info('save_figure has been dropped.')
if config.use_spot:
pass
else:
pos_label = -1
TN, FP, FN, TP, precision, recall, f1 = {}, {}, {}, {}, {}, {}, {}
for threshold_method in np_decision:
cm = confusion_matrix(y_true=abnormal_label,
y_pred=np_decision[threshold_method],
labels=[1, -1])
TN[threshold_method] = cm[0][0]
FP[threshold_method] = cm[0][1]
FN[threshold_method] = cm[1][0]
TP[threshold_method] = cm[1][1]
precision[threshold_method] = precision_score(
y_true=abnormal_label,
y_pred=np_decision[threshold_method],
pos_label=pos_label)
recall[threshold_method] = recall_score(
y_true=abnormal_label,
y_pred=np_decision[threshold_method],
pos_label=pos_label)
f1[threshold_method] = f1_score(
y_true=abnormal_label,
y_pred=np_decision[threshold_method],
pos_label=pos_label)
fpr, tpr, _ = roc_curve(y_true=abnormal_label,
y_score=np.nan_to_num(error),
pos_label=pos_label)
roc_auc = auc(fpr, tpr)
pre, re, _ = precision_recall_curve(y_true=abnormal_label,
probas_pred=np.nan_to_num(error),
pos_label=pos_label)
pr_auc = auc(re, pre)
metrics_result = MetricsResult(
TN=TN,
FP=FP,
FN=FN,
TP=TP,
precision=precision,
recall=recall,
fbeta=f1,
pr_auc=pr_auc,
roc_auc=roc_auc,
best_TN=cae_output.best_TN,
best_FP=cae_output.best_FP,
best_FN=cae_output.best_FN,
best_TP=cae_output.best_TP,
best_precision=cae_output.best_precision,
best_recall=cae_output.best_recall,
best_fbeta=cae_output.best_fbeta,
best_pr_auc=cae_output.best_pr_auc,
best_roc_auc=cae_output.best_roc_auc,
best_cks=cae_output.best_cks,
min_valid_loss=cae_output.min_valid_loss)
return metrics_result