batch_xt, batch_yt = [], []
previous_xt.append(batch_xt)
previous_yt.append(batch_xt)
prequential_acc.append(
nn_score(model_f, model_c, [batch_xs], [batch_ys], [], [],
drift_num))
no_drift_count += 1
if len(prequential_acc) > label_lag - 1 and dd.set_input(
1 - prequential_acc[-label_lag]):
#first_training_index = np.min([first_training_index, i-label_lag])
start_time = time.time()
print('CHANGE DETECTED at ' + str(i))
drift_list.append(i)
#warning_index.append(no_drift_count)
print('retrain using dataset index ' +
str(keep_last_consecutive(warning_index)))
model_f = models.Net_f(task=task, outdim=outdim).cuda()
model_c = models.Net_c_cway(task=task, outdim=outdim).cuda()
optimizer_f = torch.optim.Adam(model_f.parameters(), 0.001)
optimizer_c = torch.optim.Adam(model_c.parameters(), 0.001)
train_xs, train_ys = [], []
train_xt, train_yt = [], []
for j in keep_last_consecutive(warning_index):
train_xs.append(previous_xs[j])
train_ys.append(previous_ys[j])
if len(previous_xt[j]) == 0:
continue
train_xt.append(previous_xt[j])
train_yt.append(previous_yt[j])
train_clf(model_f, model_c, train_xs, train_ys, train_xt, train_yt,
drift_num, optimizer_f, optimizer_c)
loss_sample.backward(retain_graph=True)
#loss_sample.backward()
kfac_optim.acc_stats = False
kfac_optim.zero_grad() # clear the gradient for computing true-fisher.
loss.backward()
fisher_score = kfac_optim.step()
else:
fisher_score = 0.0
qFS_list.append(fisher_score)
prequential_acc.append(nn_score(model_f, model_c, [previous_xtogether[-1]], [previous_ytogether[-1]]))
if q1_drift == False:
if dd.set_input(q1_list[-1])==True:
q1_drift=True
if warning_index!=[]:
first_training_index = np.min([first_training_index, keep_last_consecutive(warning_index)[0]])
else:
first_training_index = np.min([first_training_index, i])
drift_1.append(i)
elif dd.is_warning_zone:
warning_index.append(i)
if q2_drift == False:
if dd_2.set_input(q2_list[-1])==True:
q2_drift=True
if warning_index_2!=[]:
first_training_index = np.min([first_training_index, keep_last_consecutive(warning_index_2)[0]])
else:
first_training_index = np.min([first_training_index, i])
drift_2.append(i)
elif dd_2.is_warning_zone:
warning_index_2.append(i)
retraining_time = 0
total_retraining_samples = 0
total_added_samples = 0
ret_ind = []
for i in range(initial_batches + label_lag + 50, n_batch):
prequential_acc.append(clf2.score(batch_X[i], batch_Y[i]))
if dd.set_input(1 - clf2.score(batch_X[i - 3], batch_Y[i - 3])):
start_time = time.time()
print('CHANGE DETECTED at ' + str(i))
drift_list.append(i)
warning_index.append(i)
print('retrain using dataset index ' +
str(keep_last_consecutive(warning_index)))
clf2 = classification_method(n_estimators=20)
ret_ind.append(keep_last_consecutive(warning_index))
train_X, train_Y = np.concatenate([
batch_X[j] for j in keep_last_consecutive(warning_index)
]), np.concatenate(
[batch_Y[j] for j in keep_last_consecutive(warning_index)])
clf2.fit(train_X, train_Y)
warning_index = []
retraining_time += (time.time() - start_time)
total_retraining_samples += train_X.shape[0]
if dd.is_warning_zone:
warning_index.append(i)
### PH in batch
if args.model == 'ph':
kfac_optim.zero_grad() # clear the gradient for computing true-fisher.
loss.backward()
fisher_score = kfac_optim.step()
qFS_list.append(fisher_score)
prequential_acc.append(
clf2.score(scaler.transform(batch_X[i]), batch_Y[i]))
if q1_drift == False:
if dd.set_input(q1_list[-1]) == True:
q1_drift = True
if warning_index != []:
first_training_index = np.min([
first_training_index,
keep_last_consecutive(warning_index)[0]
])
else:
first_training_index = np.min([first_training_index, i])
drift_1.append(i)
elif dd.is_warning_zone:
warning_index.append(i)
if q2_drift == False:
if dd_2.set_input(q2_list[-1]) == True:
q2_drift = True
if warning_index_2 != []:
first_training_index = np.min([
first_training_index,
keep_last_consecutive(warning_index_2)[0]
])
else: