def LF_experiment(angle, reps=1, ntrees=10, acorn=None):
errors = np.zeros(2)
for rep in range(reps):
print("Starting Rep {} of Angle {}".format(rep, angle))
X_base_train, y_base_train = generate_gaussian_parity(n = 100, angle_params = 0)
X_base_test, y_base_test = generate_gaussian_parity(n = 10000, angle_params = 0)
X_rotated_train, y_rotated_train = generate_gaussian_parity(n = 100, angle_params = angle)
lifelong_forest = LifeLongDNN(model = "uf", parallel = True)
lifelong_forest.new_forest(X_base_train, y_base_train, n_estimators=ntrees)
lifelong_forest.new_forest(X_rotated_train, y_rotated_train, n_estimators=ntrees)
all_predictions_test=lifelong_forest.predict(X_base_test, representation='all', decider=0)
base_predictions_test=lifelong_forest.predict(X_base_test, representation=0, decider=0)
errors[1] = errors[1]+(1 - np.mean(all_predictions_test == y_base_test))
errors[0] = errors[0]+(1 - np.mean(base_predictions_test == y_base_test))
errors = errors/reps
print("Errors For Angle {}: {}".format(angle, errors))
with open('results/angle_'+str(angle)+'.pickle', 'wb') as f:
pickle.dump(errors, f, protocol = 2)
def exp(n_sample, n_test, angle_params, n_trees, reps, acorn=None):
if acorn != None:
np.random.seed(acorn)
error = np.zeros(reps, dtype=float)
for i in range(reps):
train, label = generate_gaussian_parity(n_sample,
cov_scale=0.1,
angle_params=angle_params)
test, test_label = generate_gaussian_parity(n_test,
cov_scale=0.1,
angle_params=angle_params)
l2f = LifeLongDNN()
l2f.new_forest(train,
label,
n_estimators=n_trees,
max_samples=ceil(log2(n_sample)))
uf_task = l2f.predict(test, representation=0, decider=0)
error[i] = 1 - np.sum(uf_task == test_label) / n_test
return np.mean(error, axis=0), np.std(error, ddof=1, axis=0)
def LF_experiment(data_x,
data_y,
angle,
model,
granularity,
reps=1,
ntrees=29,
acorn=None):
if acorn is not None:
np.random.seed(acorn)
errors = np.zeros(2)
for rep in range(reps):
print("Starting Rep {} of Angle {}".format(rep, angle))
train_x1, train_y1, train_x2, train_y2, test_x, test_y = cross_val_data(
data_x, data_y, total_cls=10)
#change data angle for second task
tmp_data = train_x2.copy()
_tmp_ = np.zeros((32, 32, 3), dtype=int)
total_data = tmp_data.shape[0]
for i in range(total_data):
tmp_ = image_aug(tmp_data[i], angle)
tmp_data[i] = tmp_
if model == "uf":
train_x1 = train_x1.reshape(
(train_x1.shape[0],
train_x1.shape[1] * train_x1.shape[2] * train_x1.shape[3]))
tmp_data = tmp_data.reshape(
(tmp_data.shape[0],
tmp_data.shape[1] * tmp_data.shape[2] * tmp_data.shape[3]))
test_x = test_x.reshape(
(test_x.shape[0],
test_x.shape[1] * test_x.shape[2] * test_x.shape[3]))
with tf.device('/gpu:' + str(int(angle // granularity) % 4)):
lifelong_forest = LifeLongDNN(
model=model, parallel=True if model == "uf" else False)
lifelong_forest.new_forest(train_x1, train_y1, n_estimators=ntrees)
lifelong_forest.new_forest(tmp_data, train_y2, n_estimators=ntrees)
llf_task1 = lifelong_forest.predict(test_x,
representation='all',
decider=0)
llf_single_task = lifelong_forest.predict(test_x,
representation=0,
decider=0)
errors[1] = errors[1] + (1 - np.mean(llf_task1 == test_y))
errors[0] = errors[0] + (1 - np.mean(llf_single_task == test_y))
errors = errors / reps
print("Errors For Angle {}: {}".format(angle, errors))
with open('rotation_results/angle_' + str(angle) + '_' + model + '.pickle',
'wb') as f:
pickle.dump(errors, f, protocol=2)
def LF_experiment(data_x, data_y, ntrees, shift, slot, model, num_points_per_task, acorn=None):
df = pd.DataFrame()
shifts = []
slots = []
accuracies_across_tasks = []
train_x_task0, train_y_task0, test_x_task0, test_y_task0 = cross_val_data(data_x, data_y, num_points_per_task, total_task=10, shift=shift, slot=slot)
lifelong_forest = LifeLongDNN(model = model, parallel = True if model == "uf" else False)
lifelong_forest.new_forest(
train_x_task0,
train_y_task0,
max_depth=ceil(log2(num_points_per_task)), n_estimators=ntrees
)
task_0_predictions=lifelong_forest.predict(
test_x_task0, representation='all', decider=0
)
shifts.append(shift)
slots.append(slot)
accuracies_across_tasks.append(np.mean(
task_0_predictions == test_y_task0
))
print(accuracies_across_tasks)
for task_ii in range(29):
train_x, train_y, _, _ = cross_val_data(data_x, data_y, num_points_per_task, total_task=10, shift=shift, slot=slot, task = task_ii)
print("Starting Task {} For Fold {} For Slot {}".format(task_ii, shift, slot))
lifelong_forest.new_forest(
train_x,
train_y,
max_depth=ceil(log2(num_points_per_task)), n_estimators=ntrees
)
task_0_predictions=lifelong_forest.predict(
test_x_task0, representation='all', decider=0
)
shifts.append(shift)
slots.append(slot)
accuracies_across_tasks.append(np.mean(
task_0_predictions == test_y_task0
))
print(accuracies_across_tasks)
df['data_fold'] = shifts
df['slot'] = slots
df['accuracy'] = accuracies_across_tasks
file_to_save = 'result/'+model+str(ntrees)+'_'+str(shift)+'_'+str(slot)+'.pickle'
with open(file_to_save, 'wb') as f:
pickle.dump(df, f)
def LF_experiment(train_x,
train_y,
test_x,
test_y,
ntrees,
shift,
slot,
model,
num_points_per_task,
acorn=None):
df = pd.DataFrame()
shifts = []
base_tasks = []
accuracies_across_tasks = []
lifelong_forest = LifeLongDNN(model=model,
parallel=True if model == "uf" else False)
for task_ii in range(10):
print("Starting Task {} For Fold {}".format(task_ii, shift))
if acorn is not None:
np.random.seed(acorn)
tmp = train_y[task_ii * 5000 +
slot * num_points_per_task:task_ii * 5000 +
(slot + 1) * num_points_per_task]
if task_ii != 0:
np.random.shuffle(tmp)
#tmp=np.random.randint(low = 0, high = 10, size = num_points_per_task)
lifelong_forest.new_forest(
train_x[task_ii * 5000 +
slot * num_points_per_task:task_ii * 5000 +
(slot + 1) * num_points_per_task, :],
tmp,
max_depth=ceil(log2(num_points_per_task)),
n_estimators=ntrees)
llf_task = lifelong_forest.predict(test_x[0:1000, :],
representation='all',
decider=0)
shifts.append(shift)
base_tasks.append(task_ii + 1)
accuracies_across_tasks.append(np.mean(llf_task == test_y[0:1000]))
df['data_fold'] = shifts
df['task'] = base_tasks
df['task_1_accuracy'] = accuracies_across_tasks
file_to_save = 'result/' + model + str(ntrees) + '_' + str(
shift) + '_' + str(slot) + '.pickle'
with open(file_to_save, 'wb') as f:
pickle.dump(df, f)
def LF_experiment(num_task_1_data, rep):
l2f = LifeLongDNN(model="uf", parallel=False)
X_train_task0, y_train_task0 = generate_gaussian_parity(n=num_task_1_data,
angle_params=0,
acorn=1)
X_train_task1, y_train_task1 = generate_gaussian_parity(n=100,
angle_params=10,
acorn=1)
X_test_task0, y_test_task0 = generate_gaussian_parity(n=10000,
angle_params=0,
acorn=2)
l2f.new_forest(X_train_task0, y_train_task0, n_estimators=10)
llf_task = l2f.predict(X_test_task0, representation=0, decider=0)
single_task_accuracy = np.nanmean(llf_task == y_test_task0)
single_task_error = 1 - single_task_accuracy
l2f.new_forest(X_train_task1, y_train_task1, n_estimators=10)
llf_task = l2f.predict(X_test_task0, representation="all", decider=0)
double_task_accuracy = np.nanmean(llf_task == y_test_task0)
double_task_error = 1 - double_task_accuracy
if double_task_error == 0 or single_task_error == 0:
te = 1
else:
te = (single_task_error + 1e-6) / (double_task_error + 1e-6)
df = pd.DataFrame()
df['te'] = [te]
file_to_save = 'result/' + str(num_task_1_data) + '_' + str(
rep) + '.pickle'
with open(file_to_save, 'wb') as f:
pickle.dump(df, f)
std_accuracy_dict = {'hybrid':[],'building':[],'recruiting':[],'UF':[]}
for ns in task_10_sample:
estimation_sample_no = ceil(estimation_set*ns)
validation_sample_no = ns - estimation_sample_no
for rep in range(reps):
print("doing {} samples for {} th rep".format(ns,rep))
## estimation
l2f = LifeLongDNN(model = "uf", parallel = True)
for task in range(9):
indx = np.random.choice(num_points_per_task, num_points_per_forest, replace=False)
l2f.new_forest(
train_x_across_task[task][indx],
train_y_across_task[task][indx],
max_depth=ceil(log2(num_points_per_forest)),
n_estimators=ntrees
)
task_10_train_indx = np.random.choice(num_points_per_task, ns, replace=False)
l2f.new_forest(
train_x_across_task[9][task_10_train_indx[:estimation_sample_no]],
train_y_across_task[9][task_10_train_indx[:estimation_sample_no]],
max_depth=ceil(log2(estimation_sample_no)),
n_estimators=ntrees
)
## L2F validation
posteriors_across_trees = estimate_posteriors(
def LF_experiment(train_x, train_y, test_x, test_y, ntrees, shift, slot, num_points_per_task, acorn=None):
uf_accuracies = np.zeros(10,dtype=float)
rf_accuracies = np.zeros(10,dtype=float)
single_task_accuracies = np.zeros(10,dtype=float)
l2f_accuracies = np.zeros(10,dtype=float)
for task_ii in range(10):
single_task_learner = LifeLongDNN(model = "uf", parallel = True)
if acorn is not None:
np.random.seed(acorn)
single_task_learner.new_forest(
train_x[task_ii*5000+slot*num_points_per_task:task_ii*5000+(slot+1)*num_points_per_task,:],
train_y[task_ii*5000+slot*num_points_per_task:task_ii*5000+(slot+1)*num_points_per_task],
max_depth=ceil(log2(num_points_per_task)), n_estimators=ntrees
)
llf_task=single_task_learner.predict(
test_x[task_ii*1000:(task_ii+1)*1000,:], representation=0, decider=0
)
single_task_accuracies[task_ii] = np.mean(
llf_task == test_y[task_ii*1000:(task_ii+1)*1000]
)
lifelong_forest = LifeLongDNN(model = "uf", parallel = True)
for task_ii in range(10):
print("Starting L2F Task {} For Fold {}".format(task_ii, shift))
if acorn is not None:
np.random.seed(acorn)
lifelong_forest.new_forest(
train_x[task_ii*5000+slot*num_points_per_task:task_ii*5000+(slot+1)*num_points_per_task,:],
train_y[task_ii*5000+slot*num_points_per_task:task_ii*5000+(slot+1)*num_points_per_task],
max_depth=ceil(log2(num_points_per_task)), n_estimators=ntrees
)
llf_task=lifelong_forest.predict(
test_x[task_ii*1000:(task_ii+1)*1000,:], representation='all', decider=task_ii
)
l2f_accuracies[task_ii] = np.mean(
llf_task == test_y[task_ii*1000:(task_ii+1)*1000]
)
for task_ii in range(10):
print("Starting UF Task {} For Fold {}".format(task_ii, shift))
lifelong_forest = LifeLongDNN(model = "uf", parallel = True)
if acorn is not None:
np.random.seed(acorn)
if task_ii== 0:
train_data_x = train_x[task_ii*5000+slot*num_points_per_task:task_ii*5000+(slot+1)*num_points_per_task,:]
train_data_y = train_y[task_ii*5000+slot*num_points_per_task:task_ii*5000+(slot+1)*num_points_per_task]
else:
train_data_x = np.concatenate(
(
train_data_x,
train_x[task_ii*5000+slot*num_points_per_task:task_ii*5000+(slot+1)*num_points_per_task,:]
),
axis = 0
)
train_data_y = np.concatenate(
(
train_data_y,
train_y[task_ii*5000+slot*num_points_per_task:task_ii*5000+(slot+1)*num_points_per_task]
),
axis = 0
)
lifelong_forest.new_forest(
train_data_x,
train_data_y,
max_depth=ceil(log2(num_points_per_task*(task_ii+1))), n_estimators=(task_ii+1)*ntrees
)
llf_task=lifelong_forest.predict(
test_x[task_ii*1000:(task_ii+1)*1000,:], representation=0, decider=0
)
uf_accuracies[task_ii] = np.mean(
llf_task == test_y[task_ii*1000:(task_ii+1)*1000]
)
for task_ii in range(10):
print("Starting RF Task {} For Fold {}".format(task_ii, shift))
RF = BaggingClassifier(
DecisionTreeClassifier(
max_depth=ceil(log2(num_points_per_task*(task_ii+1))),
min_samples_leaf=1,
max_features="auto"
),
n_estimators=(task_ii+1)*ntrees,
max_samples=0.63,
n_jobs = -1
)
if acorn is not None:
np.random.seed(acorn)
if task_ii== 0:
train_data_x = train_x[task_ii*5000+slot*num_points_per_task:task_ii*5000+(slot+1)*num_points_per_task,:]
train_data_y = train_y[task_ii*5000+slot*num_points_per_task:task_ii*5000+(slot+1)*num_points_per_task]
else:
train_data_x = np.concatenate(
(
train_data_x,
train_x[task_ii*5000+slot*num_points_per_task:task_ii*5000+(slot+1)*num_points_per_task,:]
),
axis = 0
)
train_data_y = np.concatenate(
(
train_data_y,
train_y[task_ii*5000+slot*num_points_per_task:task_ii*5000+(slot+1)*num_points_per_task]
),
axis = 0
)
RF.fit(
train_data_x,
train_data_y
)
llf_task=RF.predict(
test_x[task_ii*1000:(task_ii+1)*1000,:]
)
rf_accuracies[task_ii] = np.mean(
llf_task == test_y[task_ii*1000:(task_ii+1)*1000]
)
print(rf_accuracies[task_ii])
return single_task_accuracies,uf_accuracies,rf_accuracies,l2f_accuracies
def experiment(n_xor, n_rxor, n_test, reps, n_trees, max_depth, acorn=None):
#print(1)
if n_xor == 0 and n_rxor == 0:
raise ValueError('Wake up and provide samples to train!!!')
if acorn != None:
np.random.seed(acorn)
errors = np.zeros((reps, 4), dtype=float)
for i in range(reps):
l2f = LifeLongDNN()
uf = LifeLongDNN()
#source data
xor, label_xor = generate_gaussian_parity(n_xor,
cov_scale=0.1,
angle_params=0)
test_xor, test_label_xor = generate_gaussian_parity(n_test,
cov_scale=0.1,
angle_params=0)
#target data
nxor, label_nxor = generate_gaussian_parity(n_rxor,
cov_scale=0.1,
angle_params=np.pi / 4)
test_nxor, test_label_nxor = generate_gaussian_parity(
n_test, cov_scale=0.1, angle_params=np.pi / 4)
if n_xor == 0:
l2f.new_forest(nxor,
label_nxor,
n_estimators=n_trees,
max_depth=max_depth)
errors[i, 0] = 0.5
errors[i, 1] = 0.5
uf_task2 = l2f.predict(test_nxor, representation=0, decider=0)
l2f_task2 = l2f.predict(test_nxor, representation='all', decider=0)
errors[i, 2] = 1 - np.sum(uf_task2 == test_label_nxor) / n_test
errors[i, 3] = 1 - np.sum(l2f_task2 == test_label_nxor) / n_test
elif n_rxor == 0:
l2f.new_forest(xor,
label_xor,
n_estimators=n_trees,
max_depth=max_depth)
uf_task1 = l2f.predict(test_xor, representation=0, decider=0)
l2f_task1 = l2f.predict(test_xor, representation='all', decider=0)
errors[i, 0] = 1 - np.sum(uf_task1 == test_label_xor) / n_test
errors[i, 1] = 1 - np.sum(l2f_task1 == test_label_xor) / n_test
errors[i, 2] = 0.5
errors[i, 3] = 0.5
else:
l2f.new_forest(xor,
label_xor,
n_estimators=n_trees,
max_depth=max_depth)
l2f.new_forest(nxor,
label_nxor,
n_estimators=n_trees,
max_depth=max_depth)
uf.new_forest(xor,
label_xor,
n_estimators=2 * n_trees,
max_depth=max_depth)
uf.new_forest(nxor,
label_nxor,
n_estimators=2 * n_trees,
max_depth=max_depth)
uf_task1 = uf.predict(test_xor, representation=0, decider=0)
l2f_task1 = l2f.predict(test_xor, representation='all', decider=0)
uf_task2 = uf.predict(test_nxor, representation=1, decider=1)
l2f_task2 = l2f.predict(test_nxor, representation='all', decider=1)
errors[i, 0] = 1 - np.sum(uf_task1 == test_label_xor) / n_test
errors[i, 1] = 1 - np.sum(l2f_task1 == test_label_xor) / n_test
errors[i, 2] = 1 - np.sum(uf_task2 == test_label_nxor) / n_test
errors[i, 3] = 1 - np.sum(l2f_task2 == test_label_nxor) / n_test
return np.mean(errors, axis=0)
def experiment(n_spiral3,
n_spiral5,
n_test,
reps,
n_trees,
max_depth,
acorn=None):
#print(1)
if n_spiral3 == 0 and n_rxor == 0:
raise ValueError('Wake up and provide samples to train!!!')
if acorn != None:
np.random.seed(acorn)
errors = np.zeros((reps, 4), dtype=float)
for i in range(reps):
l2f = LifeLongDNN()
uf = LifeLongDNN()
#source data
spiral3, label_spiral3 = generate_spirals(n_spiral3, 2, 3, noise=2.5)
test_spiral3, test_label_spiral3 = generate_spirals(n_test,
2,
3,
noise=2.5)
#target data
spiral5, label_spiral5 = generate_spirals(n_spiral5, 2, 5, noise=2.5)
test_spiral5, test_label_spiral5 = generate_spirals(n_test,
2,
5,
noise=2.5)
if n_spiral3 == 0:
l2f.new_forest(spiral5,
label_spiral5,
n_estimators=n_trees,
max_depth=max_depth)
errors[i, 0] = 0.5
errors[i, 1] = 0.5
uf_task2 = l2f.predict(test_spiral5, representation=0, decider=0)
l2f_task2 = l2f.predict(test_spiral5,
representation='all',
decider=0)
errors[i, 2] = 1 - np.sum(uf_task2 == test_label_spiral5) / n_test
errors[i, 3] = 1 - np.sum(l2f_task2 == test_label_spiral5) / n_test
elif n_spiral5 == 0:
l2f.new_forest(spiral3,
label_spiral3,
n_estimators=n_trees,
max_depth=max_depth)
uf_task1 = l2f.predict(test_spiral3, representation=0, decider=0)
l2f_task1 = l2f.predict(test_spiral3,
representation='all',
decider=0)
errors[i, 0] = 1 - np.sum(uf_task1 == test_label_spiral3) / n_test
errors[i, 1] = 1 - np.sum(l2f_task1 == test_label_spiral3) / n_test
errors[i, 2] = 0.5
errors[i, 3] = 0.5
else:
l2f.new_forest(spiral3,
label_spiral3,
n_estimators=n_trees,
max_depth=max_depth)
l2f.new_forest(spiral5,
label_spiral5,
n_estimators=n_trees,
max_depth=max_depth)
uf.new_forest(spiral3,
label_spiral3,
n_estimators=2 * n_trees,
max_depth=max_depth)
uf.new_forest(spiral5,
label_spiral5,
n_estimators=2 * n_trees,
max_depth=max_depth)
uf_task1 = uf.predict(test_spiral3, representation=0, decider=0)
l2f_task1 = l2f.predict(test_spiral3,
representation='all',
decider=0)
uf_task2 = uf.predict(test_spiral5, representation=1, decider=1)
l2f_task2 = l2f.predict(test_spiral5,
representation='all',
decider=1)
errors[i, 0] = 1 - np.sum(uf_task1 == test_label_spiral3) / n_test
errors[i, 1] = 1 - np.sum(l2f_task1 == test_label_spiral3) / n_test
errors[i, 2] = 1 - np.sum(uf_task2 == test_label_spiral5) / n_test
errors[i, 3] = 1 - np.sum(l2f_task2 == test_label_spiral5) / n_test
return np.mean(errors, axis=0)
def LF_experiment(train_x,
train_y,
test_x,
test_y,
ntrees,
shift,
slot,
model,
num_points_per_task,
acorn=None):
df = pd.DataFrame()
single_task_accuracies = np.zeros(10, dtype=float)
shifts = []
tasks = []
base_tasks = []
accuracies_across_tasks = []
for task_ii in range(10):
if model == "uf":
single_task_learner = LifeLongDNN(model="uf", parallel=True)
if acorn is not None:
np.random.seed(acorn)
single_task_learner.new_forest(
train_x[task_ii * 5000 +
slot * num_points_per_task:task_ii * 5000 +
(slot + 1) * num_points_per_task, :],
train_y[task_ii * 5000 +
slot * num_points_per_task:task_ii * 5000 +
(slot + 1) * num_points_per_task],
max_depth=ceil(log2(num_points_per_task)),
n_estimators=ntrees * (task_ii + 1))
llf_task = single_task_learner.predict(
test_x[task_ii * 1000:(task_ii + 1) * 1000, :],
representation=0,
decider=0)
single_task_accuracies[task_ii] = np.mean(
llf_task == test_y[task_ii * 1000:(task_ii + 1) * 1000])
lifelong_forest = LifeLongDNN(model=model,
parallel=True if model == "uf" else False)
for task_ii in range(10):
print("Starting Task {} For Fold {}".format(task_ii, shift))
if acorn is not None:
np.random.seed(acorn)
lifelong_forest.new_forest(
train_x[task_ii * 5000 +
slot * num_points_per_task:task_ii * 5000 +
(slot + 1) * num_points_per_task, :],
train_y[task_ii * 5000 +
slot * num_points_per_task:task_ii * 5000 +
(slot + 1) * num_points_per_task],
max_depth=ceil(log2(num_points_per_task)),
n_estimators=ntrees)
if model == "dnn":
llf_task = lifelong_forest.predict(
test_x[task_ii * 1000:(task_ii + 1) * 1000, :],
representation=task_ii,
decider=task_ii)
single_task_accuracies[task_ii] = np.mean(
llf_task == test_y[task_ii * 1000:(task_ii + 1) * 1000])
for task_jj in range(task_ii + 1):
llf_task = lifelong_forest.predict(
test_x[task_jj * 1000:(task_jj + 1) * 1000, :],
representation='all',
decider=task_jj)
shifts.append(shift)
tasks.append(task_jj + 1)
base_tasks.append(task_ii + 1)
accuracies_across_tasks.append(
np.mean(llf_task == test_y[task_jj * 1000:(task_jj + 1) *
1000]))
df['data_fold'] = shifts
df['task'] = tasks
df['base_task'] = base_tasks
df['accuracy'] = accuracies_across_tasks
df_single_task = pd.DataFrame()
df_single_task['task'] = range(1, 11)
df_single_task['data_fold'] = shift
df_single_task['accuracy'] = single_task_accuracies
summary = (df, df_single_task)
file_to_save = 'result/result/' + model + str(
num_points_per_task) + '_' + str(ntrees) + '_' + str(
shift) + '_' + str(slot) + '.pickle'
with open(file_to_save, 'wb') as f:
pickle.dump(summary, f)