def accuracy(trials: ts.SparseRec,
labels: np.ndarray,
transform: str = "none",
repeats: int = 1000,
train_size: int = 30,
test_size: int = 20,
validate: bool = False,
**kwargs) -> np.ndarray:
"""Give precision estimation on the estimate from a simple SVC.
The score is calculated as tp / (tp + fp)
where tp is true positivie and fp is false positivie.
A 1000 time shuffle is made on the score population from different train test splits and
the mean score is output.
Args:
record_file: recording file
labels: ids of the trials belonging to different clusters
transform: how to we get the predictors
"none": platten neuron x sample_points and take PCs
"mean": temporal mean so each neuron has one value per trial, then take PCs
"corr": pairwise correlations between neurons and take PCs
repeats: number of repeats of resampling train/test sets
Returns:
the distribution of mean prediction scores
"""
X, y = trials.values, quantize(labels, groups=1)
trial_mask = X.min(axis=2).max(axis=0) > 0
X, y = X[:, trial_mask, :], y[trial_mask]
X = scale_features(X, (0, 2))
if transform == "none":
X = np.swapaxes(X, 0, 1).reshape(X.shape[1], X.shape[0] * X.shape[2])
elif transform == "corr": # get inter-neuron correlation for each of the trials
X = np.array([
np.corrcoef(x)[np.triu_indices(x.shape[0], 1)]
for x in np.rollaxis(X, 1)
])
elif transform == "mean":
X = np.swapaxes(X.mean(-1), 0, 1)
else:
raise ValueError(
"[precision] <transform> must be one of 'none', 'corr', or 'mean'."
)
X = PCA(PC_NO).fit_transform(X) if X.shape[1] > PC_NO else X
params = {"kernel": "linear", "gamma": "auto"}
params.update(kwargs)
svc = SVC(**params)
splitter = split_data(X, y, repeats, train_size, test_size)
if validate:
result = [
accuracy_score(y_te,
svc.fit(X_tr, y_tr).predict(X_te))
for X_tr, y_tr, X_te, y_te in splitter
]
else:
result = [
accuracy_score(y_tr,
svc.fit(X_tr, y_tr).predict(X_tr))
for X_tr, y_tr, _, _ in splitter
]
return [x for x in result if (x is not None and x > 0.0)]
def get_classifier(self, traindata, kf):
x_tr, x_te, y_tr, y_te = fac.to_kfold(traindata, kf)
acc_max, bestK, acc = 0, 0, [[] for a in range(kf)]
for i in range(kf):
# print('DOAO round', i, 'begin')
# svm 00
print('test00')
clf_svm = SVC()
clf_svm.fit(x_tr[i], y_tr[i].ravel())
label_svm = clf_svm.predict(x_te[i])
acc[i].append(fac.get_acc(y_te[i], label_svm)[0])
# KNN 01
print('test01')
acc_k = []
aux_k = [3, 5, 7]
# for k in range(3, 12, 2):
for k in aux_k:
clf_knn = KNN_GPU(k=k)
clf_knn.fit(x_tr[i], y_tr[i])
label_knn = clf_knn.predict(x_te[i])
acc_k.append(fac.get_acc(y_te[i], label_knn)[0])
acc[i].append(max(acc_k))
bestK = aux_k[acc_k.index(max(acc_k))]
# LR 02
print('test02')
clf_lr = LogisticRegression()
clf_lr.fit(x_tr[i], y_tr[i])
label_LR = clf_lr.predict(x_te[i])
acc[i].append(fac.get_acc(y_te[i], label_LR)[0])
# XgBoost 03
print('test03')
clf_xgb = DecisionTreeClassifier()
clf_xgb.fit(x_tr[i], y_tr[i])
label_xgb = clf_xgb.predict(x_te[i])
acc[i].append(fac.get_acc(y_te[i], label_xgb)[0])
# RF 04
print('test04')
clf_rf = TGBMClassifier()
clf_rf.fit(x_tr[i], y_tr[i])
label_rf = clf_rf.predict(x_te[i])
acc[i].append(fac.get_acc(y_te[i], label_rf)[0])
print('DOAO round', i, 'end')
acc = np.array(acc)
acc_mean = acc.mean(axis=0)
# fun_best = np.where(acc_mean == max(acc_mean))
fun_best = np.argmax(acc_mean)
return fun_best, bestK
def get_classifier(self, train, kf):
x_tr, x_te, y_tr, y_te = fac.to_kfold(train, kf)
acc_max, bestK, acc = 0, 0, [[] for a in range(kf)]
for i in range(kf):
# print('DECOC round', i, 'begin')
# svm 00
clf_svm = SVC()
clf_svm.fit(x_tr[i], y_tr[i].ravel())
label_svm = clf_svm.predict(x_te[i])
acc[i].append(fac.get_acc(y_te[i], label_svm)[0])
# KNN 01
acc_k = []
aux_k = [3, 5, 7]
# for k in range(3, 12, 2):
for k in aux_k:
clf_knn = KNN_GPU(k=k)
clf_knn.fit(x_tr[i], y_tr[i])
label_knn = clf_knn.predict(x_te[i])
acc_k.append(fac.get_acc(y_te[i], label_knn)[0])
acc[i].append(max(acc_k))
bestK = aux_k[acc_k.index(max(acc_k))]
# # LR 02
# clf_lr = LR_GPU()
# clf_lr.fit(x_tr[i], y_tr[i])
# label_LR = clf_lr.predicted(x_te[i])
# acc[i].append(fac.get_acc(y_te[i], label_LR)[0])
# LR 02
clf_lr = LogisticRegression()
clf_lr.fit(x_tr[i], y_tr[i])
label_LR = clf_lr.predict(x_te[i])
acc[i].append(fac.get_acc(y_te[i], label_LR)[0])
# CART 03
clf_cart = DecisionTreeClassifier()
clf_cart.fit(x_tr[i], y_tr[i])
label_cart = clf_cart.predict(x_te[i])
acc[i].append(fac.get_acc(y_te[i], label_cart)[0])
# # RF 04
clf_rf = TGBMClassifier()
clf_rf.fit(x_tr[i], y_tr[i].ravel())
label_rf = clf_rf.predict(x_te[i])
acc[i].append(fac.get_acc(y_te[i], label_rf)[0])
print('DECOC round', i, 'end')
acc = np.array(acc)
acc_mean = acc.mean(axis=0)
# fun_best = np.where(acc_mean == max(acc_mean))
fun_best = np.argmax(acc_mean)
return fun_best, bestK
def fun_predict(self, x_te, C, D, L):
print('func_predict')
num = len(D)
cf = C[0]
ck = C[1]
allpre = np.zeros((len(x_te), num))
for i in range(num):
train = D[i]
traindata = train[:, 0:-1]
trainlabel = train[:, -1]
if cf[i] == 0:
# svm
print('SVM predict')
clf_svm = SVC()
clf_svm.fit(traindata, trainlabel.ravel())
label_svm = clf_svm.predict(x_te)
allpre[:, i] = label_svm
elif cf[i] == 1:
# knn
clf_knn = KNN_GPU(k=ck[i])
clf_knn.fit(traindata, trainlabel)
label_knn = clf_knn.predict(x_te)
allpre[:, i] = label_knn
elif cf[i] == 2:
# LR
print('LR predict')
clf_lr = LogisticRegression()
clf_lr.fit(traindata, trainlabel.ravel())
label_LR = clf_lr.predict(x_te)
allpre[:, i] = label_LR
elif cf[i] == 3:
# CART
print('CART predict')
clf_xgb = DecisionTreeClassifier()
clf_xgb.fit(traindata, trainlabel)
label_xgb = clf_xgb.predict(x_te)
allpre[:, i] = label_xgb
elif cf[i] == 4:
# Rf
print('RF predict')
clf_rf = TGBMClassifier()
clf_rf.fit(traindata, trainlabel.ravel())
label_rf = clf_rf.predict(x_te)
allpre[:, i] = label_rf
else:
print('error !!!! DOAO.fun_predict')
label = L[i]
for j in range(len(x_te)):
allpre[j, i] = label[0] if allpre[j, i] == 0 else label[1]
# print('predict end for')
pre = mode(allpre, axis=1)[0]
return pre
def funcPreEDOVO(self, x_test, y_test, C, D):
numC = np.asarray(C).shape[0]
num_set = len(y_test)
allpre = np.zeros([num_set, numC])
for i in range(numC):
train = D[i]
traindata = np.array(train[:, 0:-1])
trainlabel = np.array(train[:, -1], dtype='int64')
if C[i, 0] == 0:
print('test0')
# svm
clf_svm = SVC()
clf_svm.fit(traindata, trainlabel.ravel())
label_svm = clf_svm.predict(x_test)
allpre[:, i] = label_svm
elif C[i, 0] == 1:
# print('test1')
# knn
clf_knn = KNN_GPU(k=C[i][1])
# clf_knn = KNN_torch(k=C[i][1])
clf_knn.fit(traindata, trainlabel)
label_knn = clf_knn.predict(x_test)
allpre[:, i] = label_knn.ravel()
elif C[i, 0] == 2:
print('test2')
# LR
clf_lr = LogisticRegression()
clf_lr.fit(traindata, trainlabel)
label_LR = clf_lr.predict(x_test)
allpre[:, i] = label_LR
# # LR
# clf_lr = LR_GPU()
# clf_lr.fit(traindata, trainlabel)
# label_LR = clf_lr.predicted(x_test)
# allpre[:, i] = label_LR
elif C[i, 0] == 3:
print('test3')
# CART
clf_cart = DecisionTreeClassifier()
clf_cart.fit(traindata, trainlabel)
label_cart = clf_cart.predict(x_test)
allpre[:, i] = label_cart
elif C[i, 0] == 4:
print('test4')
# RandomForest
clf_ada = TGBMClassifier()
clf_ada.fit(traindata, trainlabel.ravel())
label_ada = clf_ada.predict(x_test)
allpre[:, i] = label_ada
else:
print('error !!!! DECOC.funcPreEDOVO')
return allpre
def train_kSVM(X_train, X_test, y_train, y_test, split_ID):
kSVM = SVC(kernel='rbf',
degree=3,
gamma='auto',
coef0=0.0,
C=1.0,
tol=0.001,
probability=False,
class_weight='balanced',
shrinking=False,
cache_size=None,
verbose=False,
max_iter=-1,
n_jobs=-1,
max_mem_size=-1,
random_state=None,
decision_function_shape='ovo')
kSVM_model = kSVM.fit(X_train, y_train)
kSVM_preds = kSVM_model.predict(X_test)
prec, rec, f_1, supp = prf(y_test, kSVM_preds, average=None)
class_rep = sklearn.metrics.classification_report(y_test, kSVM_preds)
exp.log_other('Classification Report' + split_ID, class_rep)
mcc = sklearn.metrics.matthews_corrcoef(y_test, kSVM_preds)
#if first iteration, report model parameters to comet
if split_ID == '0':
exp.log_parameters(kSVM_preds.get_params())
return prec, rec, f_1, supp, mcc
def funClassifier(self, x_train, y_train, type):
labels = np.unique(y_train)
nc = len(labels)
code = self.funECOCim(nc, type)
train, numn = [], []
for i in range(nc):
idi = np.where(y_train == labels[i])[0]
train.append(x_train[idi])
numn.append(len(idi))
num1 = np.size(code, 1)
ft = []
for t in range(num1):
Dt, DtLabel, flagDt, numAp, numAn, numNp, numNn = np.asarray(
[]), [], 0, 0, 0, 0, 0
for i in range(nc):
if code[i, t] == 1:
if Dt.shape[0] == 0:
Dt = train[i]
else:
Dt = np.append(Dt, train[i], axis=0)
DtLabel[flagDt:flagDt + numn[i]] = [1] * numn[i]
flagDt += numn[i]
numAp += 1
numNp += numn[i]
elif code[i, t] == -1:
if Dt.shape[0] == 0:
Dt = train[i]
else:
Dt = np.append(Dt, train[i], axis=0)
DtLabel[flagDt:flagDt + numn[i]] = [0] * numn[i]
flagDt += numn[i]
numAn += 1
numNn += numn[i]
clf_svc = SVC()
clf_svc.fit(np.array(Dt), np.array(DtLabel).ravel())
ft.append(clf_svc)
return code, ft, labels
y,
test_size=0.1,
random_state=0)
with open('settings/olivetti.json') as f:
data = json.load(f)
for s in data['models']:
print(s)
m = SVC(kernel=s.get('kernel', 'rbf'),
C=s.get('C', 10.0),
coef0=s.get('coef0', 0.0),
gamma=s.get('gamma', 'auto'),
degree=int(s.get('degree', 3)),
verbose=False)
start = time.time()
m.fit(X_train, y_train)
end = time.time()
print('fit time: ', end - start)
t = m.predict(X_train)
print('training error: ', 1 - accuracy_score(y_train, t))
start = time.time()
p = m.predict(X_test)
end = time.time()
print('prediction time: ', end - start)
print('accuracy: ', accuracy_score(y_test, p))
mesh2.to_vtp(os.path.join(output_path, '{}_d_predicted_refined.vtp'.format(i_sample[:-4])))
# upsampling
print('\tUpsampling...')
if mesh.cells.shape[0] > 100000:
target_num = 100000 # set max number of cells
ratio = 1 - target_num/mesh.cells.shape[0] # calculate ratio
mesh.mesh_decimation(ratio)
print('Original contains too many cells, simpify to {} cells'.format(mesh.cells.shape[0]))
fine_cells = mesh.cells
if upsampling_method == 'SVM':
clf = SVC(kernel='rbf', gamma='auto', gpu_id=gpu_id)
# train SVM
clf.fit(cells, np.ravel(refine_labels))
fine_labels = clf.predict(fine_cells)
fine_labels = fine_labels.reshape([mesh.cells.shape[0], 1])
elif upsampling_method == 'KNN':
neigh = KNeighborsClassifier(n_neighbors=3)
# train KNN
neigh.fit(cells, np.ravel(refine_labels))
fine_labels = neigh.predict(fine_cells)
fine_labels = fine_labels.reshape([mesh.cells.shape[0], 1])
mesh2 = Easy_Mesh()
mesh2.cells = mesh.cells
mesh2.update_cell_ids_and_points()
mesh2.cell_attributes['Label'] = fine_labels
mesh2.to_vtp(os.path.join(output_path, '{}_predicted_refined.vtp'.format(i_sample[:-4])))
print("acc %s" % score)
"""
if __name__ == "__main__":
with open("./data/comment_new/vedio_vector_svm", "rb") as f:
vedio_vector = pickle.load(f)
with open("./train", "rb") as f:
train_D = pickle.load(f)
with open("./test", "rb") as f:
test_D = pickle.load(f)
train_X = []
train_Y = []
for f, L1 in train_D:
train_X.append(vedio_vector[f])
train_Y.append(m[L1])
test_X = []
test_Y = []
for f, L1 in test_D:
test_X.append(vedio_vector[f])
test_Y.append(m[L1])
train_X = np.array(train_X)
train_Y = np.array(train_Y)
test_X = np.array(test_X)
test_Y = np.array(test_Y)
clf = SVC()
print(train_X.shape)
print(train_Y.shape)
clf.fit(train_X, train_Y)
print("acc %s" % clf.score(test_X, test_Y))
class ModelBasedOption(object):
def __init__(self,
*,
name,
parent,
mdp,
global_solver,
global_value_learner,
buffer_length,
global_init,
gestation_period,
timeout,
max_steps,
device,
use_vf,
use_global_vf,
use_model,
dense_reward,
option_idx,
lr_c,
lr_a,
max_num_children=2,
target_salient_event=None,
path_to_model="",
multithread_mpc=False):
self.mdp = mdp
self.name = name
self.lr_c = lr_c
self.lr_a = lr_a
self.parent = parent
self.device = device
self.use_vf = use_vf
self.global_solver = global_solver
self.use_global_vf = use_global_vf
self.timeout = timeout
self.use_model = use_model
self.max_steps = max_steps
self.global_init = global_init
self.dense_reward = dense_reward
self.buffer_length = buffer_length
self.max_num_children = max_num_children
self.target_salient_event = target_salient_event
self.multithread_mpc = multithread_mpc
# TODO
self.overall_mdp = mdp
self.seed = 0
self.option_idx = option_idx
self.num_goal_hits = 0
self.num_executions = 0
self.gestation_period = gestation_period
self.positive_examples = []
self.negative_examples = []
self.optimistic_classifier = None
self.pessimistic_classifier = None
# In the model-free setting, the output norm doesn't seem to work
# But it seems to stabilize off policy value function learning
# Therefore, only use output norm if we are using MPC for action selection
use_output_norm = self.use_model
if not self.use_global_vf or global_init:
self.value_learner = TD3(state_dim=self.mdp.state_space_size() + 2,
action_dim=self.mdp.action_space_size(),
max_action=1.,
name=f"{name}-td3-agent",
device=self.device,
lr_c=lr_c,
lr_a=lr_a,
use_output_normalization=use_output_norm)
self.global_value_learner = global_value_learner if not self.global_init else None # type: TD3
if use_model:
print(f"Using model-based controller for {name}")
self.solver = self._get_model_based_solver()
else:
print(f"Using model-free controller for {name}")
self.solver = self._get_model_free_solver()
self.children = []
self.success_curve = []
self.effect_set = []
if path_to_model:
print(f"Loading model from {path_to_model} for {self.name}")
self.solver.load_model(path_to_model)
if self.use_vf and not self.use_global_vf and self.parent is not None:
self.initialize_value_function_with_global_value_function()
print(
f"Created model-based option {self.name} with option_idx={self.option_idx}"
)
self.is_last_option = False
def _get_model_based_solver(self):
assert self.use_model
if self.global_init:
return MPC(mdp=self.mdp,
state_size=self.mdp.state_space_size(),
action_size=self.mdp.action_space_size(),
dense_reward=self.dense_reward,
device=self.device,
multithread=self.multithread_mpc)
assert self.global_solver is not None
return self.global_solver
def _get_model_free_solver(self):
assert not self.use_model
assert self.use_vf
# Global option creates its own VF solver
if self.global_init:
assert self.value_learner is not None
return self.value_learner
# Local option either uses the global VF..
if self.use_global_vf:
assert self.global_value_learner is not None
return self.global_value_learner
# .. or uses its own local VF as solver
assert self.value_learner is not None
return self.value_learner
# ------------------------------------------------------------
# Learning Phase Methods
# ------------------------------------------------------------
def get_training_phase(self):
if self.num_goal_hits < self.gestation_period:
return "gestation"
return "initiation_done"
def extract_features_for_initiation_classifier(self, state):
features = state if isinstance(state, np.ndarray) else state.features()
if "push" in self.mdp.env_name:
return features[:4]
return features[:2]
def is_init_true(self, state):
if self.global_init or self.get_training_phase() == "gestation":
return True
if self.is_last_option and self.mdp.get_start_state_salient_event()(
state):
return True
features = self.extract_features_for_initiation_classifier(state)
return self.optimistic_classifier.predict(
[features])[0] == 1 or self.pessimistic_is_init_true(state)
def is_term_true(self, state):
if self.parent is None:
return self.target_salient_event(state)
# TODO change
return self.parent.pessimistic_is_init_true(state)
def pessimistic_is_init_true(self, state):
if self.global_init or self.get_training_phase() == "gestation":
return True
features = self.extract_features_for_initiation_classifier(state)
return self.pessimistic_classifier.predict([features])[0] == 1
def is_at_local_goal(self, state, goal):
""" Goal-conditioned termination condition. """
reached_goal = self.mdp.sparse_gc_reward_function(state, goal, {})[1]
reached_term = self.is_term_true(state) or state.is_terminal()
return reached_goal and reached_term
# ------------------------------------------------------------
# Control Loop Methods
# ------------------------------------------------------------
def _get_epsilon(self):
if self.use_model:
return 0.1
if not self.dense_reward and self.num_goal_hits <= 3:
return 0.8
return 0.2
def act(self, state, goal):
""" Epsilon-greedy action selection. """
if random.random() < self._get_epsilon():
return self.mdp.sample_random_action()
if self.use_model:
assert isinstance(self.solver, MPC), f"{type(self.solver)}"
vf = self.value_function if self.use_vf else None
return self.solver.act(state, goal, vf=vf)
assert isinstance(self.solver, TD3), f"{type(self.solver)}"
augmented_state = self.get_augmented_state(state, goal)
return self.solver.act(augmented_state, evaluation_mode=False)
def update_model(self, state, action, reward, next_state):
""" Learning update for option model/actor/critic. """
self.solver.step(state.features(), action, reward,
next_state.features(), next_state.is_terminal())
def get_goal_for_rollout(self):
""" Sample goal to pursue for option rollout. """
if self.parent is None and self.target_salient_event is not None:
return self.target_salient_event.get_target_position()
sampled_goal = self.parent.sample_from_initiation_region_fast_and_epsilon(
)
assert sampled_goal is not None
if isinstance(sampled_goal, np.ndarray):
return sampled_goal.squeeze()
return self.extract_goal_dimensions(sampled_goal)
def rollout(self, step_number, rollout_goal=None, eval_mode=False):
""" Main option control loop. """
start_state = deepcopy(self.mdp.cur_state)
assert self.is_init_true(start_state)
num_steps = 0
total_reward = 0
visited_states = []
option_transitions = []
state = deepcopy(self.mdp.cur_state)
goal = self.get_goal_for_rollout(
) if rollout_goal is None else rollout_goal
print(
f"[Step: {step_number}] Rolling out {self.name}, from {state.position} targeting {goal}"
)
self.num_executions += 1
while not self.is_at_local_goal(
state, goal
) and step_number < self.max_steps and num_steps < self.timeout:
# Control
action = self.act(state, goal)
reward, next_state = self.mdp.execute_agent_action(action)
if self.use_model:
self.update_model(state, action, reward, next_state)
# Logging
num_steps += 1
step_number += 1
total_reward += reward
visited_states.append(state)
option_transitions.append((state, action, reward, next_state))
state = deepcopy(self.mdp.cur_state)
visited_states.append(state)
self.success_curve.append(self.is_term_true(state))
self.effect_set.append(state.features())
if self.is_term_true(state):
self.num_goal_hits += 1
if self.use_vf and not eval_mode:
self.update_value_function(
option_transitions,
pursued_goal=goal,
reached_goal=self.extract_goal_dimensions(state))
self.derive_positive_and_negative_examples(visited_states)
# Always be refining your initiation classifier
if not self.global_init and not eval_mode:
self.fit_initiation_classifier()
return option_transitions, total_reward
# ------------------------------------------------------------
# Hindsight Experience Replay
# ------------------------------------------------------------
def update_value_function(self, option_transitions, reached_goal,
pursued_goal):
""" Update the goal-conditioned option value function. """
self.experience_replay(option_transitions, pursued_goal)
self.experience_replay(option_transitions, reached_goal)
def initialize_value_function_with_global_value_function(self):
self.value_learner.actor.load_state_dict(
self.global_value_learner.actor.state_dict())
self.value_learner.critic.load_state_dict(
self.global_value_learner.critic.state_dict())
self.value_learner.target_actor.load_state_dict(
self.global_value_learner.target_actor.state_dict())
self.value_learner.target_critic.load_state_dict(
self.global_value_learner.target_critic.state_dict())
def extract_goal_dimensions(self, goal):
goal_features = goal if isinstance(goal,
np.ndarray) else goal.features()
if "ant" in self.mdp.env_name:
return goal_features[:2]
raise NotImplementedError(f"{self.mdp.env_name}")
def get_augmented_state(self, state, goal):
assert goal is not None and isinstance(goal, np.ndarray)
goal_position = self.extract_goal_dimensions(goal)
return np.concatenate((state.features(), goal_position))
def experience_replay(self, trajectory, goal_state):
for state, action, reward, next_state in trajectory:
augmented_state = self.get_augmented_state(state, goal=goal_state)
augmented_next_state = self.get_augmented_state(next_state,
goal=goal_state)
done = self.is_at_local_goal(next_state, goal_state)
reward_func = self.overall_mdp.dense_gc_reward_function if self.dense_reward \
else self.overall_mdp.sparse_gc_reward_function
reward, global_done = reward_func(next_state, goal_state, info={})
if not self.use_global_vf or self.global_init:
self.value_learner.step(augmented_state, action, reward,
augmented_next_state, done)
# Off-policy updates to the global option value function
if not self.global_init:
assert self.global_value_learner is not None
self.global_value_learner.step(augmented_state, action, reward,
augmented_next_state,
global_done)
def value_function(self, states, goals):
assert isinstance(states, np.ndarray)
assert isinstance(goals, np.ndarray)
if len(states.shape) == 1:
states = states[None, ...]
if len(goals.shape) == 1:
goals = goals[None, ...]
goal_positions = goals[:, :2]
augmented_states = np.concatenate((states, goal_positions), axis=1)
augmented_states = torch.as_tensor(augmented_states).float().to(
self.device)
if self.use_global_vf and not self.global_init:
values = self.global_value_learner.get_values(augmented_states)
else:
values = self.value_learner.get_values(augmented_states)
return values
# ------------------------------------------------------------
# Learning Initiation Classifiers
# ------------------------------------------------------------
def get_first_state_in_classifier(self,
trajectory,
classifier_type="pessimistic"):
""" Extract the first state in the trajectory that is inside the initiation classifier. """
assert classifier_type in ("pessimistic",
"optimistic"), classifier_type
classifier = self.pessimistic_is_init_true if classifier_type == "pessimistic" else self.is_init_true
for state in trajectory:
if classifier(state):
return state
return None
def sample_from_initiation_region_fast(self):
""" Sample from the pessimistic initiation classifier. """
num_tries = 0
sampled_state = None
while sampled_state is None and num_tries < 200:
num_tries = num_tries + 1
sampled_trajectory_idx = random.choice(
range(len(self.positive_examples)))
sampled_trajectory = self.positive_examples[sampled_trajectory_idx]
sampled_state = self.get_first_state_in_classifier(
sampled_trajectory)
return sampled_state
def sample_from_initiation_region_fast_and_epsilon(self):
""" Sample from the pessimistic initiation classifier. """
def compile_states(s):
pos0 = self.mdp.get_position(s)
pos1 = np.copy(pos0)
pos1[0] -= self.target_salient_event.tolerance
pos2 = np.copy(pos0)
pos2[0] += self.target_salient_event.tolerance
pos3 = np.copy(pos0)
pos3[1] -= self.target_salient_event.tolerance
pos4 = np.copy(pos0)
pos4[1] += self.target_salient_event.tolerance
return pos0, pos1, pos2, pos3, pos4
idxs = [i for i in range(len(self.positive_examples))]
random.shuffle(idxs)
for idx in idxs:
sampled_trajectory = self.positive_examples[idx]
states = []
for s in sampled_trajectory:
states.extend(compile_states(s))
position_matrix = np.vstack(states)
# optimistic_predictions = self.optimistic_classifier.predict(position_matrix) == 1
# pessimistic_predictions = self.pessimistic_classifier.predict(position_matrix) == 1
# predictions = np.logical_or(optimistic_predictions, pessimistic_predictions)
predictions = self.pessimistic_classifier.predict(
position_matrix) == 1
predictions = np.reshape(predictions, (-1, 5))
valid = np.all(predictions, axis=1)
indices = np.argwhere(valid == True)
if len(indices) > 0:
return sampled_trajectory[indices[0][0]]
return self.sample_from_initiation_region_fast()
def derive_positive_and_negative_examples(self, visited_states):
start_state = visited_states[0]
final_state = visited_states[-1]
if self.is_term_true(final_state):
positive_states = [start_state
] + visited_states[-self.buffer_length:]
self.positive_examples.append(positive_states)
else:
negative_examples = [start_state]
self.negative_examples.append(negative_examples)
def should_change_negative_examples(self):
should_change = []
for negative_example in self.negative_examples:
should_change += [
self.does_model_rollout_reach_goal(negative_example[0])
]
return should_change
def does_model_rollout_reach_goal(self, state):
sampled_goal = self.get_goal_for_rollout()
final_states, actions, costs = self.solver.simulate(
state, sampled_goal, num_rollouts=14000, num_steps=self.timeout)
farthest_position = final_states[:, :2].max(axis=0)
return self.is_term_true(farthest_position)
def fit_initiation_classifier(self):
if len(self.negative_examples) > 0 and len(self.positive_examples) > 0:
self.train_two_class_classifier()
elif len(self.positive_examples) > 0:
self.train_one_class_svm()
def construct_feature_matrix(self, examples):
states = list(itertools.chain.from_iterable(examples))
positions = [
self.extract_features_for_initiation_classifier(state)
for state in states
]
return np.array(positions)
def train_one_class_svm(self,
nu=0.1
): # TODO: Implement gamma="auto" for thundersvm
positive_feature_matrix = self.construct_feature_matrix(
self.positive_examples)
self.pessimistic_classifier = OneClassSVM(kernel="rbf", nu=nu)
self.pessimistic_classifier.fit(positive_feature_matrix)
self.optimistic_classifier = OneClassSVM(kernel="rbf", nu=nu / 10.)
self.optimistic_classifier.fit(positive_feature_matrix)
def train_two_class_classifier(self, nu=0.1):
positive_feature_matrix = self.construct_feature_matrix(
self.positive_examples)
negative_feature_matrix = self.construct_feature_matrix(
self.negative_examples)
positive_labels = [1] * positive_feature_matrix.shape[0]
negative_labels = [0] * negative_feature_matrix.shape[0]
X = np.concatenate((positive_feature_matrix, negative_feature_matrix))
Y = np.concatenate((positive_labels, negative_labels))
if negative_feature_matrix.shape[
0] >= 10: # TODO: Implement gamma="auto" for thundersvm
kwargs = {
"kernel": "rbf",
"gamma": "auto",
"class_weight": "balanced"
}
else:
kwargs = {"kernel": "rbf", "gamma": "auto"}
self.optimistic_classifier = SVC(**kwargs)
self.optimistic_classifier.fit(X, Y)
training_predictions = self.optimistic_classifier.predict(X)
positive_training_examples = X[training_predictions == 1]
if positive_training_examples.shape[0] > 0:
self.pessimistic_classifier = OneClassSVM(kernel="rbf", nu=nu)
self.pessimistic_classifier.fit(positive_training_examples)
# ------------------------------------------------------------
# Distance functions
# ------------------------------------------------------------
def get_states_inside_pessimistic_classifier_region(self):
point_array = self.construct_feature_matrix(self.positive_examples)
point_array_predictions = self.pessimistic_classifier.predict(
point_array)
positive_point_array = point_array[point_array_predictions == 1]
return positive_point_array
def distance_to_state(self, state, metric="euclidean"):
""" Compute the distance between the current option and the input `state`. """
assert metric in ("euclidean", "value"), metric
if metric == "euclidean":
return self._euclidean_distance_to_state(state)
return self._value_distance_to_state(state)
def _euclidean_distance_to_state(self, state):
point = self.mdp.get_position(state)
assert isinstance(point, np.ndarray)
assert point.shape == (2, ), point.shape
positive_point_array = self.get_states_inside_pessimistic_classifier_region(
)
distances = distance.cdist(point[None, :], positive_point_array)
return np.median(distances)
def _value_distance_to_state(self, state):
features = state.features() if not isinstance(state,
np.ndarray) else state
goals = self.get_states_inside_pessimistic_classifier_region()
distances = self.value_function(features, goals)
distances[distances > 0] = 0.
return np.median(np.abs(distances))
# ------------------------------------------------------------
# Convenience functions
# ------------------------------------------------------------
def get_option_success_rate(self):
if self.num_executions > 0:
return self.num_goal_hits / self.num_executions
return 1.
def get_success_rate(self):
if len(self.success_curve) == 0:
return 0.
return np.mean(self.success_curve)
def __str__(self):
return self.name
def __repr__(self):
return str(self)
def __eq__(self, other):
if isinstance(other, ModelBasedOption):
return self.name == other.name
return False
def run(self):
model = self.exp_data['model']
base_workspace = self.exp_data['base_workspace']
if model == 'ridge':
space = [Real(1e-3, 1e+3, prior='log-uniform')]
optimize_types = ['alpha']
minimizer = MagnitudeAxisMinimizer(base_workspace, optimize_types,
gp_minimize)
minimizer.model = Ridge
x0 = [1.0]
elif model == 'kernel_ridge':
space = [
Categorical(['poly', 'rbf', 'sigmoid']),
Real(1e-3, 1e+3, prior='log-uniform'),
Integer(1, 8),
Real(1e-6, 1e+1, prior='log-uniform'),
Real(-10, 10)
]
optimize_types = ['kernel', 'alpha', 'degree', 'gamma', 'coef0']
minimizer = MagnitudeAxisMinimizer(base_workspace, optimize_types,
gp_minimize)
minimizer.model = KernelRidge
x0 = ['poly', 1.0, 3, 1 / 300, 0]
elif model == 'kernel_ridge_separation':
space = [Real(1e-3, 1e+3, prior='log-uniform')]
optimize_types = ['alpha']
minimizer = MagnitudeAxisMinimizer(base_workspace, optimize_types,
gp_minimize)
minimizer.model_fixed_params = {
'kernel': 'poly',
'degree': 3,
'gamma': 1 / 300,
'coef0': 0
}
minimizer.model = KernelRidge
x0 = [1.0]
elif model == 'kernel_ridge_random':
space = [Real(1e-3, 1e+3, prior='log-uniform')]
optimize_types = ['alpha']
minimizer = MagnitudeAxisMinimizer(base_workspace, optimize_types,
gp_minimize)
kernel = random.choice(
['poly', 'rbf', 'laplacian', 'sigmoid', 'cosine'])
degree = random.randint(1, 10)
coef = random.uniform(-5, 5)
minimizer.model_fixed_params = {
'kernel': kernel,
'degree': degree,
'gamma': None,
'coef0': coef
}
minimizer.model = KernelRidge
x0 = [1.0]
elif model == 'nn':
space = [
Integer(16, 256),
Real(1e-5, 1, prior='log-uniform'),
]
optimize_types = ['n_hidden_units', 'lr']
minimizer = MagnitudeAxisMinimizer(base_workspace, optimize_types,
gp_minimize)
minimizer.model_fixed_params = {
'build_fn': self.build_nn,
'epochs': 20,
'batch_size': 256
}
minimizer.model = KerasRegressor
x0 = [64, 0.001]
elif model == 'kernel_svm':
space = [
Categorical(['poly', 'rbf', 'sigmoid']),
Real(1e-3, 1e+3, prior='log-uniform'),
Integer(1, 8),
Real(1e-6, 1e+1, prior='log-uniform'),
Real(-10, 10)
]
optimize_types = ['kernel', 'C', 'degree', 'gamma', 'coef0']
minimizer = MagnitudeAxisMinimizer(base_workspace, optimize_types,
gp_minimize)
minimizer.model_fixed_params = {
'cache_size': 8000,
'max_iter': 10000
}
minimizer.model = SVR
x0 = ['poly', 1.0, 3, 1 / 300, 0]
elif model == 'pca':
if not hasattr(self, 'pca_component'):
pca = PCA(n_components=1)
_, X_nums, *_ = prepare_separation_data(
self.name.split('_')[0] + '.txt')
pca.fit(X_nums)
# pca.fit(np.concatenate([base_workspace['X'],base_workspace['X_test']]))
self.pca_component = pca.components_ # 1xd
error = self.evaluate_w(base_workspace['X_test'],
self.pca_component,
base_workspace['y_test'])
return error
elif model == 'proj_pca':
if not hasattr(self, 'proj_pca_component'):
X, X_nums, y_label, _ = prepare_separation_data(
self.name.split('_')[0] + '.txt')
svc = SVC(kernel='linear',
degree=3,
gamma=1 / 300,
coef0=0,
C=1,
cache_size=4000,
class_weight='balanced',
verbose=True)
svc.fit(X, y_label)
beta = svc.coef_ # 1xd
# X_pred = svc.decision_function(X) nx1
# X_nums = np.concatenate([base_workspace['X'],base_workspace['X_test']])
X_proj = X_nums - (
(svc.decision_function(X_nums) / beta @ beta.T) @ beta
) # nxd
pca = PCA(n_components=1)
pca.fit(X_proj)
self.proj_pca_component = pca.components_ # 1xd
error = self.evaluate_w(base_workspace['X_test'],
self.proj_pca_component,
base_workspace['y_test'])
return error
elif model == 'kernel_proj_pca':
pass
else:
assert False
res = minimizer.minimize(space, n_calls=40, verbose=True, x0=x0)
if self.save_results:
skopt.dump(res, self.name + '.pkl', store_objective=False)
params = {type: v for type, v in zip(minimizer.optimize_types, res.x)}
if hasattr(minimizer, 'model_fixed_params'):
params = {**params, **minimizer.model_fixed_params}
error = self.fit_test_best_model(minimizer.model, base_workspace['X'],
base_workspace['y'],
base_workspace['X_test'],
base_workspace['y_test'], **params)
return error
!wget https://developer.nvidia.com/compute/cuda/9.0/Prod/local_installers/cuda-repo-ubuntu1704-9-0-local_9.0.176-1_amd64-deb
!ls # Check if required cuda 9.0 amd64-deb file is downloaded
!dpkg -i cuda-repo-ubuntu1704-9-0-local_9.0.176-1_amd64-deb
!ls /var/cuda-repo-9-0-local | grep .pub
!apt-key add /var/cuda-repo-9-0-local/7fa2af80.pub
!apt-get update
!sudo apt-get install cuda-9.0
!pip install thundersvm
#import thundersvm
from thundersvm import SVC
model = SVC(C=100, kernel='rbf')
model.fit(first_tensor, final_label)
svm_prediction = model.predict(first_tensor1)
svm_probability = model.predict_proba(first_tensor1)
label_test2 = []
for image , label in testloader:
label1 = []
for i in label:
if i!= 1:
j= 0
label1.append(j)
else:
label1.append(i)
def polynomialSVM(whid):
# This function models the SVM of orders in relation (the features) to each warehouse based on the percentage
# of products the warehouse of focus can supply and the distance to warehouse of focus
# Retrieve Dataset
dataset = _is_order_optimized_('./assets/warehouse_' + str(whid) +
'.csv', )
X = dataset.drop('classifier', axis=1)
Y = dataset["classifier"]
# Splitting data into train and test
xTrain, xTest, yTrain, yTest = train_test_split(X,
Y,
test_size=0.20,
random_state=0)
# Feature scaling
sc = StandardScaler()
xTrain = sc.fit_transform(xTrain)
xTest = sc.fit_transform(xTest)
# Fitting Kernel SVM to the Training set.
svcClassifier = SVC(
kernel='rbf',
random_state=0,
max_mem_size=50000,
n_jobs=8,
C=100 # This value can vary for whether the margin is too 'hard' or too 'soft'
)
# pickling the files (serializing them and storing them)
# This way, the model can run the data against other data
svcClassifier.fit(xTrain, yTrain)
svc_pickle = './assets/sv_pickle_rbf_' + str(whid) + '.sav'
pickle.dump(svcClassifier, open(svc_pickle, 'wb'))
# Predicting the test results
polyPred = svcClassifier.predict(xTest)
print(polyPred)
# Confusion Matrix Print: SVM Classifier polyTest against the Test Labeled Data yTest
print("Confusion Matrix")
print(confusion_matrix(yTest, polyPred))
print("\n")
# Classification report
print("Classification Report")
print(classification_report(yTest, polyPred))
print("\n")
# Applying k-fold cross validation for accuracy purposes
accuracies = cross_val_score(estimator=svcClassifier,
X=xTrain,
y=yTrain,
cv=10)
print(accuracies.mean())
print(accuracies.std())
# Visualising the Test set results
from matplotlib.colors import ListedColormap
X_set, y_set = xTest, yTest
X1, X2 = np.meshgrid(
np.arange(start=X_set[:, 0].min() - 1,
stop=X_set[:, 0].max() + 1,
step=0.01),
np.arange(start=X_set[:, 1].min() - 1,
stop=X_set[:, 1].max() + 1,
step=0.01))
plt.contourf( # Creating the contouring lines
X1,
X2,
svcClassifier.\
predict(np.array([X1.ravel(), X2.ravel()]).T).\
reshape(X1.shape),
alpha = 0.5,
cmap = ListedColormap(('blue', 'black'))
)
plt.xlim(X1.min(), X1.max())
plt.ylim(X2.min(), X2.max())
for i, j in enumerate(np.unique(y_set)): # Creating the scatter plots
plt.scatter(X_set[y_set == j, 0],
X_set[y_set == j, 1],
c=ListedColormap(('red', 'green'))(i),
label=j)
# labeling each plot
plt.title('Kernel SVM (Training Data) Warehouse: ' + str(whid))
plt.xlabel('Distance From Warehouse')
plt.ylabel('Percentage of Available Products')
plt.legend()
plt.savefig('./assets/RBF' + str(whid) + '_' + str(int(time.time())) +
'.png')
plt.close()
return False
mesh._polydata.GetCell(i).GetPointId(
2)) # don't need to copy
fine_cells = cells
barycenters = mesh3.cellCenters() # don't need to copy
fine_barycenters = mesh.cellCenters() # don't need to copy
if upsampling_method == 'SVM':
#clf = SVC(kernel='rbf', gamma='auto', probability=True, gpu_id=gpu_id)
clf = SVC(kernel='rbf', gamma='auto', gpu_id=gpu_id)
# train SVM
#clf.fit(mesh2.cells, np.ravel(refine_labels))
#fine_labels = clf.predict(fine_cells)
clf.fit(barycenters, np.ravel(refine_labels))
fine_labels = clf.predict(fine_barycenters)
fine_labels = fine_labels.reshape(-1, 1)
elif upsampling_method == 'KNN':
neigh = KNeighborsClassifier(n_neighbors=3)
# train KNN
#neigh.fit(mesh2.cells, np.ravel(refine_labels))
#fine_labels = neigh.predict(fine_cells)
neigh.fit(barycenters, np.ravel(refine_labels))
fine_labels = neigh.predict(fine_barycenters)
fine_labels = fine_labels.reshape(-1, 1)
mesh.addCellArray(fine_labels, 'Label')
vedo.write(
mesh,