def __init__(self, n_hidden=50, activation_func='tanh', max_memory=5000):
self.n_hidden = n_hidden
self.activation_func = activation_func
self.model = OSELMRegressor(n_hidden=n_hidden,
activation_func=activation_func)
self.model_sec = SGDRegressor(D=6)
self.memory = ExperienceReplay(max_memory)
self.fitted = False
def __init__(self, n_hidden=50, activation_func='tanh'):
self.n_hidden = n_hidden
self.activation_func = activation_func
self.model = OSELMRegressor(n_hidden=n_hidden,
activation_func=activation_func)
self.model_sec = SGDRegressor(D=6)
self.buffer = []
self.fitted = False
class ELMRegressorACExperience(object):
def __init__(self, n_hidden=50, activation_func='tanh', max_memory=5000):
self.n_hidden = n_hidden
self.activation_func = activation_func
self.model = OSELMRegressor(n_hidden=n_hidden,
activation_func=activation_func)
self.model_sec = SGDRegressor(D=6)
self.memory = ExperienceReplay(max_memory)
self.fitted = False
self.steps = 0
def partial_fit(self, x, y, e=None):
# TODO: hacer algo con el valor de e!
self.memory.add_state(tuple(x), y)
if self.fitted is False or self.steps > self.n_hidden:
# Then we have to wait until a number of 'n_hidden' samples arrives
self.model_sec.partial_fit(x, y, e)
if len(self.memory) >= self.n_hidden:
# Now we can fit the model for first time, and from here start online learning
train_data = self.memory.get_sample(max_items=self.n_hidden)
x, y = [], []
for x_i, y_i in train_data:
x.append(x_i)
y.append(y_i)
self.model.fit(X=x, y=np.asarray(y).ravel())
self.fitted = True
self.steps = 0
else:
# We can make online learning
self.model.fit(X=x, y=np.asarray(y).ravel())
self.steps += 1
def predict(self, x):
if self.fitted:
y = self.model.predict(x)
if isinstance(y, np.ndarray):
y = y[0]
else:
y = self.model_sec.predict(x)
if isinstance(y, np.ndarray):
y = y[0]
return y
class ELMRegressorAC(object):
def __init__(self, n_hidden=50, activation_func='tanh'):
self.n_hidden = n_hidden
self.activation_func = activation_func
self.model = OSELMRegressor(n_hidden=n_hidden,
activation_func=activation_func)
self.model_sec = SGDRegressor(D=6)
self.buffer = []
self.fitted = False
def partial_fit(self, x, y, e=None):
# TODO: hacer algo con el valor de e!
if self.fitted is False:
# Then we have to wait until a number of 'n_hidden' samples arrives
self.model_sec.partial_fit(x, y, e)
self.buffer.append((x, y))
if len(self.buffer) >= self.n_hidden:
# Now we can fit the model for first time, and from here start online learning
x, y = [], []
for x_i, y_i in self.buffer:
x.append(x_i)
y.append(y_i)
self.model.fit(X=x, y=np.asarray(y).ravel())
self.fitted = True
else:
# We can make online learning
self.model.fit(X=x, y=np.asarray(y).ravel())
def predict(self, x):
if self.fitted:
y = self.model.predict(x)
if isinstance(y, np.ndarray):
y = y[0]
else:
y = self.model_sec.predict(x)
if isinstance(y, np.ndarray):
y = y[0]
return y
class ELMRegressorExperience(object):
def __init__(self, n_hidden=50, activation_func='tanh', max_memory=5000):
self.n_hidden = n_hidden
self.activation_func = activation_func
self.model = OSELMRegressor(n_hidden=n_hidden,
activation_func=activation_func)
self.model_sec = SGDRegressor(D=6)
self.memory = ExperienceReplay(max_memory)
self.fitted = False
def partial_fit(self, x, y, e=None):
# TODO: hacer algo con el valor de e!
self.model_sec.partial_fit(x, y, e)
if x[0] < 1.0:
# De-normalize
x = [int(x_i * 255) for x_i in x]
x = tuple(x)
self.memory.add_state(x, y)
if len(self.memory) >= self.n_hidden:
train_data = self.memory.get_sample(max_items=self.n_hidden)
x, y = [], []
for x_i, y_i in train_data:
# Normalize
x.append([i / 255.0 for i in x_i])
y.append(y_i)
self.model.fit(X=x, y=np.asarray(y).ravel())
self.fitted = True
def predict(self, x):
if self.fitted:
y = self.model.predict(x)
if isinstance(y, np.ndarray):
y = y[0]
else:
y = self.model_sec.predict(x)
if isinstance(y, np.ndarray):
y = y[0]
return y
def test_oselm_regression_sequential(n_samples=2000,
n_hidden=20,
activation_func='tanh',
plot=True):
x, y = make_regression(n_samples=n_samples, n_targets=1, n_features=10)
oselmr = OSELMRegressor(n_hidden=n_hidden, activation_func=activation_func)
oselmr, y_pred = make_sequential_predictions(model=oselmr, x=x, y=y)
if plot is True:
import matplotlib.pyplot as plt
axis_x = range(len(y))
plt.plot(axis_x, y, axis_x, y_pred)
plt.show()
def __init__(self, n_hidden=50, max_memory=5000):
self.n_hidden = n_hidden
self.model = OSELMRegressor(n_hidden=n_hidden)
self.model_sec = SGDRegressor(D=6)
self.memory = ExperienceReplay(max_memory)
self.fitted = False