def train_single(self, input_vector, target_vector):
""""
Forward Propagation
input: input_vector [784], target_vector [10]
H1: output_vector [100], weights_in_hidden[100, 784], output_hidden[100]
output: output_vector2 [10], weight_hidden_output[10, 100], output_network[10]
loss scalar 2.3
Backward Propagation
gradient [10], derived [10], tmp2 [10], hidden_errors [100,10], tmp5 [100,10]
"""
input_vector = np.array(input_vector, ndmin=2).T
target_vector = np.array(target_vector, ndmin=2).T
output_vector1 = np.dot(self.weights_in_hidden, input_vector)
output_hidden = Activation.leakyReLU(output_vector1)
output_vector2 = np.dot(self.weights_hidden_output, output_hidden)
output_network = Activation.leakyReLU(output_vector2)
loss = Cross_Entropy.calc(output_network, target_vector)
gradient = Cross_Entropy.derived_calc(output_network, target_vector)
# update the weights:
derived1 = Derivative.leakyReLU(gradient)
tmp2 = derived1 * (loss * gradient)
# calculate hidden errors:
hidden_errors = np.dot(self.weights_hidden_output.T, loss * derived1)
# update the weights:
tmp5 = hidden_errors * Derivative.leakyReLU(output_hidden)
self.weights_hidden_output += self.learning_rate * np.dot(
tmp2, output_hidden.T)
self.weights_in_hidden += self.learning_rate * np.dot(
tmp5, input_vector.T)
class Net:
def __init__(self, method=Method.Sigmoid):
self.weights = [] # Current weights
self.old_weights = [] # Last time weights
self.output = 0.0 # Neuron output
self.inputted_features = [] # Inputted features
self.summed_signal = 0.0 # Summed singal (the summation of input)
self.learning_rate = 0.8 # Learning rate
self.activition = Activation() # Activation function
# 加总信号
def summarize_inputs(self, features=[]):
if not features:
return 0.0
self.inputted_features = copy.deepcopy(features)
self.summed_signal = np.dot(self.inputted_features,
self.weights) # a = X * W
self.output = self.activition.activate(self.summed_signal) # b = f(a)
return self.output
# 更新权重
def update_weights(self, error_value=0.0):
self.old_weights = copy.deepcopy(self.weights)
for index, old_weight in enumerate(self.old_weights):
# new weight = old weight + learning rate * error_value(i) * input
new_weight = old_weight + self.learning_rate * error_value * self.inputted_features[
index]
self.weights[index] = new_weight
def differential_activition(self):
return self.activition.differentiate(self.output)
def __init__(self, num_neurons, input_shape):
print(
'Adding Layer: input_shape: {}, number of neurons: {}, output_shape: {}'
.format(input_shape, num_neurons, num_neurons))
# Let's initialize the weights in interval [0,1) for respective synaptic inputs
self.weights = np.random.uniform(low=0, high=1, size=input_shape)
# Lets initialize the biases all with value '1' for every neuron in current layer
self.biases = np.ones(num_neurons)
# Lets initialize the activation_potentials all with value '0' for every neuron in current layer
self.activation_potentials = np.zeros(num_neurons)
# Outputs of this layer
self.outputs = np.zeros(num_neurons)
# Local Gradients of all the neurons in current layer
self.local_gradients = np.zeros(num_neurons)
# And finally the activation function, for non-linearity of outputs
self.activation = Activation()
# Inputs to this layer -> Outputs from previous layer
self.previous_layers_outputs = []
print('Added Layer ... ')
def back_prop(self,incoming_grad,lr = 0.01):
"""
:param incoming_grad: should be a 1d vector
:return:
"""
if self.activation !="softmax":
if self.activation == "sigmoid":
self.local_grad = A.grad_sigmoid(self.sum_of_incoming)
elif self.activation == "relu":
self.local_grad = A.grad_relu(self.sum_of_incoming)
self.local_grad *= incoming_grad #element wise multiplication
else:
self.local_grad = incoming_grad
temp_to_pass_back = [self.local_grad for _ in range(self.weights.shape[0])]
temp_to_pass_back = np.asarray(temp_to_pass_back)
bias_grad = self.local_grad
weight_grad = np.matmul(self.input.reshape((self.input.shape[0],1)),
self.local_grad.reshape((1,self.local_grad.shape[0])))
back_grad = self.weights * temp_to_pass_back
self.biases =self.biases - lr * bias_grad
self.weights = self.weights - lr * weight_grad
##propogate gradient to previous layer
return np.sum(back_grad,axis=1)
def __init__(self, start=0, end=1, frequency = 50, duty_cycle = 0.4, scaling = 1, non_linearity = -1, shape_="monophasic"):
self.start = start
self.end = end
self.lit_data = DataLoader()
self.a = Activation(frequency, duty_cycle, scaling, non_linearity)
self.a.get_activation_signal(self.lit_data.activation_function(), shape=shape_)
self.a_sol = Activation(frequency, duty_cycle, scaling, non_linearity)
self.a_sol.get_activation_signal(self.lit_data.activation_function_soleus(), shape=shape_)
rest_length_soleus = self.soleus_length(23.7*np.pi/180)*1.015
rest_length_tibialis = self.tibialis_length(-37.4*np.pi/180)*0.9158 # lower is earlier activation
print(rest_length_soleus)
print(rest_length_tibialis)
soleus_f0m = 2600.06
self.soleus = HillTypeMuscle(soleus_f0m, .1342*rest_length_soleus, .8658*rest_length_soleus)
self.tibialis = HillTypeMuscle(605.3465, .2206*rest_length_tibialis, .7794*rest_length_tibialis)
# theta, velocity, initial CE length of soleus, initial CE length of TA
self.initial_state = np.array([self.lit_data.ankle_angle(self.start)[0]*np.pi/180,
self.lit_data.ankle_velocity(self.start)[0]*np.pi/180,
0.827034,
1.050905])
print(self.initial_state)
self.time = None
self.x1 = None
self.x2 = None
self.x3 = None
self.x4 = None
def train_single(self, input_vector, target_vector):
"""
input_vector and target_vector can be tuple,
list or ndarray
"""
input_vector = np.array(input_vector, ndmin=2).T
target_vector = np.array(target_vector, ndmin=2).T
output_vector1 = np.dot(self.weights_in_hidden, input_vector)
output_hidden = Activation.reLU(output_vector1)
output_vector2 = np.dot(self.weights_hidden_output, output_hidden)
#output_network = Activation.sigmoid(output_vector2)
output_network = Activation.reLU(output_vector2)
output_errors = target_vector - output_network
# update the weights:
#tmp = output_errors * Derivative.sigmoid(output_network)
tmp = output_errors * Derivative.reLU(output_network)
tmp = self.learning_rate * np.dot(tmp, output_hidden.T)
self.weights_hidden_output += tmp
# calculate hidden errors:
hidden_errors = np.dot(self.weights_hidden_output.T, output_errors)
# ----------------------------------------------------------------------
# update the weights:
tmp = hidden_errors * Derivative.reLU(output_hidden)
# -----------------------------------------------------------------------
self.weights_in_hidden += self.learning_rate * np.dot(
tmp, input_vector.T)
def __init__(self, method=Method.Sigmoid):
self.weights = [] # Current weights
self.old_weights = [] # Last time weights
self.output = 0.0 # Neuron output
self.inputted_features = [] # Inputted features
self.summed_signal = 0.0 # Summed singal (the summation of input)
self.learning_rate = 0.8 # Learning rate
self.activition = Activation() # Activation function
def __init__(self):
self.tag = self.__class__.__name__
self.samples = [] # 所有的训练样本（特征值)
self.targets = [] # 范本的目标输出
self.weights = [] # 权重
self.bias = 0.0 # 偏权值
self.learning_rate = 1.0 # 学习速率
self.max_iteration = 1 # 最大迭代数
self.convergence = 0.001 # 收敛误差
self.activation = Activation()
def __init__(self, has_recurrent=False):
self.weights = [] # <number>
self.recurrent_weights = [] # <number>
self.bias = 0.0
self.delta_value = 0.0 # Current delta value will be next delta value.
self.has_recurrent = has_recurrent # Has recurrent inputs ? (hidden net has recurrent, but output net not.
self.activation = Activation(
) # 活化函式的 Get, Set 都在这里： net.activation.method.
# 有另外开 self.activation_method 来方便存取
self.output = NetOutput()
self.timesteps = [] # <Timestep Object>
def run(self, input_vector):
# input_vector can be tuple, list or ndarray
input_vector = np.array(input_vector, ndmin=2).T
# 1st layer
output_vector = np.dot(self.weights_in_hidden, input_vector)
output_vector = Activation.leakyReLU(output_vector)
# 2nd layer
output_vector = np.dot(self.weights_hidden_output, output_vector)
output_vector = Activation.leakyReLU(output_vector)
return output_vector
def test_activation_functions():
# Ensure the correct values are calculated by the member functions
# We compare against sklearn functions
from sklearn.neural_network._base import tanh, relu
from scipy.special import expit as sigmoid
N = 100
act = Activation(function='sigmoid')
x = np.random.uniform(-10.0, 10.0, size=(N, 1))
assert act(x) == pytest.approx(sigmoid(x))
act.set(function='tanh')
x = np.random.uniform(-10.0, 10.0, size=(N, 1))
assert act(x) == pytest.approx(tanh(x))
act.set(function='relu')
x = np.random.uniform(-10.0, 10.0, size=(N, 1))
assert act(x) == pytest.approx(relu(x))
alpha = 2.5082958
act.set(function='leakyrelu', alpha=alpha)
x = np.random.uniform(-10.0, 10.0, size=(N, 1))
assert act(x) == pytest.approx((x >= 0.0) * x + (x < 0.0) * alpha * x)
def compute_gradients(self, test_input, test_output):
test_input = np.array(test_input)
test_output = np.array(test_output)
deltas = list()
slopes = list()
error = -(test_output - self.predict(test_input))
for i, (activation, inputs) in enumerate(zip(reversed(self._activations[:-1]), reversed(self._layer_inputs))):
if i == 0:
deltas.append(np.multiply(error, Activation.sigmoid_prime(inputs)))
else:
deltas.append(np.dot(deltas[-1], self._layer_weights[-i].T) * Activation.sigmoid_prime(inputs))
slopes.append(np.multiply(activation.T, deltas[-1]))
slopes = [slope.ravel() for slope in reversed(slopes)]
return np.concatenate(slopes)
def feed_forward(self, x): # aka forward propagation
l_actvtns = x
for l_ix in range(1, self.l_number):
l_values = np.dot(l_actvtns, self.l_weights[l_ix]) \
+ self.l_biases[l_ix]
l_actvtns = aa.fn(self.actvtn_types[l_ix], l_values)
self.l_actvtns[l_ix] = l_actvtns
def predict(self, input_matrix):
input_matrix = np.array(input_matrix)
self._activations = [input_matrix]
self._layer_inputs = [input_matrix]
for layer_weight in self._layer_weights:
self._layer_inputs.append(np.dot(self._activations[-1], layer_weight))
self._activations.append(Activation.sigmoid(self._layer_inputs[-1]))
return self._activations[-1]
def test_activation_init():
# Ensure the setup is handled correctly when initializing an instance
# of the activation class
act = Activation()
# Default values
assert act.function == act._sigmoid
assert act.alpha == pytest.approx(0.01)
# String to correct function conversion
act = Activation(function='tanh')
assert act.function == act._tanh
act = Activation(function='relu')
assert act.function == act._relu
act = Activation(function='leakyrelu')
assert act.function == act._leakyrelu
act = Activation(function='sigmoid')
assert act.function == act._sigmoid
# Check wrong string error is handled correctly
caught = False
try:
act = Activation(function='this_is_not_an_allowed_string')
except ValueError as e:
caught = True
assert caught == True
# Check alpha value specification is handled correctly
alpha = 0.867
act = Activation(function='relu', alpha=alpha)
assert act.alpha == pytest.approx(alpha)
def prediction(self, x):
l_actvtns = x
for l_ix in range(1, self.l_number):
l_values = np.dot(l_actvtns, self.l_weights[l_ix]) \
+ self.l_biases[l_ix]
l_actvtns = aa.fn(self.actvtn_types[l_ix], l_values)
return l_actvtns
def predict(w, b, X):
m = X.shape[1]
Y_prediction = np.zeros((1, m))
w = w.reshape(X.shape[0], 1)
A = Activation.sigmoid(np.dot(w.T, X) + b)
for i in range(A.shape[1]):
Y_prediction[0, i] = 1 if A[0, i] > 0.5 else 0
return Y_prediction
def predict(self, input_matrix):
input_matrix = np.array(input_matrix)
self._activations = [input_matrix]
self._layer_inputs = [input_matrix]
for layer_weight in self._layer_weights:
self._layer_inputs.append(
np.dot(self._activations[-1], layer_weight))
self._activations.append(Activation.sigmoid(
self._layer_inputs[-1]))
return self._activations[-1]
def propagate(w, b, x, y):
m = x.shape[1]
A = Activation.sigmoid(np.dot(w.T, x) + b)
cost = (-1 / m) * np.sum(y * np.log(A) + (1 - y) * (np.log(1 - A)))
dz = A - y
dw = (1 / m) * np.dot(x, dz.T)
db = (1 / m) * np.sum(dz)
cost = np.squeeze(cost)
grads = {"dw": dw, "db": db}
return grads, cost
def compute_gradients(self, test_input, test_output):
test_input = np.array(test_input)
test_output = np.array(test_output)
deltas = list()
slopes = list()
error = -(test_output - self.predict(test_input))
for i, (activation, inputs) in enumerate(
zip(reversed(self._activations[:-1]),
reversed(self._layer_inputs))):
if i == 0:
deltas.append(
np.multiply(error, Activation.sigmoid_prime(inputs)))
else:
deltas.append(
np.dot(deltas[-1], self._layer_weights[-i].T) *
Activation.sigmoid_prime(inputs))
slopes.append(np.multiply(activation.T, deltas[-1]))
slopes = [slope.ravel() for slope in reversed(slopes)]
return np.concatenate(slopes)
def __init__(self) -> None:
super().__init__()
# hyper params
self.alpha: float = 1
self.lambda_: float = 0
self.c: float = 0
self.gamma: float = 0
# model params
self.layers = []
# engine params
self.act: Activation() = None
self.reg: Regularization() = None
self.opt: Optimizer() = None
def __init__(self,
architecture=[784, 100, 10],
activation='sigmoid',
learning_rate=0.1,
momentum=0.5,
weight_decay=1e-4,
dropout=0.5,
early_stopping=True,
seed=99):
"""
Neural network model initializer.
"""
# Attributes
self.architecture = architecture
self.activation = activation
self.learning_rate = learning_rate
self.momentum = momentum
self.weight_decay = weight_decay
self.dropout = dropout
self.early_stopping = early_stopping
self.seed = seed
# Turn `activation` and `learning_rate` to class instances
if not isinstance(self.activation, Activation):
self.activation = Activation(self.activation)
if not isinstance(self.learning_rate, LearningRate):
self.learning_rate = LearningRate(self.learning_rate)
# Initialize a list of layers
self.layers = []
for i, (n_in,
n_out) in enumerate(zip(architecture[:-2],
architecture[1:-1])):
l = HiddenLayer('layer{}'.format(i), n_in, n_out, self.activation,
self.learning_rate, self.momentum,
self.weight_decay, self.dropout, self.seed + i)
self.layers.append(l)
# Output layer
n_in, n_out = architecture[-2], architecture[-1]
l = OutputLayer('output_layer', n_in, n_out, self.learning_rate,
self.momentum, self.weight_decay, self.dropout,
self.seed + i + 1)
self.layers.append(l)
# Training updates
self.epoch = 0
self.training_error = []
self.validation_error = []
self.training_loss = []
self.validation_loss = []
def store_grad(self,incoming_grad,lr = 0.01):
"""
:param incoming_grad: should be a 1d vector
:return:
"""
if self.activation !="softmax":
if self.activation == "sigmoid":
self.local_grad = A.grad_sigmoid(self.sum_of_incoming)
elif self.activation == "relu":
self.local_grad = A.grad_relu(self.sum_of_incoming)
elif self.activation == "swish":
self.local_grad = A.grad_swish(self.sum_of_incoming)
elif self.activation == "tanh":
self.local_grad = A.grad_tanh(self.sum_of_incoming)
# print("local grad is:",self.local_grad)
self.local_grad *= incoming_grad #element wise multiplication
# print("incoming grad is:",self.local_grad)
else:
self.local_grad = incoming_grad
# print("for softmax incoming grad is:",self.local_grad)
temp_to_pass_back = [self.local_grad for _ in range(self.weights.shape[0])]
temp_to_pass_back = np.asarray(temp_to_pass_back)
temp = np.matmul(self.input.reshape((self.input.shape[0],1)),
self.local_grad.reshape((1,self.local_grad.shape[0])))
self.bias_grad += self.local_grad
# print("temp is ",temp)
# print("weight grad is ",self.weight_grad)
self.weight_grad += temp
# print("weight grad is ",self.weight_grad)
##propogate gradient to previous layer
back_grad = self.weights * temp_to_pass_back
# if np.linalg.norm(back_grad)>1.0:
# print("grad explosion , inside store_grad function of layer")
return np.sum(back_grad,axis=1)
def train_single(self, input_vector, target_vector):
"""
input_vector and target_vector can be tuple,
list or ndarray
"""
input_vector = np.array(input_vector, ndmin=2).T
target_vector = np.array(target_vector, ndmin=2).T
output_vector1 = np.dot(self.weights_in_hidden, input_vector)
output_hidden = Activation.reLU(output_vector1)
output_hidden *= Dropout.get_mask(output_vector1)
output_vector2 = np.dot(self.weights_hidden_output, output_hidden)
output_network = Activation.reLU(output_vector2)
output_network *= Dropout.get_mask(output_vector2)
output_errors = target_vector - output_network
# update the weights:
#tmp = output_errors * Derivative.sigmoid(output_network)
try:
tmp = output_errors * Derivative.reLU(output_network)
tmp = self.learning_rate * np.dot(tmp, output_hidden.T)
self.weights_hidden_output += tmp
except:
print("Something went wrong when writing to the file")
# calculate hidden errors:
try:
hidden_errors = np.dot(self.weights_hidden_output.T, output_errors)
except:
print("Something went wrong when writing to the file")
# ----------------------------------------------------------------------
# update the weights:
tmp1 = Derivative.reLU(output_hidden)
tmp = hidden_errors * tmp1
# -----------------------------------------------------------------------
self.weights_in_hidden += self.learning_rate * np.dot(
tmp, input_vector.T)
def feed_forward(self, layer):
"""Feeds forward the layers values"""
for i in range(len(self.bias.weights)):
layer.nodes[i].value = 0
for i in range(len(self.nodes)):
for weight in range(len(self.nodes[i].weights)):
layer.nodes[weight].value += self.nodes[i].value * self.nodes[
i].weights[weight]
for weight in range(len(self.bias.weights)):
layer.nodes[weight].value += self.bias.weights[weight]
for w in range(len(layer.nodes)):
# use tanh as our activation function
layer.nodes[w].value = Activation.tanh(layer.nodes[w].value)
def back_propagtn(self, x, y):
learng_coeff = self.learng_eta / len(x) # scaling
l_losses = cc.dv(self.cost_type, y, self.l_actvtns[self.l_number-1]) \
+ rr.dv(self.reg_type, y, self.l_actvtns[self.l_number-1])
for l_ix in range(self.l_number - 1, 0, -1):
l_slopes = aa.dv(self.actvtn_types[l_ix], self.l_actvtns[l_ix])
l_deltas = l_losses * l_slopes
weights_nabla = self.l_actvtns[l_ix-1].T.dot(l_deltas)
biases_nabla = np.sum(l_deltas, axis=0, keepdims=True)[0]
self.l_weights[l_ix] = self.l_weights[l_ix] \
+ weights_nabla * learng_coeff
self.l_biases[l_ix] = self.l_biases[l_ix] \
+ biases_nabla * learng_coeff
l_losses = l_deltas.dot(self.l_weights[l_ix].T)
def activate_tag_callback(tag):
if display_manager.active_window() != prompt:
return
print "Activating tag: %s" % tag
display_manager.launch(load_window)
res = urllib2.urlopen("http://192.168.1.10:8000/api/activate?tag=%s" %
tag).read()
print res
res_dict = json.loads(res)
if (res_dict["status"] == "error"):
load_window.finish()
else:
name = res_dict["player"]["first_name"] + " " + res_dict["player"][
"last_name"]
team = res_dict["player"]["team"]
rule = res_dict["player"]["rule_text"]
activation = Activation(name, team, rule)
display_manager.launch_consume(activation)
def apply_activation_fun(data,activation="relu"):
if activation=="relu":
return A.relu(data)
elif activation == "softmax":
return A.softmax(data)
elif activation == "tanh":
return A.tanh(data)
elif activation == "softplus":
return A.softplus(data)
elif activation == "swish":
return A.swish(data)
elif activation == "sigmoid":
return A.sigmoid(data)
class Perceptron(Neuron):
def __init__(self, learning_rate: float, X: np.array, Y: np.array,
func: str):
super().__init__(learning_rate, X, Y)
self._activer = Activation(func)
def run(self, steps=1001, show=100):
for step in range(steps):
cost = 0
for x_n, y_n in zip(self._data, self._label):
y_pred = self._predict(x_n)
y_pred = self._activer.chooser(y_pred)
diff = y_n - int(y_pred)
self._update_w_online(x_n, diff)
self._update_bias(diff)
cost += diff**2
if step % show == 0:
print('step {0}: {1}'.format(step, cost))
self.print_status(step)
class Sigmoid(Neuron):
def __init__(self, learning_rate: float, X: np.array, Y: np.array,
func: str):
super().__init__(learning_rate, X, Y)
self._activer = Activation(func)
def run(self, steps=1001, show=100):
for step in range(steps):
z = np.dot(self._data, self._w.T) + self._bias
y_pred = self._activer.chooser(z)
error = self._label - y_pred
self._update_w_offline(self._data, error.T)
self._update_bias(error.sum())
if step % show == 0:
cost = np.mean(-self._label * np.log(y_pred) -
(1 - self._label) * np.log(1 - y_pred))
print('step {0}: {1}'.format(step, cost))
self.print_status(step)
def __init__(self,
input_layer,
num_units,
init_stddev,
activation_fun=Activation('relu')):
self.num_units = num_units
self.activation_fun = activation_fun
# the input shape will be of size (batch_size, num_units_prev)
# where num_units_prev is the number of units in the input
# (previous) layer
self.input_shape = input_layer.output_size()
# TODO ################################
# TODO: implement weight initialization
# TODO ################################
# this is the weight matrix it should have shape: (num_units_prev, num_units)
# use normal distrbution with mean 0 and stdv = init_stddev
self.W = np.random.normal(0, init_stddev,
(self.input_shape[1], num_units)) #FIXME
# and this is the bias vector of shape: (num_units)
self.b = np.random.normal(0, init_stddev, (num_units, )) #FIXME
# create dummy variables for parameter gradients
# no need to change these here!
self.dW = None
self.db = None
def set_activation(self, frequency, duty_cycle, scaling, non_linearity, shape_):
self.a = Activation(frequency, duty_cycle, scaling, non_linearity)
self.a.get_activation_signal(self.lit_data.activation_function(), shape=shape_)
self.a.plot()
