def train(self,
X,
y,
num_passes=1000,
lr=0.01,
regularization=0.01,
to_print=True):
# add gates
m_Gate = MultiplyGate()
a_Gate = AddGate()
# activate nonlinear layer
if self.activation_func == 'sigmoid':
layer = Sigmoid()
elif self.activation_func == 'tanh':
layer = Tanh()
# activate output layer
if self.output_func == 'softmax':
output = Softmax()
elif self.output_func == 'lse':
output = LSE()
# for each epoch
for epoch in range(num_passes):
# Forward propagation
input = X
forward = [(None, None, input)]
# for each layer except the last one
for i in range(len(self.W)):
mul = m_Gate.forward(self.W[i], input)
add = a_Gate.forward(mul, self.b[i])
input = layer.forward(add)
forward.append((mul, add, input))
# last output of forward propagation is an array: num_samples * num_neurons_last_layer
# Back propagation
# derivative of cumulative error from output layer
dfunc = output.calc_diff(forward[len(forward) - 1][2], y)
for i in range(len(forward) - 1, 0, -1):
# 1 layer consists of mul, add and layer
dadd = layer.backward(forward[i][1], dfunc)
# dLdb and dLdmul are functions of dLdadd
db, dmul = a_Gate.backward(forward[i][0], self.b[i - 1], dadd)
dW, dfunc = m_Gate.backward(self.W[i - 1], forward[i - 1][2],
dmul)
# Add regularization terms (b1 and b2 don't have regularization terms)
dW += regularization * self.W[i - 1]
# Gradient descent parameter update
self.b[i - 1] += -lr * db
self.W[i - 1] += -lr * dW
if to_print and epoch % 100 == 0:
print("Loss after iteration %i: %f" %
(epoch, self.calculate_loss(X, y)))
def bptt(self, x, y):
assert len(x) == len(y)
output = Softmax()
layers = self.forward_propagation(x)
dU = np.zeros(self.U.shape)
dV = np.zeros(self.V.shape)
dW = np.zeros(self.W.shape)
T = len(layers)
prev_s_t = np.zeros(self.hidden_dim)
diff_s = np.zeros(self.hidden_dim)
delta = np.zeros(self.hidden_dim)
for k in range(0, T):
t = T - k - 1
input = np.zeros(self.word_dim)
input[x[t]] = 1
if t == 0:
prev_s_t = np.zeros(self.hidden_dim)
else:
prev_s_t = layers[t - 1].s
dmulv = output.diff(layers[t].mulv, y[t])
delta, dU_t, dW_t, dV_t = layers[t].backward1(
input, prev_s_t, self.U, self.W, self.V, delta, dmulv)
dV += dV_t
dU += dU_t
dW += dW_t
return (dU, dW, dV)
def bptt(self, x, y):
assert len(x) == len(y)
output = Softmax()
layers = self.forward_propagation(x)
dU = np.zeros(self.U.shape)
dV = np.zeros(self.V.shape)
dW = np.zeros(self.W.shape)
T = len(layers)
prev_s_t = np.zeros(self.hidden_dim)
diff_s = np.zeros(self.hidden_dim)
for t in range(0, T):
dmulv = output.diff(layers[t].mulv, y[t])
input = np.zeros(self.word_dim)
input[x[t]] = 1
dprev_s, dU_t, dW_t, dV_t = layers[t].backward(input, prev_s_t, self.U, self.W, self.V, diff_s, dmulv)
prev_s_t = layers[t].s
dmulv = np.zeros(self.word_dim)
for i in range(t-1, max(-1, t-self.bptt_truncate-1), -1):
input = np.zeros(self.word_dim)
input[x[i]] = 1
prev_s_i = np.zeros(self.hidden_dim) if i == 0 else layers[i-1].s
dprev_s, dU_i, dW_i, dV_i = layers[i].backward(input, prev_s_i, self.U, self.W, self.V, dprev_s, dmulv)
dU_t += dU_i
dW_t += dW_i
dV += dV_t
dU += dU_t
dW += dW_t
return (dU, dW, dV)
def calculate_loss(self, X, y):
m_Gate = MultiplyGate()
a_Gate = AddGate()
if self.activation_func == 'sigmoid':
layer = Sigmoid()
elif self.activation_func == 'tanh':
layer = Tanh()
if self.output_func == 'softmax':
output = Softmax()
elif self.output_func == 'lse':
output = LSE()
input = X
# loop through each layer
for i in range(len(self.W)):
# X*W
mul = m_Gate.forward(self.W[i], input)
# X*W + b
add = a_Gate.forward(mul, self.b[i])
# nonlinear activation
input = layer.forward(add)
return output.eval_error(input, y)
def predict(self, X):
m_Gate = MultiplyGate()
a_Gate = AddGate()
if self.activation_func == 'sigmoid':
layer = Sigmoid()
elif self.activation_func == 'tanh':
layer = Tanh()
if self.output_func == 'softmax':
output = Softmax()
elif self.output_func == 'lse':
output = LSE()
input = X
for i in range(len(self.W)):
mul = m_Gate.forward(self.W[i], input)
add = a_Gate.forward(mul, self.b[i])
input = layer.forward(add)
if self.output_func == 'softmax':
probs = output.eval(input)
return np.argmax(probs, axis=1)
elif self.output_func == 'lse':
return (np.greater(input, 0.5)) * 1
def calculate_loss(self, x, y):
output = Softmax()
layers = self.forward_propagation(x)
loss = 0.0
for i, layer in enumerate(layers):
loss += output.loss(layer.mulv, y[i])
return loss / float(len(y))
def calculate_loss(self, X, y):
mulGate = MultiplyGate()
addGate = AddGate()
layer = Tanh()
softmaxOutput = Softmax()
input = X
for i in range(len(self.W)):
mul = mulGate.forward(self.W[i], input)
add = addGate.forward(mul, self.b[i])
input = layer.forward(add)
return softmaxOutput.loss(input, y)
def predict(self, X):
mulGate = MultiplyGate()
addGate = AddGate()
layer = Tanh()
softmaxOutput = Softmax()
input = X
for i in range(len(self.W)):
mul = mulGate.forward(self.W[i], input)
add = addGate.forward(mul, self.b[i])
input = layer.forward(add)
probs = softmaxOutput.predict(input)
return np.argmax(probs, axis=1)
def train(self,
X,
y,
num_passes=20000,
epsilon=0.01,
reg_lambda=0.01,
print_loss=False):
mulGate = MultiplyGate()
addGate = AddGate()
layer = Tanh()
softmaxOutput = Softmax()
for epoch in range(num_passes):
# Forward propagation
input = X
forward = [(None, None, input)]
for i in range(len(self.W)):
mul = mulGate.forward(self.W[i], input)
add = addGate.forward(mul, self.b[i])
input = layer.forward(add)
forward.append((mul, add, input))
# Back propagation
dtanh = softmaxOutput.diff(forward[len(forward) - 1][2], y)
for i in range(len(forward) - 1, 0, -1):
dadd = layer.backward(forward[i][1], dtanh)
db, dmul = addGate.backward(forward[i][0], self.b[i - 1], dadd)
dW, dtanh = mulGate.backward(self.W[i - 1], forward[i - 1][2],
dmul)
# Add regularization terms (b1 and b2 don't have regularization terms)
dW += reg_lambda * self.W[i - 1]
# Gradient descent parameter update
self.b[i - 1] += -epsilon * db
self.W[i - 1] += -epsilon * dW
# write log
nn_log_instance.w = self.W
nn_log_instance.b = self.b
nn_log_instance.forward = forward
nn_log_instance.write_log()
if print_loss and epoch % 1000 == 0:
print("Loss after iteration %i: %f" %
(epoch, self.calculate_loss(X, y)))
def predict(self, x):
output = Softmax()
layers = self.forward_propagation(x)
return [np.argmax(output.predict(layer.mulv)) for layer in layers]
def predict(self, x):
"""Calculate the predictions given data x."""
output = Softmax()
layers = self.forward_propagation(x)
return [np.argmax(output.predict(layer.mulv)) for layer in layers]