def update_learner(self,example):
self.layers[0][:] = example[0]
# fprop
for h in range(self.n_hidden_layers):
mllin.product_matrix_vector(self.Ws[h],self.layers[h],self.layer_acts[h+1])
self.layer_acts[h+1] += self.cs[h]
mlnonlin.sigmoid(self.layer_acts[h+1],self.layers[h+1])
mllin.product_matrix_vector(self.U,self.layers[-1],self.output_act)
self.output_act += self.d
mlnonlin.softmax(self.output_act,self.output)
self.doutput_act[:] = self.output
self.doutput_act[example[1]] -= 1
self.doutput_act *= self.learning_rate/(1.+self.decrease_constant*self.n_updates)
self.dd[:] = self.doutput_act
mllin.outer(self.doutput_act,self.layers[-1],self.dU)
mllin.product_matrix_vector(self.U.T,self.doutput_act,self.dlayers[-1])
mlnonlin.dsigmoid(self.layers[-1],self.dlayers[-1],self.dlayer_acts[-1])
for h in range(self.n_hidden_layers-1,-1,-1):
self.dcs[h][:] = self.dlayer_acts[h+1]
mllin.outer(self.dlayer_acts[h+1],self.layers[h],self.dWs[h])
mllin.product_matrix_vector(self.Ws[h].T,self.dlayer_acts[h+1],self.dlayers[h])
mlnonlin.dsigmoid(self.layers[h],self.dlayers[h],self.dlayer_acts[h])
self.U -= self.dU
self.d -= self.dd
for h in range(self.n_hidden_layers-1,-1,-1):
self.Ws[h] -= self.dWs[h]
self.cs[h] -= self.dcs[h]
self.n_updates += 1
def update_learner(self, example):
self.input[self.input_order] = example
# fprop
np.multiply(self.input, self.W, self.input_times_W)
np.add.accumulate(self.input_times_W[:, :-1],
axis=1,
out=self.acc_input_times_W[:, 1:])
self.acc_input_times_W[:, 0] = 0
self.acc_input_times_W += self.c[:, np.newaxis]
mlnonlin.sigmoid(self.acc_input_times_W, self.hid)
if self.untied_weights:
np.multiply(self.hid, self.V, self.Whid)
else:
np.multiply(self.hid, self.W, self.Whid)
mllin.sum_columns(self.Whid, self.recact)
self.recact += self.b
mlnonlin.sigmoid(self.recact, self.rec)
# bprop
np.subtract(self.rec, self.input, self.drec)
self.drec *= self.alpha
self.db[:] = self.drec
if self.untied_weights:
np.multiply(self.drec, self.hid, self.dV)
np.multiply(self.drec, self.V, self.dhid)
self.dW[:] = 0
else:
np.multiply(self.drec, self.hid, self.dW)
np.multiply(self.drec, self.W, self.dhid)
mlnonlin.dsigmoid(self.hid, self.dhid, self.dacc_input_times_W)
mllin.sum_rows(self.dacc_input_times_W, self.dc)
np.add.accumulate(self.dacc_input_times_W[:, :0:-1],
axis=1,
out=self.dWenc[:, -2::-1])
self.dWenc[:, -1] = 0
self.dWenc *= self.input
self.dW += self.dWenc
self.dW *= self.learning_rate / (
1. + self.decrease_constant * self.n_updates)
self.db *= self.learning_rate / (
1. + self.decrease_constant * self.n_updates)
self.dc *= self.learning_rate / (
1. + self.decrease_constant * self.n_updates)
self.W -= self.dW
self.b -= self.db
self.c -= self.dc
if self.untied_weights:
self.dV *= self.learning_rate / (
1. + self.decrease_constant * self.n_updates)
self.V -= self.dV
self.n_updates += 1
def bprop(self,target):
"""
Computes the loss derivatives with respect to all parameters
times the current learning rate. It assumes that
``self.fprop(input)`` was called first. All the derivatives are
put in their corresponding object attributes (i.e. ``self.d*``).
"""
self.doutput_act[:] = self.output
self.doutput_act[target] -= 1
self.doutput_act *= self.learning_rate/(1.+self.decrease_constant*self.n_updates)
self.dd[:] = self.doutput_act
for k in range(self.n_k_means):
c = self.cluster_indices[k]
idx = c + k*self.n_clusters
mllin.outer(self.doutput_act,self.layers[k],self.dVs[idx])
mllin.product_matrix_vector(self.Vs[idx].T,self.doutput_act,self.dlayers[k])
#mlnonlin.dsigmoid(self.layers[k],self.dlayers[k],self.dlayer_acts[k])
if self.activation_function == 'sigmoid':
mlnonlin.dsigmoid(self.layers[k],self.dlayers[k],self.dlayer_acts[k])
elif self.activation_function == 'tanh':
mlnonlin.dtanh(self.layers[k],self.dlayers[k],self.dlayer_acts[k])
elif self.activation_function == 'reclin':
mlnonlin.dreclin(self.layers[k],self.dlayers[k],self.dlayer_acts[k])
else:
raise ValueError('activation_function must be either \'sigmoid\', \'tanh\' or \'reclin\'')
self.dcs[idx][:] = self.dlayer_acts[k]
mllin.outer(self.dlayer_acts[k],self.input,self.dWs[idx])
if self.autoencoder_regularization != 0:
self.dae_doutput_act[:] = self.dae_output
self.dae_doutput_act[:] -= self.input
self.dae_doutput_act *= 2*self.autoencoder_regularization*self.learning_rate/(1.+self.decrease_constant*self.n_updates)
self.dae_dd[:] = self.dae_doutput_act
for k in range(self.n_k_means):
c = self.cluster_indices[k]
idx = c + k*self.n_clusters
mllin.outer(self.dae_doutput_act,self.dae_layers[k],self.dae_dWsT[idx])
self.dWs[idx] += self.dae_dWsT[idx].T
mllin.product_matrix_vector(self.Ws[idx],self.dae_doutput_act,self.dae_dlayers[k])
#mlnonlin.dsigmoid(self.dae_layers[k],self.dae_dlayers[k],self.dae_dlayer_acts[k])
if self.activation_function == 'sigmoid':
mlnonlin.dsigmoid(self.dae_layers[k],self.dae_dlayers[k],self.dae_dlayer_acts[k])
elif self.activation_function == 'tanh':
mlnonlin.dtanh(self.dae_layers[k],self.dae_dlayers[k],self.dae_dlayer_acts[k])
elif self.activation_function == 'reclin':
mlnonlin.dreclin(self.dae_layers[k],self.dae_dlayers[k],self.dae_dlayer_acts[k])
else:
raise ValueError('activation_function must be either \'sigmoid\', \'tanh\' or \'reclin\'')
self.dcs[idx] += self.dae_dlayer_acts[k]
mllin.outer(self.dae_dlayer_acts[k],self.dae_input,self.dae_dWs[idx])
self.dWs[idx] += self.dae_dWs[idx]
def update_learner(self,example):
self.layers[0][:] = example[0]
# fprop
for h in range(self.n_hidden_layers):
mllin.product_matrix_vector(self.Ws[h],self.layers[h],self.layer_acts[h+1])
self.layer_acts[h+1] += self.cs[h]
if self.activation_function == 'sigmoid':
mlnonlin.sigmoid(self.layer_acts[h+1],self.layers[h+1])
elif self.activation_function == 'tanh':
mlnonlin.tanh(self.layer_acts[h+1],self.layers[h+1])
elif self.activation_function == 'reclin':
mlnonlin.reclin(self.layer_acts[h+1],self.layers[h+1])
else:
raise ValueError('activation_function must be either \'sigmoid\', \'tanh\' or \'reclin\'')
mllin.product_matrix_vector(self.U,self.layers[-1],self.output_act)
self.output_act += self.d
mlnonlin.softmax(self.output_act,self.output)
self.doutput_act[:] = self.output
self.doutput_act[example[1]] -= 1
self.doutput_act *= self.learning_rate/(1.+self.decrease_constant*self.n_updates)
self.dd[:] = self.doutput_act
mllin.outer(self.doutput_act,self.layers[-1],self.dU)
mllin.product_matrix_vector(self.U.T,self.doutput_act,self.dlayers[-1])
if self.activation_function == 'sigmoid':
mlnonlin.dsigmoid(self.layers[-1],self.dlayers[-1],self.dlayer_acts[-1])
elif self.activation_function == 'tanh':
mlnonlin.dtanh(self.layers[-1],self.dlayers[-1],self.dlayer_acts[-1])
elif self.activation_function == 'reclin':
mlnonlin.dreclin(self.layers[-1],self.dlayers[-1],self.dlayer_acts[-1])
else:
raise ValueError('activation_function must be either \'sigmoid\', \'tanh\' or \'reclin\'')
for h in range(self.n_hidden_layers-1,-1,-1):
self.dcs[h][:] = self.dlayer_acts[h+1]
mllin.outer(self.dlayer_acts[h+1],self.layers[h],self.dWs[h])
mllin.product_matrix_vector(self.Ws[h].T,self.dlayer_acts[h+1],self.dlayers[h])
if self.activation_function == 'sigmoid':
mlnonlin.dsigmoid(self.layers[h],self.dlayers[h],self.dlayer_acts[h])
elif self.activation_function == 'tanh':
mlnonlin.dtanh(self.layers[h],self.dlayers[h],self.dlayer_acts[h])
elif self.activation_function == 'reclin':
mlnonlin.dreclin(self.layers[h],self.dlayers[h],self.dlayer_acts[h])
else:
raise ValueError('activation_function must be either \'sigmoid\', \'tanh\' or \'reclin\'')
self.U -= self.dU
self.d -= self.dd
for h in range(self.n_hidden_layers-1,-1,-1):
self.Ws[h] -= self.dWs[h]
self.cs[h] -= self.dcs[h]
self.n_updates += 1
def test_dsigmoid():
"""
Testing nonlinear sigmoid deriv.
"""
input = np.random.randn(30, 20)
output = np.zeros((30, 20))
nonlinear.sigmoid(input, output)
dinput = np.zeros((30, 20))
doutput = np.random.randn(30, 20)
nonlinear.dsigmoid(output, doutput, dinput)
assert np.sum(np.abs(dinput - doutput * output * (1 - output))) < 1e-12
def test_dsigmoid():
"""
Testing nonlinear sigmoid deriv.
"""
input = np.random.randn(30,20)
output = np.zeros((30,20))
nonlinear.sigmoid(input,output)
dinput = np.zeros((30,20))
doutput = np.random.randn(30,20)
nonlinear.dsigmoid(output,doutput,dinput)
assert np.sum(np.abs(dinput-doutput*output*(1-output))) < 1e-12
def update_learner(self,example):
self.input[self.input_order] = example
# fprop
np.multiply(self.input,self.W,self.input_times_W)
np.add.accumulate(self.input_times_W[:,:-1],axis=1,out=self.acc_input_times_W[:,1:])
self.acc_input_times_W[:,0] = 0
self.acc_input_times_W += self.c[:,np.newaxis]
mlnonlin.sigmoid(self.acc_input_times_W,self.hid)
if self.untied_weights:
np.multiply(self.hid,self.V,self.Whid)
else:
np.multiply(self.hid,self.W,self.Whid)
mllin.sum_columns(self.Whid,self.recact)
self.recact += self.b
mlnonlin.sigmoid(self.recact,self.rec)
# bprop
np.subtract(self.rec,self.input,self.drec)
self.drec *= self.alpha
self.db[:] = self.drec
if self.untied_weights:
np.multiply(self.drec,self.hid,self.dV)
np.multiply(self.drec,self.V,self.dhid)
self.dW[:] = 0
else:
np.multiply(self.drec,self.hid,self.dW)
np.multiply(self.drec,self.W,self.dhid)
mlnonlin.dsigmoid(self.hid,self.dhid,self.dacc_input_times_W)
mllin.sum_rows(self.dacc_input_times_W,self.dc)
np.add.accumulate(self.dacc_input_times_W[:,:0:-1],axis=1,out=self.dWenc[:,-2::-1])
self.dWenc[:,-1] = 0
self.dWenc *= self.input
self.dW += self.dWenc
self.dW *= self.learning_rate/(1.+self.decrease_constant*self.n_updates)
self.db *= self.learning_rate/(1.+self.decrease_constant*self.n_updates)
self.dc *= self.learning_rate/(1.+self.decrease_constant*self.n_updates)
self.W -= self.dW
self.b -= self.db
self.c -= self.dc
if self.untied_weights:
self.dV *= self.learning_rate/(1.+self.decrease_constant*self.n_updates)
self.V -= self.dV
self.n_updates += 1
def apply_dactivation(self, output, doutput, dinput):
"""
Apply the derivative of the activatiun fonction
"""
if self.activation_function == "sigmoid":
mlnonlin.dsigmoid(output, doutput, dinput)
elif self.activation_function == "tanh":
mlnonlin.dtanh(output, doutput, dinput)
elif self.activation_function == "reclin":
mlnonlin.dreclin(output, doutput, dinput)
elif self.activation_function == "softmax":
dinput[:] = output * (doutput - (doutput * output).sum(axis=1).reshape((-1, 1)))
else:
raise ValueError("activation_function must be either 'sigmoid', 'tanh' or 'reclin'")
def apply_dactivation(self, output, doutput, dinput):
"""
Apply the derivative of the activatiun fonction
"""
if self.activation_function == 'sigmoid':
mlnonlin.dsigmoid(output,doutput,dinput)
elif self.activation_function == 'tanh':
mlnonlin.dtanh(output,doutput,dinput)
elif self.activation_function == 'reclin':
mlnonlin.dreclin(output,doutput,dinput)
elif self.activation_function == 'softmax':
dinput[:] = output*(doutput-(doutput*output).sum(axis=1).reshape((-1,1)))
else:
raise ValueError('activation_function must be either \'sigmoid\', \'tanh\' or \'reclin\'')
def update_learner(self, example):
self.layers[0][:] = example[0]
# fprop
for h in range(self.n_hidden_layers):
mllin.product_matrix_vector(self.Ws[h], self.layers[h],
self.layer_acts[h + 1])
self.layer_acts[h + 1] += self.cs[h]
mlnonlin.sigmoid(self.layer_acts[h + 1], self.layers[h + 1])
mllin.product_matrix_vector(self.U, self.layers[-1], self.output_act)
self.output_act += self.d
mlnonlin.softmax(self.output_act, self.output)
self.doutput_act[:] = self.output
self.doutput_act[example[1]] -= 1
self.doutput_act *= self.learning_rate / (
1. + self.decrease_constant * self.n_updates)
self.dd[:] = self.doutput_act
mllin.outer(self.doutput_act, self.layers[-1], self.dU)
mllin.product_matrix_vector(self.U.T, self.doutput_act,
self.dlayers[-1])
mlnonlin.dsigmoid(self.layers[-1], self.dlayers[-1],
self.dlayer_acts[-1])
for h in range(self.n_hidden_layers - 1, -1, -1):
self.dcs[h][:] = self.dlayer_acts[h + 1]
mllin.outer(self.dlayer_acts[h + 1], self.layers[h], self.dWs[h])
mllin.product_matrix_vector(self.Ws[h].T, self.dlayer_acts[h + 1],
self.dlayers[h])
mlnonlin.dsigmoid(self.layers[h], self.dlayers[h],
self.dlayer_acts[h])
self.U -= self.dU
self.d -= self.dd
for h in range(self.n_hidden_layers - 1, -1, -1):
self.Ws[h] -= self.dWs[h]
self.cs[h] -= self.dcs[h]
self.n_updates += 1
def update_learner(self, vec_input):
self.vec_input[self.input_order] = vec_input
#fprop
self.fprop()
#bprob, computing gradient of -log p(vec_input)
np.subtract(self.vec_recProb,self.vec_input,self.vec_grad_bias_inp)
np.multiply(self.vec_grad_bias_inp,self.mat_h,self.mat_grad_V)
np.multiply(self.vec_grad_bias_inp,self.mat_V,self.mat_grad_h)
mlnonlin.dsigmoid(self.mat_h,self.mat_grad_h,self.mat_grad_temp)
mllin.sum_rows(self.mat_grad_temp,self.vec_grad_bias_h)
np.add.accumulate(self.mat_grad_temp[:,:0:-1],axis=1,out=self.mat_grad_W[:,-2::-1])
self.mat_grad_W[:,-1] = 0
self.mat_grad_W *= self.vec_input
#update
self.vec_bias_inp -= self.learning_rate*self.vec_grad_bias_inp
self.vec_bias_h -= self.learning_rate*self.vec_grad_bias_h
self.mat_W -= self.learning_rate*self.mat_grad_W
self.mat_V -= self.learning_rate*self.mat_grad_V
def update_learner(self, vec_input):
self.vec_input[self.input_order] = vec_input
#fprop
self.fprop()
#bprob, computing gradient of -log p(vec_input)
np.subtract(self.vec_recProb, self.vec_input, self.vec_grad_bias_inp)
np.multiply(self.vec_grad_bias_inp, self.mat_h, self.mat_grad_V)
np.multiply(self.vec_grad_bias_inp, self.mat_V, self.mat_grad_h)
mlnonlin.dsigmoid(self.mat_h, self.mat_grad_h, self.mat_grad_temp)
mllin.sum_rows(self.mat_grad_temp, self.vec_grad_bias_h)
np.add.accumulate(self.mat_grad_temp[:, :0:-1],
axis=1,
out=self.mat_grad_W[:, -2::-1])
self.mat_grad_W[:, -1] = 0
self.mat_grad_W *= self.vec_input
#update
self.vec_bias_inp -= self.learning_rate * self.vec_grad_bias_inp
self.vec_bias_h -= self.learning_rate * self.vec_grad_bias_h
self.mat_W -= self.learning_rate * self.mat_grad_W
self.mat_V -= self.learning_rate * self.mat_grad_V
P2, L2, U2 = scipy.linalg.lu(A)
print "Scipy vs mathutils.linalg diff. P:", np.sum(np.abs(P - P2))
print "Scipy vs mathutils.linalg diff. L:", np.sum(np.abs(L - L2))
print "Scipy vs mathutils.linalg diff. U:", np.sum(np.abs(U - U2))
print 'Testing nonlinear sigmoid'
input = np.random.randn(30, 20)
output = np.zeros((30, 20))
nonlinear.sigmoid(input, output)
print 'NumPy vs mathutils.nonlinear diff. output:', np.sum(
np.abs(output - 1 / (1 + np.exp(-input))))
print 'Testing nonlinear sigmoid deriv.'
dinput = np.zeros((30, 20))
doutput = np.random.randn(30, 20)
nonlinear.dsigmoid(output, doutput, dinput)
print 'NumPy vs mathutils.nonlinear diff. output:', np.sum(
np.abs(dinput - doutput * output * (1 - output)))
print 'Testing nonlinear softmax'
input = np.random.randn(20)
output = np.zeros((20))
nonlinear.softmax(input, output)
print 'NumPy vs mathutils.nonlinear diff. output:', np.sum(
np.abs(output - np.exp(input) / np.sum(np.exp(input))))
print 'Testing nonlinear softplus'
input = np.random.randn(20)
output = np.zeros((20))
nonlinear.softplus(input, output)
print 'NumPy vs mathutils.nonlinear diff. output:', np.sum(
def update_learner(self, example):
# apply example to the inputs
self.layers[0][:] = example[0]
# forward propagation: compute activation values of all units
# hidden layers
for h in range(self.n_hidden_layers):
mllin.product_matrix_vector(self.Ws[h], self.layers[h], self.layer_acts[h + 1])
self.layer_acts[h + 1] += self.cs[h]
mlnonlin.sigmoid(self.layer_acts[h + 1], self.layers[h + 1])
# output layer
mllin.product_matrix_vector(self.U, self.layers[-1], self.output_act)
self.output_act += self.d
mlnonlin.softmax(self.output_act, self.output)
# back propagation: compute delta errors and updates to weights and
# biases
# TA:begin
if self.cost_function == 'CE':
self.doutput_act[:] = self.output
self.doutput_act[example[1]] -= 1
elif self.cost_function == 'SSE':
y = self.output.copy()
t = np.zeros(np.shape(y))
t[example[1]] = 1
# nr of classes
c = np.size(y)
T2 = (y-t)*y
T2 = np.array([T2])
T2 = T2.T
T2 = np.tile(T2,[1,c])
T3 = np.eye(c,c)
T3 = T3 - np.tile(y,[c,1])
# delta error at output layer
self.doutput_act = np.sum(T2*T3,axis=0)
elif self.cost_function == 'EXP':
y = self.output.copy()
t = np.zeros(np.shape(y))
t[example[1]] = 1
# nr of classes
c = np.size(y)
T1 = y-t
T1 = np.square(T1)
T1 = np.sum(T1)
T1 = T1/self.tau
T1 = np.exp(T1)
T1 = 2*T1
T2 = (y-t)*y
T2 = np.array([T2])
T2 = T2.T
T2 = np.tile(T2,[1,c])
T3 = np.eye(c,c)
T3 = T3 - np.tile(y,[c,1])
# delta error at output layer
self.doutput_act = T1 * np.sum(T2*T3,axis=0)
# TA:end
self.doutput_act *= self.learning_rate / (1. + self.decrease_constant * self.n_updates)
self.dd[:] = self.doutput_act
mllin.outer(self.doutput_act, self.layers[-1], self.dU)
mllin.product_matrix_vector(self.U.T, self.doutput_act, self.dlayers[-1])
"""
The description and argument names of dsigmoid() are unclear. In
practice, dsigmoid(s,dx,ds) computes s*(1-s)*dx element-wise and puts
the result in ds. [TA]
"""
mlnonlin.dsigmoid(self.layers[-1], self.dlayers[-1], self.dlayer_acts[-1])
for h in range(self.n_hidden_layers - 1, -1, -1):
self.dcs[h][:] = self.dlayer_acts[h + 1]
mllin.outer(self.dlayer_acts[h + 1], self.layers[h], self.dWs[h])
mllin.product_matrix_vector(self.Ws[h].T, self.dlayer_acts[h + 1], self.dlayers[h])
mlnonlin.dsigmoid(self.layers[h], self.dlayers[h], self.dlayer_acts[h])
#TA:
if not self.freeze_Ws_cs:
# update output weights and biases
self.U -= self.dU
self.d -= self.dd
# update all hidden weights and biases
for h in range(self.n_hidden_layers - 1, -1, -1):
self.Ws[h] -= self.dWs[h]
self.cs[h] -= self.dcs[h]
else:
# update output weights and biases
self.U -= self.dU
self.d -= self.dd
# # update only highest hidden layer
# h = self.n_hidden_layers - 1
# self.Ws[h] -= self.dWs[h]
# self.cs[h] -= self.dcs[h]
self.n_updates += 1
P_row[p_el] = 1 P2,L2,U2 = scipy.linalg.lu(A) print "Scipy vs mathutils.linalg diff. P:",np.sum(np.abs(P-P2)) print "Scipy vs mathutils.linalg diff. L:",np.sum(np.abs(L-L2)) print "Scipy vs mathutils.linalg diff. U:",np.sum(np.abs(U-U2)) print 'Testing nonlinear sigmoid' input = np.random.randn(30,20) output = np.zeros((30,20)) nonlinear.sigmoid(input,output) print 'NumPy vs mathutils.nonlinear diff. output:',np.sum(np.abs(output-1/(1+np.exp(-input)))) print 'Testing nonlinear sigmoid deriv.' dinput = np.zeros((30,20)) doutput = np.random.randn(30,20) nonlinear.dsigmoid(output,doutput,dinput) print 'NumPy vs mathutils.nonlinear diff. output:',np.sum(np.abs(dinput-doutput*output*(1-output))) print 'Testing nonlinear softmax' input = np.random.randn(20) output = np.zeros((20)) nonlinear.softmax(input,output) print 'NumPy vs mathutils.nonlinear diff. output:',np.sum(np.abs(output-np.exp(input)/np.sum(np.exp(input)))) print 'Testing nonlinear softplus' input = np.random.randn(20) output = np.zeros((20)) nonlinear.softplus(input,output) print 'NumPy vs mathutils.nonlinear diff. output:',np.sum(np.abs(output-np.log(1+np.exp(input)))) print 'Testing nonlinear reclin'
