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 cost(self, outputs, example):
hidden = outputs[0]
self.input[:] = 0
self.input[example[1]] = example[0]
mllin.product_matrix_vector(self.W.T, hidden, self.neg_input_act)
self.neg_input_act += self.b
mlnonlin.softmax(self.neg_input_act, self.neg_input_prob)
return [np.sum((self.input - self.input.sum() * self.neg_input_prob) ** 2)]
def cost(self,outputs,example):
hidden = outputs[0]
self.input[:] = 0
self.input[example[1]] = example[0]
mllin.product_matrix_vector(self.W.T,hidden,self.neg_input_act)
self.neg_input_act += self.b
mlnonlin.softmax(self.neg_input_act,self.neg_input_prob)
return [ np.sum((self.input-self.input.sum()*self.neg_input_prob)**2) ]
def test_softmax():
"""
Testing nonlinear softmax.
"""
input = np.random.randn(20)
output = np.zeros((20))
nonlinear.softmax(input,output)
assert np.sum(np.abs(output-np.exp(input)/np.sum(np.exp(input)))) < 1e-12
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_softmax():
"""
Testing nonlinear softmax.
"""
input = np.random.randn(20)
output = np.zeros((20))
nonlinear.softmax(input, output)
assert np.sum(
np.abs(output - np.exp(input) / np.sum(np.exp(input)))) < 1e-12
def fprop(self,input):
"""
Computes the output given some input. Puts the result in ``self.output``
"""
self.input[:] = input
self.output_act[:] = self.d
for k in range(self.n_k_means):
if self.n_k_means_inputs == self.input_size:
c = self.clusterings[k].compute_cluster(self.input)
else:
c = self.clusterings[k].compute_cluster(self.input[self.k_means_subset_inputs[k]])
idx = c + k*self.n_clusters
self.cluster_indices[k] = c
mllin.product_matrix_vector(self.Ws[idx],self.input,self.layer_acts[k])
self.layer_acts[k] += self.cs[idx]
#mlnonlin.sigmoid(self.layer_acts[k],self.layers[k])
if self.activation_function == 'sigmoid':
mlnonlin.sigmoid(self.layer_acts[k],self.layers[k])
elif self.activation_function == 'tanh':
mlnonlin.tanh(self.layer_acts[k],self.layers[k])
elif self.activation_function == 'reclin':
mlnonlin.reclin(self.layer_acts[k],self.layers[k])
else:
raise ValueError('activation_function must be either \'sigmoid\', \'tanh\' or \'reclin\'')
mllin.product_matrix_vector(self.Vs[idx],self.layers[k],self.output_acts[k])
self.output_act += self.output_acts[k]
mlnonlin.softmax(self.output_act,self.output)
if self.autoencoder_regularization != 0:
self.dae_input[:] = input
self.rng.shuffle(self.input_idx)
self.dae_input[self.input_idx[:int(self.autoencoder_missing_fraction*self.input_size)]] = 0
self.dae_output_act[:] = self.dae_d
for k in range(self.n_k_means):
idx = self.cluster_indices[k] + k*self.n_clusters
mllin.product_matrix_vector(self.Ws[idx],self.dae_input,self.dae_layer_acts[k])
self.dae_layer_acts[k] += self.cs[idx]
#mlnonlin.sigmoid(self.dae_layer_acts[k],self.dae_layers[k])
if self.activation_function == 'sigmoid':
mlnonlin.sigmoid(self.dae_layer_acts[k],self.dae_layers[k])
elif self.activation_function == 'tanh':
mlnonlin.tanh(self.dae_layer_acts[k],self.dae_layers[k])
elif self.activation_function == 'reclin':
mlnonlin.reclin(self.dae_layer_acts[k],self.dae_layers[k])
else:
raise ValueError('activation_function must be either \'sigmoid\', \'tanh\' or \'reclin\'')
mllin.product_matrix_vector(self.Ws[idx].T,self.dae_layers[k],self.dae_output_acts[k])
self.dae_output_act += self.dae_output_acts[k]
self.dae_output[:] = self.dae_output_act
def use_learner(self, example):
output = np.zeros((self.n_classes))
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, output)
return [output.argmax(), output]
def use_learner(self, example):
output = np.zeros((self.n_classes))
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, output)
return [output.argmax(), output]
def update_learner(self, example):
self.input[:] = 0
self.input[example[1]] = example[0]
n_words = int(self.input.sum())
# Performing CD-k
mllin.product_matrix_vector(self.W, self.input, self.hidden_act)
self.hidden_act += self.c * n_words
mlnonlin.sigmoid(self.hidden_act, self.hidden_prob)
self.neg_hidden_prob[:] = self.hidden_prob
for k in range(self.k_contrastive_divergence_steps):
if self.mean_field:
self.hidden[:] = self.neg_hidden_prob
else:
np.less(self.rng.rand(self.hidden_size), self.neg_hidden_prob, self.hidden)
mllin.product_matrix_vector(self.W.T, self.hidden, self.neg_input_act)
self.neg_input_act += self.b
mlnonlin.softmax(self.neg_input_act, self.neg_input_prob)
if self.mean_field:
self.neg_input[:] = n_words * self.neg_input_prob
else:
self.neg_input[:] = self.rng.multinomial(n_words, self.neg_input_prob)
mllin.product_matrix_vector(self.W, self.neg_input, self.neg_hidden_act)
self.neg_hidden_act += self.c * n_words
mlnonlin.sigmoid(self.neg_hidden_act, self.neg_hidden_prob)
mllin.outer(self.hidden_prob, self.input, self.deltaW)
mllin.outer(self.neg_hidden_prob, self.neg_input, self.neg_stats)
self.deltaW -= self.neg_stats
np.subtract(self.input, self.neg_input, self.deltab)
np.subtract(self.hidden_prob, self.neg_hidden_prob, self.deltac)
self.deltaW *= self.learning_rate / (1.0 + self.decrease_constant * self.n_updates)
self.deltab *= self.learning_rate / (1.0 + self.decrease_constant * self.n_updates)
self.deltac *= n_words * self.learning_rate / (1.0 + self.decrease_constant * self.n_updates)
self.W += self.deltaW
self.b += self.deltab
self.c += self.deltac
self.n_updates += 1
def update_learner(self,example):
self.input[:] = 0
self.input[example[1]] = example[0]
n_words = int(self.input.sum())
# Performing CD-k
mllin.product_matrix_vector(self.W,self.input,self.hidden_act)
self.hidden_act += self.c*n_words
mlnonlin.sigmoid(self.hidden_act,self.hidden_prob)
self.neg_hidden_prob[:] = self.hidden_prob
for k in range(self.k_contrastive_divergence_steps):
if self.mean_field:
self.hidden[:] = self.neg_hidden_prob
else:
np.less(self.rng.rand(self.hidden_size),self.neg_hidden_prob,self.hidden)
mllin.product_matrix_vector(self.W.T,self.hidden,self.neg_input_act)
self.neg_input_act += self.b
mlnonlin.softmax(self.neg_input_act,self.neg_input_prob)
if self.mean_field:
self.neg_input[:] = n_words*self.neg_input_prob
else:
self.neg_input[:] = self.rng.multinomial(n_words,self.neg_input_prob)
mllin.product_matrix_vector(self.W,self.neg_input,self.neg_hidden_act)
self.neg_hidden_act += self.c*n_words
mlnonlin.sigmoid(self.neg_hidden_act,self.neg_hidden_prob)
mllin.outer(self.hidden_prob,self.input,self.deltaW)
mllin.outer(self.neg_hidden_prob,self.neg_input,self.neg_stats)
self.deltaW -= self.neg_stats
np.subtract(self.input,self.neg_input,self.deltab)
np.subtract(self.hidden_prob,self.neg_hidden_prob,self.deltac)
self.deltaW *= self.learning_rate/(1.+self.decrease_constant*self.n_updates)
self.deltab *= self.learning_rate/(1.+self.decrease_constant*self.n_updates)
self.deltac *= n_words*self.learning_rate/(1.+self.decrease_constant*self.n_updates)
self.W += self.deltaW
self.b += self.deltab
self.c += self.deltac
self.n_updates += 1
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 use_learner(self,example):
output = np.zeros((self.n_classes))
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,output)
return [output.argmax(),output]
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'
input = np.random.randn(30, 20)
output = np.zeros((30, 20))
nonlinear.reclin(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
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' input = np.random.randn(30,20) output = np.zeros((30,20)) nonlinear.reclin(input,output) print 'NumPy vs mathutils.nonlinear diff. output:',np.sum(np.abs(output-(input>0)*input)) print 'Testing nonlinear reclin deriv.'
