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 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 apply_activation(self, input_data, output):
"""
Apply the activation function
"""
if self.activation_function == "sigmoid":
mlnonlin.sigmoid(input_data, output)
elif self.activation_function == "tanh":
mlnonlin.tanh(input_data, output)
elif self.activation_function == "reclin":
mlnonlin.reclin(input_data, output)
elif self.activation_function == "softmax":
m = input_data.max(axis=1)
output[:] = np.exp(input_data - m.reshape((-1, 1)))
output[:] /= output.sum(axis=1).reshape((-1, 1))
else:
raise ValueError("activation_function must be either 'sigmoid', 'tanh' or 'reclin'")
def apply_activation(self, input_data, output):
"""
Apply the activation function
"""
if self.activation_function == 'sigmoid':
mlnonlin.sigmoid(input_data,output)
elif self.activation_function == 'tanh':
mlnonlin.tanh(input_data,output)
elif self.activation_function == 'reclin':
mlnonlin.reclin(input_data,output)
elif self.activation_function == 'softmax':
m = input_data.max(axis=1)
output[:] = np.exp(input_data-m.reshape((-1,1)))
output[:] /= output.sum(axis=1).reshape((-1,1))
else:
raise ValueError('activation_function must be either \'sigmoid\', \'tanh\' or \'reclin\'')
def test_softplus():
"""
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.'
dinput = np.zeros((30,20))
doutput = np.random.randn(30,20)
nonlinear.dreclin(output,doutput,dinput)
print 'NumPy vs mathutils.nonlinear diff. output:',np.sum(np.abs(dinput-(input>0)*doutput))
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]
def test_softplus():
"""
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.'
dinput = np.zeros((30, 20))
doutput = np.random.randn(30, 20)
nonlinear.dreclin(output, doutput, dinput)
print 'NumPy vs mathutils.nonlinear diff. output:', np.sum(
np.abs(dinput - (input > 0) * doutput))
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.'
dinput = np.zeros((30, 20))
doutput = np.random.randn(30, 20)
nonlinear.dreclin(output, doutput, dinput)
print 'NumPy vs mathutils.nonlinear diff. output:', np.sum(
np.abs(dinput - (input > 0) * doutput))
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.' dinput = np.zeros((30,20)) doutput = np.random.randn(30,20) nonlinear.dreclin(output,doutput,dinput) print 'NumPy vs mathutils.nonlinear diff. output:',np.sum(np.abs(dinput-(input>0)*doutput))