def tanh_trans(X, mode='ff'):
"""Compute feedforward and backprop for tanh nonlinearity."""
if (mode == 'ff'):
F = gp.tanh(X)
if (mode == 'bp'):
F = (1.0 - X['A']**2.0) * X['dLdA']
return F
def tanh_trans(X, mode='ff'):
"""Compute feedforward and backprop for tanh nonlinearity."""
if (mode == 'ff'):
F = gp.tanh(X)
if (mode == 'bp'):
F = (1.0 - X['A']**2.0) * X['dLdA']
return F
def feedforward(self, train_set_x):
self.activations = []
self.activations.append(train_set_x)
for i in range(self.n_layers):
current_activations = gnp.tanh(gnp.dot(self.activations[i], self.W_params[i]) + self.b_params[i])
self.activations.append(current_activations)
#output layers
self.final_layer_output = gnp.dot(self.activations[self.n_layers], self.W_params[self.n_layers]) + self.b_params[self.n_layers]
def parameter_prediction(self, test_set_x):
test_set_x = gnp.as_garray(test_set_x)
current_activations = test_set_x
for i in range(self.n_layers):
current_activations = gnp.tanh(gnp.dot(current_activations, self.W_params[i]) + self.b_params[i])
final_layer_output = gnp.dot(current_activations, self.W_params[self.n_layers]) + self.b_params[self.n_layers]
return final_layer_output.as_numpy_array()
def parameter_prediction(self, test_set_x):
test_set_x = gnp.as_garray(test_set_x)
current_activations = test_set_x
for i in range(self.n_layers):
current_activations = gnp.tanh(gnp.dot(current_activations, self.W_params[i]) + self.b_params[i])
final_layer_output = gnp.dot(current_activations, self.W_params[self.n_layers]) + self.b_params[self.n_layers]
return final_layer_output.as_numpy_array()
def feedforward(self, train_set_x):
self.activations = []
self.activations.append(train_set_x)
for i in range(self.n_layers):
current_activations = gnp.tanh(gnp.dot(self.activations[i], self.W_params[i]) + self.b_params[i])
self.activations.append(current_activations)
#output layers
self.final_layer_output = gnp.dot(self.activations[self.n_layers], self.W_params[self.n_layers]) + self.b_params[self.n_layers]
def __call__(self, X):
return g.tanh(X)
def parameter_prediction_trajectory(self, test_set_x, test_set_y,
mean_matrix, std_matrix):
test_set_x = gnp.garray(test_set_x)
current_activations = test_set_x
for i in range(self.n_layers):
input_data = current_activations
current_activations = gnp.tanh(
gnp.dot(input_data, self.W_params[i]) + self.b_params[i])
final_layer_output = gnp.dot(
current_activations,
self.W_params[self.n_layers]) + self.b_params[self.n_layers]
final_layer_output = final_layer_output * gnp.garray(
std_matrix) + gnp.garray(mean_matrix)
frame_number = final_layer_output.shape[0]
final_layer_output = final_layer_output.T
obs_mat = gnp.zeros((60, frame_number * 3))
traj_err_mat = gnp.zeros((60, frame_number))
var_base = np.zeros((60, 3))
static_indice = []
delta_indice = []
acc_indice = []
for i in range(60):
static_indice.append(i)
delta_indice.append(i + 60)
acc_indice.append(i + 120)
obs_mat[:, 0:frame_number] = final_layer_output[static_indice, :]
obs_mat[:, frame_number:frame_number *
2] = final_layer_output[delta_indice, :]
obs_mat[:, frame_number * 2:frame_number *
3] = final_layer_output[acc_indice, :]
var_base[:, 0] = std_matrix[0, static_indice].T
var_base[:, 1] = std_matrix[0, delta_indice].T
var_base[:, 2] = std_matrix[0, acc_indice].T
var_base = np.reshape(var_base, (60 * 3, 1))
var_base = var_base**2
sub_dim_list = []
for i in range(60):
sub_dim_list.append(1)
sub_dim_start = 0
for sub_dim in sub_dim_list:
wuw_mat, wu_mat = self.pre_wuw_wu(
frame_number, sub_dim,
var_base[sub_dim_start * 3:sub_dim_start * 3 + sub_dim * 3])
obs_mu = obs_mat[sub_dim_start:sub_dim_start + sub_dim, :].reshape(
(frame_number * 3 * sub_dim, 1))
wuwwu = gnp.dot(wuw_mat, wu_mat)
mlpg_traj = gnp.dot(wuwwu, obs_mu)
sub_std_mat = std_matrix[:, static_indice].T
sub_mu_mat = mean_matrix[:, static_indice].T
sub_std_mat = sub_std_mat[sub_dim_start:sub_dim_start +
sub_dim, :].reshape(
(frame_number * sub_dim, 1))
sub_mu_mat = sub_mu_mat[sub_dim_start:sub_dim_start +
sub_dim, :].reshape(
(frame_number * sub_dim, 1))
ref_y = test_set_y[:, static_indice].T
ref_y = ref_y[sub_dim_start:sub_dim_start + sub_dim, :].reshape(
(frame_number * sub_dim, 1))
ref_y = ref_y * sub_std_mat + sub_mu_mat
traj_err = (mlpg_traj - ref_y) # mlpg_traj ref_y
traj_err_mat[sub_dim_start:sub_dim_start +
sub_dim, :] = traj_err.reshape(
(sub_dim, frame_number))
sub_dim_start = sub_dim_start + sub_dim
validation_losses = gnp.sum(traj_err_mat[1:60, :].T**2, axis=1)
validation_losses = validation_losses**0.5
return validation_losses.as_numpy_array()
def activation(self, netInput):
return gnp.tanh(netInput)
def forward(self, A):
return gnp.tanh(A)
def activation(self, netInput):
return gnp.tanh(netInput)
def forward_prop(self, x):
return gnp.tanh(x)
def backwardOneStep(self, a):
z = gp.dot(a, self.W2) + self.b2
return gp.tanh(z)
def forwardOneStep(self, a):
z = gp.dot(a, self.W1) + self.b1
return gp.tanh(z)
def forward_prop(self, x):
return gnp.tanh(x)
def parameter_prediction_trajectory(self, test_set_x, test_set_y, mean_matrix, std_matrix):
test_set_x = gnp.garray(test_set_x)
current_activations = test_set_x
for i in xrange(self.n_layers):
input_data = current_activations
current_activations = gnp.tanh(gnp.dot(input_data, self.W_params[i]) + self.b_params[i])
final_layer_output = gnp.dot(current_activations, self.W_params[self.n_layers]) + self.b_params[self.n_layers]
final_layer_output = final_layer_output * gnp.garray(std_matrix) + gnp.garray(mean_matrix)
frame_number = final_layer_output.shape[0]
final_layer_output = final_layer_output.T
obs_mat = gnp.zeros((60, frame_number*3))
traj_err_mat = gnp.zeros((60, frame_number))
var_base = np.zeros((60, 3))
static_indice = []
delta_indice = []
acc_indice = []
for i in xrange(60):
static_indice.append(i)
delta_indice.append(i+60)
acc_indice.append(i+120)
obs_mat[:, 0:frame_number] = final_layer_output[static_indice, :]
obs_mat[:, frame_number:frame_number*2] = final_layer_output[delta_indice, :]
obs_mat[:, frame_number*2:frame_number*3] = final_layer_output[acc_indice, :]
var_base[:, 0] = std_matrix[0, static_indice].T
var_base[:, 1] = std_matrix[0, delta_indice].T
var_base[:, 2] = std_matrix[0, acc_indice].T
var_base = np.reshape(var_base, (60*3, 1))
var_base = var_base ** 2
sub_dim_list = []
for i in xrange(60):
sub_dim_list.append(1)
sub_dim_start = 0
for sub_dim in sub_dim_list:
wuw_mat, wu_mat = self.pre_wuw_wu(frame_number, sub_dim, var_base[sub_dim_start*3:sub_dim_start*3+sub_dim*3])
obs_mu = obs_mat[sub_dim_start:sub_dim_start+sub_dim, :].reshape((frame_number*3*sub_dim, 1))
wuwwu = gnp.dot(wuw_mat, wu_mat)
mlpg_traj = gnp.dot(wuwwu, obs_mu)
sub_std_mat = std_matrix[:, static_indice].T
sub_mu_mat = mean_matrix[:, static_indice].T
sub_std_mat = sub_std_mat[sub_dim_start:sub_dim_start+sub_dim, :].reshape((frame_number*sub_dim, 1))
sub_mu_mat = sub_mu_mat[sub_dim_start:sub_dim_start+sub_dim, :].reshape((frame_number*sub_dim, 1))
ref_y = test_set_y[:, static_indice].T
ref_y = ref_y[sub_dim_start:sub_dim_start+sub_dim, :].reshape((frame_number*sub_dim, 1))
ref_y = ref_y * sub_std_mat + sub_mu_mat
traj_err = (mlpg_traj - ref_y) #mlpg_traj ref_y
traj_err_mat[sub_dim_start:sub_dim_start+sub_dim, :] = traj_err.reshape((sub_dim, frame_number))
sub_dim_start = sub_dim_start + sub_dim
validation_losses = gnp.sum(traj_err_mat[1:60, :].T ** 2, axis=1)
validation_losses = validation_losses ** 0.5
return validation_losses.as_numpy_array()
def activate(self, net_input):
return gnp.tanh(net_input)
def forward(self, A):
return gnp.tanh(A)