def __call__(self, y_true, y_pred):
result = None
if self.loss_type == LossType.mae:
result = objectives.mean_absolute_error(y_true, y_pred)
elif self.loss_type == LossType.mse:
result = objectives.mean_squared_error(y_true, y_pred)
elif self.loss_type == LossType.rmse:
result = K.sqrt(objectives.mean_squared_error(y_true, y_pred))
elif self.loss_type == LossType.variance:
result = K.sqrt(objectives.mean_squared_error(
y_true, y_pred)) - objectives.mean_absolute_error(
y_true, y_pred)
elif self.loss_type == LossType.weighted_mae_mse:
loss1 = objectives.mean_absolute_error(y_true, y_pred)
loss2 = objectives.mean_squared_error(y_true, y_pred)
result = self.loss_ratio * loss1 + (1.0 - self.loss_ratio) * loss2
elif self.loss_type == LossType.weighted_mae_rmse:
loss1 = objectives.mean_absolute_error(y_true, y_pred)
loss2 = K.sqrt(objectives.mean_squared_error(y_true, y_pred))
result = self.loss_ratio * loss1 + (1.0 - self.loss_ratio) * loss2
elif self.loss_type == LossType.binary_crossentropy:
result = objectives.binary_crossentropy(y_true, y_pred)
elif self.loss_type == LossType.weighted_tanhmse_mse:
loss1 = losses.mean_squared_error(
K.tanh(self.data_input_scale * y_true),
K.tanh(self.data_input_scale * y_pred))
loss2 = losses.mean_squared_error(y_true, y_pred)
result = self.loss_ratio * loss1 + (1.0 - self.loss_ratio) * loss2
else:
assert False, ("Loss function not supported")
return result
def tailored_loss(true_output, loss_input):
flux_loss = objectives.mean_squared_error(
loss_input[:, 0, :, :] * true_output[:, 2, :, :], true_output[:,
0, :, :])
flux_err_loss = objectives.mean_squared_error(
loss_input[:, 1, :, :] * true_output[:, 2, :, :], true_output[:,
1, :, :])
return 0.5 * flux_loss + 0.5 * flux_err_loss
def feature_matching(self, out_true, out_pred):
"Feature matching objective function for use in G."
mse = mean_squared_error(out_true, out_pred)
activations = K.function([self.D.layers[0].input, K.learning_phase()],
[self.D.layers[1].output,])
# Inputs to theano functions must be numeric not tensors like out_true
pred_batch = self.D.predict(self.batch, batch_size=self.batch_size)
pred_activ = activations([pred_batch, 0])[0]
true_activ = activations([self.batch, 0])[0]
feat_match = mean_squared_error(true_activ, pred_activ)
return mse + self.gamma * feat_match
def loss_function_multiple_distance_and_area(y_true, y_pred):
# 点損失
error = mean_squared_error(y_true, y_pred)
# 線の損失
for i in range(8):
error += mean_squared_error(y_true[:, ((i+1)*2):18]-y_true[:, 0:(16-i*2)], y_pred[:, ((i+1)*2):18]-y_pred[:, 0:(16-i*2)])
# 右耳と左耳のエリア
left_ear_true, left_ear_pred = y_true[:, 6:12], y_pred[:, 6:12]
right_ear_true, right_ear_pred = y_true[:, 12:18], y_pred[:, 12:18]
error += calc_area_loss(left_ear_true, left_ear_pred)
error += calc_area_loss(right_ear_true, right_ear_pred)
return error
def feature_matching(self, out_true, out_pred):
"Feature matching objective function for use in G."
mse = mean_squared_error(out_true, out_pred)
activations = K.function([self.D.layers[0].input,
K.learning_phase()], [
self.D.layers[1].output,
])
# Inputs to theano functions must be numeric not tensors like out_true
pred_batch = self.D.predict(self.batch, batch_size=self.batch_size)
pred_activ = activations([pred_batch, 0])[0]
true_activ = activations([self.batch, 0])[0]
feat_match = mean_squared_error(true_activ, pred_activ)
return mse + self.gamma * feat_match
def vae_loss(x, x_decoded_mean):
# NOTE: binary_crossentropy expects a batch_size by dim for x and x_decoded_mean, so we MUST flatten these!
x = K.flatten(x)
x_decoded_mean = K.flatten(x_decoded_mean)
xent_loss = objectives.mean_squared_error(x, x_decoded_mean)
kl_loss = - 0.5 * K.mean(1 + z_log_var - K.square(z_mean) - K.exp(z_log_var), axis=-1)
return xent_loss + kl_loss
def __init_model(self):
'''
Initializes model, does not load weights.
'''
## get args
data_dim = self.data_dim
label_dim = self.label_dim
latent_dim = self.latent_dim
n_hidden = self.n_hidden
batch_size = self.batch_size
kl_coef = self.kl_coef
## encoder
x = Input(shape=(data_dim,))
condition = Input(shape=(label_dim,))
inputs = concatenate([x, condition])
x_encoded = Dense(n_hidden, activation='relu')(inputs)
x_encoded = Dense(n_hidden//2, activation='relu')(x_encoded)
x_encoded = Dense(n_hidden//4, activation='relu')(x_encoded)
mu = Dense(latent_dim, activation='linear')(x_encoded)
log_var = Dense(latent_dim, activation='linear')(x_encoded)
## latent sampler
def sampling(args):
mu, log_var = args
eps = K.random_normal(shape=(batch_size, latent_dim), mean=0., stddev=1.)
return mu + K.exp(log_var/2.) * eps
## sample
z = Lambda(sampling, output_shape=(latent_dim,))([mu, log_var])
z_cond = concatenate([z, condition])
## decoder
z_decoder1 = Dense(n_hidden//4, activation='relu')
z_decoder2 = Dense(n_hidden//2, activation='relu')
z_decoder3 = Dense(n_hidden, activation='relu')
y_decoder = Dense(data_dim, activation='linear')
z_decoded = z_decoder1(z_cond)
z_decoded = z_decoder2(z_decoded)
z_decoded = z_decoder3(z_decoded)
y = y_decoder(z_decoded)
## loss
reconstruction_loss = objectives.mean_squared_error(x, y)
kl_loss = .5 * K.mean(K.square(mu) + K.exp(log_var) - log_var - 1, axis=-1)
cvae_loss = reconstruction_loss + kl_coef * kl_loss
## define full model
cvae = Model([x, condition], y)
cvae.add_loss(cvae_loss)
cvae.compile(optimizer='adam')
cvae.summary()
self.cvae = cvae
## define encoder model
encoder = Model([x, condition], mu)
self.encoder = encoder
## define decoder model
decoder_input = Input(shape=(latent_dim + label_dim,))
_z_decoded = z_decoder1(decoder_input)
_z_decoded = z_decoder2(_z_decoded)
_z_decoded = z_decoder3(_z_decoded)
_y = y_decoder(_z_decoded)
generator = Model(decoder_input, _y)
generator.summary()
self.generator = generator
pass
def vae_loss(x, x_decoded_mean):
Z = T.transpose(K.repeat(z, n_centroid), [0, 2, 1])
z_mean_t = T.transpose(K.repeat(z_mean, n_centroid), [0, 2, 1])
z_log_var_t = T.transpose(K.repeat(z_log_var, n_centroid), [0, 2, 1])
u_tensor3 = T.repeat(u_p.dimshuffle('x', 0, 1), batch_size, axis=0)
lambda_tensor3 = T.repeat(lambda_p.dimshuffle('x', 0, 1),
batch_size,
axis=0)
theta_tensor3 = theta_p.dimshuffle('x', 'x', 0) * T.ones(
(batch_size, latent_dim, n_centroid))
p_c_z=K.exp(K.sum((K.log(theta_tensor3)-0.5*K.log(2*math.pi*lambda_tensor3)-\
K.square(Z-u_tensor3)/(2*lambda_tensor3)),axis=1))+1e-10
gamma = p_c_z / K.sum(p_c_z, axis=-1, keepdims=True)
gamma_t = K.repeat(gamma, latent_dim)
if datatype == 'sigmoid':
loss=alpha*original_dim * objectives.binary_crossentropy(x, x_decoded_mean)\
+K.sum(0.5*gamma_t*(latent_dim*K.log(math.pi*2)+K.log(lambda_tensor3)+K.exp(z_log_var_t)/lambda_tensor3+K.square(z_mean_t-u_tensor3)/lambda_tensor3),axis=(1,2))\
-0.5*K.sum(z_log_var+1,axis=-1)\
-K.sum(K.log(K.repeat_elements(theta_p.dimshuffle('x',0),batch_size,0))*gamma,axis=-1)\
+K.sum(K.log(gamma)*gamma,axis=-1)
else:
loss=alpha*original_dim * objectives.mean_squared_error(x, x_decoded_mean)\
+K.sum(0.5*gamma_t*(latent_dim*K.log(math.pi*2)+K.log(lambda_tensor3)+K.exp(z_log_var_t)/lambda_tensor3+K.square(z_mean_t-u_tensor3)/lambda_tensor3),axis=(1,2))\
-0.5*K.sum(z_log_var+1,axis=-1)\
-K.sum(K.log(K.repeat_elements(theta_p.dimshuffle('x',0),batch_size,0))*gamma,axis=-1)\
+K.sum(K.log(gamma)*gamma,axis=-1)
return loss
def vae_loss(x, x_decoded_mean):
x_d = Flatten()(x)
x_dec_d = Flatten()(x_decoded_mean)
xent_loss = input_dim * objectives.mean_squared_error(x_d, x_dec_d)
kl_loss = -0.5 * K.sum(1 + z_log_var - K.square(z_mean) - K.exp(z_log_var),
axis=-1)
return xent_loss + kl_loss
def vae_loss(x, x_decoded_mean):
z_mean_t = K.permute_dimensions(K.repeat(z_mean, n_centroid),[0,2,1])
z_log_var_t = K.permute_dimensions(K.repeat(z_log_var, n_centroid),[0,2,1])
u_p = vade.get_layer('latent').u_p
theta_p = vade.get_layer('latent').theta_p
lambda_p = vade.get_layer('latent').lambda_p
u_tensor3 = K.repeat_elements(K.expand_dims(u_p, 0), batch_size,axis=0)
lambda_tensor3 = K.repeat_elements(K.expand_dims(lambda_p, 0), batch_size,axis=0)
gamma = vade.get_layer('latent').gamma
gamma_t = K.repeat(gamma, latent_dim)
# We are trying to minimize the ELBO, i.e. maximizing its opposite
if datatype == 'sigmoid':
loss = alpha*original_dim * objectives.binary_crossentropy(x, x_decoded_mean)\
+ K.sum(0.5*gamma_t*(latent_dim*K.log(math.pi*2)+K.log(lambda_tensor3)+K.exp(z_log_var_t)/lambda_tensor3 + K.square(z_mean_t-u_tensor3)/lambda_tensor3), axis=(1, 2))\
- 0.5*K.sum(z_log_var+1,axis=-1)\
- K.sum(K.log(K.repeat_elements(K.expand_dims(theta_p, 0),batch_size,0))*gamma,axis=-1)\
+ K.sum(K.log(gamma)*gamma,axis=-1)
else:
loss=alpha*original_dim * objectives.mean_squared_error(x, x_decoded_mean)\
+ K.sum(0.5*gamma_t*(latent_dim*K.log(math.pi*2)+K.log(lambda_tensor3)+K.exp(z_log_var_t)/lambda_tensor3+K.square(z_mean_t-u_tensor3)/lambda_tensor3),axis=(1,2))\
- 0.5*K.sum(z_log_var+1,axis=-1)\
- K.sum(K.log(K.repeat_elements(K.expand_dims(theta_p, 0),batch_size,0))*gamma,axis=-1)\
+ K.sum(K.log(gamma)*gamma,axis=-1)
return loss
def vae_loss(real_image, generated_image):
gen_loss = K.mean(
objectives.mean_squared_error(real_image, generated_image))
kl_loss = -0.5 * K.mean(
1 + std_vector - K.square(mean_vector) - K.exp(std_vector),
axis=-1)
# kl_loss = 0.5 * K.mean(K.square(std_vector) + K.square(mean_vector) - K.log(K.square(std_vector)) -1, axis=-1)
return gen_loss + kl_loss
def print_regression_model_summary(prefix, y_test, y_pred, parmsFromNormalization):
y_test = (y_test*parmsFromNormalization.std*parmsFromNormalization.sqrtx2) + parmsFromNormalization.mean
y_pred = (y_pred*parmsFromNormalization.std*parmsFromNormalization.sqrtx2) + parmsFromNormalization.mean
mse = mean_squared_error(y_test, y_pred)
error_AC, rmsep, mape, rmse = almost_correct_based_accuracy(y_test, y_pred, 10)
rmsle = calculate_rmsle(y_test, y_pred)
print ">> %s AC_errorRate=%.1f RMSEP=%.6f MAPE=%6f RMSE=%6f mse=%f rmsle=%.5f" %(prefix, error_AC, rmsep, mape, rmse, mse, rmsle)
log.write("%s AC_errorRate=%.1f RMSEP=%.6f MAPE=%6f RMSE=%6f mse=%f rmsle=%.5f" %(prefix, error_AC, rmsep, mape, rmse, mse, rmsle))
def rec_loss(self, x, y):
ec = int(self.edge_clip)
x_f = K.flatten(x[:, :, ec:-1 * ec, ec:-1 * ec, ec:-1 * ec])
y_f = K.flatten(y[:, :, ec:-1 * ec, ec:-1 * ec, ec:-1 * ec])
reconstruction_loss = objectives.mean_squared_error(
x_f, y_f) * 16**3 * 4**3 * self.rec_loss_factor
return reconstruction_loss
def vae_loss(x, x_decoded_mean):
x = K.flatten(x)
x_decoded_mean = K.flatten(x_decoded_mean)
xent_loss = SEQUENCE_LENGTH * objectives.mean_squared_error(
x, x_decoded_mean)
kl_loss = -0.5 * K.mean(
1 + z_log_var - K.square(z_mean) - K.exp(z_log_var), axis=-1)
return xent_loss + kl_loss
def vade_loss_function(self, args):
inputs = self.inputs
#reconst = self.x_decoded
reconst, z, z_mean, z_log_var = args[:self.num_views], args[-3], args[
-2], args[-1]
Z = T.transpose(K.repeat(z, self.n_centroid), [0, 2, 1])
z_mean_t = T.transpose(K.repeat(z_mean, self.n_centroid), [0, 2, 1])
z_log_var_t = T.transpose(K.repeat(z_log_var, self.n_centroid),
[0, 2, 1])
u_tensor3 = T.repeat(self.u_p.dimshuffle('x', 0, 1),
self.batch_size,
axis=0)
lambda_tensor3 = T.repeat(self.lambda_p.dimshuffle('x', 0, 1),
self.batch_size,
axis=0)
#version1
theta_tensor3 = self.theta_p.dimshuffle('x', 'x', 0) * T.ones(
(self.batch_size, self.latent_dim, self.n_centroid))
p_c_z = K.exp(K.sum((K.log(theta_tensor3) - 0.5 * K.log(2 * math.pi * lambda_tensor3) - \
K.square(Z - u_tensor3) / (2 * lambda_tensor3)), axis=1)) + 1e-10
gamma = p_c_z / K.sum(p_c_z, axis=-1, keepdims=True)
gamma_t = K.repeat(gamma, self.latent_dim)
#version2
# theta_tensor2 = self.theta_p.dimshuffle('x', 0) * T.ones((self.batch_size, self.n_centroid))
# p_c_z = K.exp(K.log(theta_tensor2) - K.sum((0.5 * K.log(2 * math.pi * lambda_tensor3) + \
# K.square(Z - u_tensor3) / (
# 2 * lambda_tensor3)), axis=1)) + 1e-10
# gamma = p_c_z / K.sum(p_c_z, axis=1, keepdims=True)
reconst_loss = 0
for i in range(0, num_views):
#version 1
r_loss = self.original_dim[i] * objectives.mean_squared_error(
inputs[i], reconst[i])
#version 2
#r_loss = self.original_dim[i]*objectives.binary_crossentropy(inputs[i], reconst[i])
reconst_loss += r_loss
#version 1
loss = reconst_loss + self.alpha * (K.sum(0.5 * gamma_t * (
self.latent_dim * K.log(math.pi * 2) + K.log(lambda_tensor3) + K.exp(
z_log_var_t) / lambda_tensor3 + K.square(z_mean_t - u_tensor3) / lambda_tensor3), axis=(1, 2)) \
- 0.5 * K.sum(z_log_var + 1, axis=-1) \
- K.sum(
K.log(K.repeat_elements(self.theta_p.dimshuffle('x', 0), self.batch_size, 0)) * gamma, axis=-1) \
+ K.sum(K.log(gamma) * gamma, axis=-1))
#version2
# loss = reconst_loss + self.alpha * (K.sum(0.5 * gamma * K.sum(
# K.log(math.pi * 2) + K.log(lambda_tensor3) + K.exp(
# z_log_var_t) / lambda_tensor3 + K.square(z_mean_t - u_tensor3) / lambda_tensor3, axis=1), axis=1) \
# - 0.5 * K.sum(z_log_var + 1 + K.log(2*math.pi), axis=1) \
# - K.sum(
# K.log(theta_tensor2)* gamma, axis=1) \
# + K.sum(K.log(gamma) * gamma, axis=1))
return loss
def loss_function_multiple_distance_and_triangle(y_true, y_pred):
# 点損失
error = mean_squared_error(y_true, y_pred)
# 線の損失
for i in range(8):
error += mean_squared_error(y_true[:, ((i+1)*2):18]-y_true[:, 0:(16-i*2)], y_pred[:, ((i+1)*2):18]-y_pred[:, 0:(16-i*2)])
# 面の損失
for comb in combinations(range(9), 3):
s_true = sarrus_formula(
y_true[:, (comb[0]*2):(comb[0]*2+2)],
y_true[:, (comb[1]*2):(comb[1]*2+2)],
y_true[:, (comb[2]*2):(comb[2]*2+2)]
)
s_pred = sarrus_formula(
y_pred[:, (comb[0]*2):(comb[0]*2+2)],
y_pred[:, (comb[1]*2):(comb[1]*2+2)],
y_pred[:, (comb[2]*2):(comb[2]*2+2)]
)
error += K.abs(s_true - s_pred)
return error
def vae_loss(x, x_decoded_mean):
x = K.flatten(x)
x_decoded_mean = K.flatten(x_decoded_mean)
rec_loss = objectives.mean_squared_error(x, x_decoded_mean)
kl_loss = -0.5 * K.mean(1 + self.z_logvar - K.square(self.z_mean) -
K.exp(self.z_logvar),
axis=-1)
return rec_loss + kl_loss
def cnn_loss(x, x_decoded_mean):
#N = tf.convert_to_tensor(DPParam, dtype=tf.float32)
gamma = tf.convert_to_tensor(DPParam['LPMtx'], dtype=tf.float32)
N = tf.convert_to_tensor(DPParam['Nvec'], dtype=tf.float32)
m = tf.convert_to_tensor(DPParam['m'], dtype=tf.float32)
W = tf.convert_to_tensor(DPParam['B'], dtype=tf.float32)
v = tf.convert_to_tensor(DPParam['nu'], dtype=tf.float32)
num_cluster = N.shape[0]
z_mean_1_last = tf.expand_dims(z_mean, -1) # bs, latent_dim, 1
z_mean_1_mid = tf.expand_dims(z_mean, 1) # bs, 1, latent_dim
for k in range(num_cluster):
gamma_k_rep = tf.squeeze(
K.repeat(tf.expand_dims(gamma[:, k], -1), latent_dim))
z_k_bar = 1 / N[k] * K.sum(tf.multiply(gamma_k_rep, z_mean),
axis=0) #(latent_dim, )
z_k_bar_batch = tf.squeeze(
K.repeat(tf.expand_dims(z_k_bar, 0), batch_size))
#tf.transpose(z_k_bar_batch, perm=[1, 0])
z_k_bar_batch_1_last = tf.expand_dims(z_k_bar_batch,
-1) # bs, latent_dim, 1
z_k_bar_batch_1_mid = tf.expand_dims(z_k_bar_batch,
1) # bs, 1, latent_dim
# TODO:!
S_k = 1 / N[k] * K.sum(K.batch_dot(
tf.multiply(tf.expand_dims(gamma_k_rep, -1),
(z_mean_1_last - z_k_bar_batch_1_last)),
z_mean_1_mid - z_k_bar_batch_1_mid),
axis=0) # (latent_dim, latent_dim)
temp = tf.linalg.trace(tf.linalg.solve(W[k], S_k))
temp2 = tf.matmul(tf.expand_dims((z_k_bar - m[k]), 0),
tf.linalg.inv(W[k]))
temp3 = tf.squeeze(
tf.matmul(temp2, tf.expand_dims((z_k_bar - m[k]), -1)))
if k == 0:
e = 0.5 * N[k] * (v[k] * (temp + temp3))
else:
e += 0.5 * N[k] * (v[k] * (temp + temp3))
loss_ = alpha * original_dim * objectives.mean_squared_error(
K.flatten(x), K.flatten(x_decoded_mean)) - scale * K.sum(
(z_log_var + 1), axis=-1)
loss_ = K.sum(loss_, axis=0) + e
# loss = K.sum(loss_, axis = 0)
#for i in range(5):
# loss_ += N
#return loss_
return loss_
def validate(self, x_train, y_train):
preds = self.predict(x_train)
answers = y_train
for ans, pred in zip(answers, preds):
row = "ans: \t{0: 3.3f}, predict: \t{1: 3.3f}, diff:\t{2: 3.3f}"
print(row.format(ans[0], pred[0], pred[0] - ans[0]))
ans_t = answers.reshape(answers.shape[1], answers.shape[0])
pred_t = preds.reshape(preds.shape[1], preds.shape[0])
mse = K.eval(mean_squared_error(ans_t, pred_t))
print("mean squared error is {0}".format(mse))
def gan_loss(y_true, y_pred):
#trade-off coefficient
alpha_recip = 0.05
y_true_flat = K.batch_flatten(y_true)
y_pred_flat = K.batch_flatten(y_pred)
#L_adv = objectives.binary_crossentropy(y_true_flat, y_pred_flat)
L_adv = objectives.mean_squared_error(y_true_flat, y_pred_flat)
L_seg = gen_dice_multilabel(labels, fake_labels)
return alpha_recip * L_adv + L_seg
def vae_loss(self, x, x_decoded_mean, args):
"""
We train with two loss functions. Reconstruction loss force decoded samples to match to X (just like
autoencoder). KL loss (latent loss) calculates the divergence between the learned latent distribution
derived from z_mean and z_logsigma and the original distribution of X.
"""
z_mean, z_logsigma = args
print('z_mean', z_mean)
print('logsig', z_logsigma)
reconstruction_loss = objectives.mean_squared_error(x, x_decoded_mean)
latent_loss = -0.50 * K.mean(
1 + z_logsigma - K.square(z_mean) - K.exp(z_logsigma), axis=-1)
print('RL ', reconstruction_loss)
print('LL ', latent_loss)
return K.mean(reconstruction_loss + latent_loss)
def vae_loss(x,
x_decoded_mean,
z,
z_mean,
z_log_var,
u_p,
theta_p,
lambda_p,
alpha=1,
datatype='sigmoid'):
Z = tf.transpose(K.repeat(z, n_centroid), [0, 2, 1])
z_mean_t = tf.transpose(K.repeat(z_mean, n_centroid), [0, 2, 1])
z_log_var_t = tf.transpose(K.repeat(z_log_var, n_centroid), [0, 2, 1])
u_tensor3 = tf.tile(tf.expand_dims(u_p, [0]), [batch_size, 1, 1])
# u_tensor3 = T.repeat(tf.expand_dims(u_p,[0]), batch_size, axis=0)
# lambda_tensor3 = T.repeat(tf.expand_dims(lambda_p,[0]), batch_size, axis=0)
lambda_tensor3 = tf.tile(tf.expand_dims(lambda_p, [0]), [batch_size, 1, 1])
temp_theta_p = tf.expand_dims(theta_p, [0])
temp_theta_p = tf.expand_dims(temp_theta_p, [0])
# theta_tensor3 = temp_theta_p * T.ones((batch_size, z_dim, n_centroid))
theta_tensor3 = tf.tile(temp_theta_p, [batch_size, z_dim, 1])
#@TODO
#PROBLEM HERE ? add theta z_dim times for each cluster?
p_c_z = K.exp(K.sum((K.log(theta_tensor3) - 0.5 * K.log(2 * math.pi * lambda_tensor3) - \
K.square(Z - u_tensor3) / (2 * lambda_tensor3)), axis=1)) + 1e-10
gamma = p_c_z / K.sum(p_c_z, axis=-1, keepdims=True)
gamma_t = K.repeat(gamma, z_dim)
if datatype == 'sigmoid':
loss = alpha * original_dim * objectives.binary_crossentropy(x, x_decoded_mean) \
+ K.sum(0.5 * gamma_t * (
z_dim * K.log(math.pi * 2) + K.log(lambda_tensor3) + K.exp(z_log_var_t) / lambda_tensor3 + K.square(
z_mean_t - u_tensor3) / lambda_tensor3), axis=(1, 2)) \
- 0.5 * K.sum(z_log_var + 1, axis=-1) \
- K.sum(K.log(K.repeat_elements(tf.expand_dims(theta_p, [0]), batch_size, 0)) * gamma, axis=-1) \
+ K.sum(K.log(gamma) * gamma, axis=-1)
else:
loss = alpha * original_dim * objectives.mean_squared_error(x, x_decoded_mean) \
+ K.sum(0.5 * gamma_t * (
z_dim * K.log(math.pi * 2) + K.log(lambda_tensor3) + K.exp(z_log_var_t) / lambda_tensor3 + K.square(
z_mean_t - u_tensor3) / lambda_tensor3), axis=(1, 2)) \
- 0.5 * K.sum(z_log_var + 1, axis=-1) \
- K.sum(K.log(K.repeat_elements(tf.expand_dims(theta_p, [0]), batch_size, 0)) * gamma, axis=-1) \
+ K.sum(K.log(gamma) * gamma, axis=-1)
return tf.reduce_mean(loss)
def print_regression_model_summary(prefix, y_test, y_pred,
parmsFromNormalization):
y_test = (y_test * parmsFromNormalization.std *
parmsFromNormalization.sqrtx2) + parmsFromNormalization.mean
y_pred = (y_pred * parmsFromNormalization.std *
parmsFromNormalization.sqrtx2) + parmsFromNormalization.mean
mse = mean_squared_error(y_test, y_pred)
error_AC, rmsep, mape, rmse = almost_correct_based_accuracy(
y_test, y_pred, 10)
rmsle = calculate_rmsle(y_test, y_pred)
print ">> %s AC_errorRate=%.1f RMSEP=%.6f MAPE=%6f RMSE=%6f mse=%f rmsle=%.5f" % (
prefix, error_AC, rmsep, mape, rmse, mse, rmsle)
log.write(
"%s AC_errorRate=%.1f RMSEP=%.6f MAPE=%6f RMSE=%6f mse=%f rmsle=%.5f" %
(prefix, error_AC, rmsep, mape, rmse, mse, rmsle))
def cvae_loss(self, x, y):
ec = int(self.edge_clip)
x_f = K.flatten(x[:, :, ec:-1 * ec, ec:-1 * ec, ec:-1 * ec])
y_f = K.flatten(y[:, :, ec:-1 * ec, ec:-1 * ec, ec:-1 * ec])
reconstruction_loss = objectives.mean_squared_error(
x_f, y_f) * 16**3 * 4**3 * self.rec_loss_factor
kl_loss = 0.5 * K.sum(
K.square(self.mu) + K.exp(self.log_var) - self.log_var - 1,
axis=-1)
# out_o_bounds = tf.reduce_sum(tf.cast(tf.logical_and(y[:,0,ec:-1*ec,ec:-1*ec,ec:-1*ec]>.2,y[:,-1,ec:-1*ec,ec:-1*ec,ec:-1*ec]<4.0),tf.float32))
rr = (y[:, 0, ec:-1 * ec, ec:-1 * ec, ec:-1 * ec] - 1.0 * np.ones(
(32, 44, 44, 44)))**2 + (
y[:, -1, ec:-1 * ec, ec:-1 * ec, ec:-1 * ec] - 3.5 * np.ones(
(32, 44, 44, 44)))**2
out_o_bounds = tf.reduce_sum(tf.math.exp(-1 * rr / 0.15) * 5)
loss = K.mean(reconstruction_loss + kl_loss)
return loss
def create_loss_optimizer(self):
z_mean_t = K.permute_dimensions(K.repeat(self.z_mean, self.n_clusters),
[0, 2, 1])
z_log_var_t = K.permute_dimensions(
K.repeat(self.z_log_var, self.n_clusters), [0, 2, 1])
u_tensor3 = K.repeat_elements(K.expand_dims(self.u_p, 0),
self.batch_size,
axis=0)
lambda_tensor3 = K.repeat_elements(K.expand_dims(self.lambda_p, 0),
self.batch_size,
axis=0)
gamma_t = K.repeat(self.gamma, self.dimensions[-1])
if self.datatype == 'binary':
self.loss = self.alpha * self.dimensions[0] * objectives.binary_crossentropy(self.x, self.output) \
+ K.sum(0.5 * gamma_t * (self.dimensions[-1] * K.log(math.pi * 2) + K.log(lambda_tensor3) + K.exp(z_log_var_t) / lambda_tensor3 + K.square(z_mean_t - u_tensor3) / lambda_tensor3), axis=(1, 2)) \
- 0.5 * K.sum(self.z_log_var + 1, axis=-1) \
- K.sum(K.log(K.repeat_elements(K.expand_dims(self.theta_p, 0), self.dimensions[0], 0)) * self.gamma, axis=-1) \
+ K.sum(K.log(self.gamma) * self.gamma, axis=-1)
else:
self.loss = self.alpha * self.dimensions[0] * objectives.mean_squared_error(self.x, self.output) \
+ K.sum(0.5 * gamma_t * (self.dimensions[-1] * K.log(math.pi * 2) + K.log(lambda_tensor3) + K.exp(z_log_var_t) / lambda_tensor3 + K.square(z_mean_t - u_tensor3) / lambda_tensor3), axis=(1, 2)) \
- 0.5 * K.sum(self.z_log_var + 1, axis=-1) \
- K.sum(K.log(K.repeat_elements(K.expand_dims(self.theta_p_norm, 0), self.batch_size, 0)) * self.gamma, axis=-1) \
+ K.sum(K.log(self.gamma) * self.gamma, axis=-1)
self.cost = tf.reduce_mean(self.loss)
self.optimizer = tf.train.AdamOptimizer(
learning_rate=self.learning_rate,
beta1=0.9,
beta2=0.999,
epsilon=5).minimize(self.cost)
self.normalize = tf.assign(self.theta_p_norm,
self.theta_p / tf.reduce_sum(self.theta_p))
print('Loss optimizer created.')
def c_loss(x_true_c, x_pred_c):
max_length_f = 1.0 * max_length
x_true_conn = x_true_c[:, :, conn_dim_start:]
x_pred_conn = x_pred_c[:, :, conn_dim_start:]
#variant 1
#x_true_conn = 0.5 * x_true_conn + 0.5
#x_pred_conn = 0.5 * K.round(x_pred_conn * max_length_f) / max_length_f + 0.5
#variant 2
#x_pred_conn = 0.5 * x_pred_conn + 0.5
#variant 3
x_true_conn = K.round(x_true_conn * max_length_f)
x_pred_conn = K.round(x_pred_conn * max_length_f)
#x_pred_conn.sort(axis=2)
#x_true_conn.sort(axis=2)
#x_pred_conn = K.round(x_pred_conn * max_length_f) / max_length_f
x_true_conn = K.flatten(x_true_conn)
x_pred_conn = K.flatten(x_pred_conn)
return objectives.mean_squared_error(x_true_conn,
x_pred_conn) / max_length_f
def loss_function_with_distance(y_true, y_pred):
point_mse = mean_squared_error(y_true, y_pred)
distance_mse = mean_squared_error(y_true[:, 2:18]-y_true[:, 0:16], y_pred[:, 2:18]-y_pred[:, 0:16])
return point_mse + distance_mse
def loss_function_simple(y_true, y_pred):
return mean_squared_error(y_true, y_pred)
if mode == "new":
print("saving model weights...")
model.save_weights(netfile, overwrite = True)
exit(0)
# are we prepared for this???
if len(sys.argv) < 4:
print(usage)
exit(0)
# prepare the model for learning
print("compiling model...")
#rms = SGD()
rms = RMSprop()
#model.compile(optimizer=rms, loss="mean_squared_error")
KObjs.allzero = lambda x,y: KObjs.mean_squared_error(x, y)\
* rnd.choice([-0.01, 0.01])
#model.compile(optimizer=rms, loss={"output": "mean_squared_error"})
model.compile(optimizer = rms,\
loss = {"output": "mean_squared_error", "memo-out": "mean_squared_error"})
# load weights if available
print("loading weights...")
model.load_weights(sys.argv[2])
audiofile = sys.argv[3]
if mode == "train":
print("loading audio input...")
data, sample_rate = soundfile.read(audiofile)
data = data[:,0] # extract the 0th channel
def objective(self, out_true, out_pred):
mse = mean_squared_error(out_true, out_pred)
return mse
def loss_function_with_multiple_distance(y_true, y_pred):
error = mean_squared_error(y_true, y_pred)
for i in range(8):
error += mean_squared_error(y_true[:, ((i+1)*2):18]-y_true[:, 0:(16-i*2)], y_pred[:, ((i+1)*2):18]-y_pred[:, 0:(16-i*2)])
return error
def mse_loss(x, x_decoded):
original_dim = np.float32(np.prod((IMG_DIM, IMG_DIM, 3)))
return K.mean(original_dim * mean_squared_error(x, x_decoded))
for i in range(nb_test_samples):
if(random.random() < 0.5):#50 % of the time make feature[2] = 1 & the output also = 1
test_features[i, 2] = 1
test_labels[i] = 1
test_ones = test_ones + 1
else:
test_features[i,2] = 0
print "Expect "+str(test_ones)+" ones in test set"
###Objective function:
#from keras.objectives import categorical_crossentropy
#loss = tf.reduce_mean(categorical_crossentropy(labels, output_layer))
from keras.objectives import mean_squared_error
loss = tf.reduce_mean(mean_squared_error(labels, output_layer))
train_step = tf.train.GradientDescentOptimizer(0.5).minimize(loss)
###training:
with sess.as_default():
for i in range(nb_train_samples): #Batching this might help also.
train_step.run(feed_dict={
features:train_features[i,:].reshape(1,sample_width),
labels: train_labels[i,:].reshape(1,1),
K.learning_phase(): 1
})
###evaluation:
correct_prediction = tf.equal(tf.round(output_layer), labels)
def poi_gau_mix(y_true, log_y_pred):
return log_poisson(y_true, log_y_pred) + 0.01 * mean_squared_error(
y_true, K.exp(log_y_pred))
def clipped_mse(y_true, y_pred):
y_true = K.clip(y_true, -1.0, 1.0)
y_pred = K.clip(y_pred, -1.0, 1.0)
return objectives.mean_squared_error(y_true, y_pred)
def mse_steer_angle(y_true, y_pred):
return mean_squared_error(y_true[0], y_pred[0])
def rmse(y_true, y_pred):
from keras import backend as k
from keras.objectives import mean_squared_error
return k.sqrt(mean_squared_error(y_true, y_pred))
