def model_regression_dropout_var(filtros=32, dropout_rate=0.00, l2_rate=1e-8, activation='relu', n_filters=64, l2_reg=1e-12,
input_channel_num=1, out_ch=1, start_ch=32, inc_rate=2., depth=2,
batchnorm=False, maxpool=True, upconv=True, residual=False,disableDo=False):
# UNet: code from https://github.com/pietz/unet-keras
def _conv_block(m, dim, acti, bn, res, do=0):
n = Conv2D(dim, 3, padding='same',kernel_initializer='he_normal', activation=acti)(m)
n = MCDropout(do)(n)
n = Conv2D(dim, 3, padding='same',kernel_initializer='he_normal', activation=acti)(n)
n = MCDropout(do)(n)
return Concatenate()([m, n]) if res else n
def _level_block(m, dim, depth, inc, acti, do, bn, mp, up, res):
if depth > 0:
n = _conv_block(m, dim, acti, bn, res,do)
m = MaxPooling2D()(n) if mp else Conv2D(dim, 3, strides=2, padding='same')(n)
m = _level_block(m, int(inc * dim), depth - 1, inc, acti, do, bn, mp, up, res)
if up:
m = UpSampling2D()(m)
m = Conv2D(dim, 2, activation=acti, padding='same')(m)
else:
m = Conv2DTranspose(dim, 3, strides=2, activation=acti, padding='same')(m)
n = Concatenate()([n, m])
m = _conv_block(n, dim, acti, bn, res,do)
else:
m = _conv_block(m, dim, acti, bn, res, do)
return m
def _level_block2(m, dim, depth, inc, acti, do, bn, mp, up, res):
m = _conv_block(m, dim, acti, bn, res, do)
m = _conv_block(m, dim, acti, bn, res, do)
m = _conv_block(m, dim, acti, bn, res, do)
m = _conv_block(m, dim, acti, bn, res, do)
return m
input_tensor = Input(shape=(None, None, 1),name='input')
labels_err_tensor = Input(shape=(None,None, 1), name='label_err')
# Good old output
o1 = _level_block(input_tensor, 32, depth, inc_rate, activation, dropout_rate, batchnorm, maxpool, upconv, residual)
linear_output = Conv2D(1,1,activation="linear",name='linear_output')(o1)
# Data-dependent uncertainty outainty
o2 = _level_block2(input_tensor, 32, 0, inc_rate, activation, dropout_rate, batchnorm, maxpool, upconv, residual)
o2 = Conv2D(1,1, activation="linear")(o2)
# We had to force to start close to the solution because working in log has this problem
variance_output = Lambda(lambda x: -K.abs(x)-12,name='variance_output')(o2)
model = Model(inputs=[input_tensor, labels_err_tensor], outputs=[variance_output, linear_output])
model_prediction = Model(inputs=input_tensor, outputs=[variance_output, linear_output])
mse_var_ext = mse_var_wrapper(linear_output, labels_err_tensor)
mse_lin_ext = mse_lin_wrapper(variance_output, labels_err_tensor)
return model, model_prediction, mse_lin_ext, mse_var_ext
def custom_train_step(self, data):
"""
Custom training logic
:param data:
:return:
"""
data = data_adapter.expand_1d(data)
x, y, sample_weight = data_adapter.unpack_x_y_sample_weight(data)
with tf.GradientTape() as tape:
y_pred = self.keras_model(x, training=True)
loss = self.keras_model.compiled_loss(
y,
y_pred,
sample_weight,
regularization_losses=self.keras_model.losses)
if self.task == 'regression':
variance_loss = mse_var_wrapper(y_pred[0], x['labels_err'])
output_loss = mse_lin_wrapper(y_pred[1], x['labels_err'])
elif self.task == 'classification':
output_loss = bayesian_categorical_crossentropy_wrapper(
y_pred[1])
variance_loss = bayesian_categorical_crossentropy_var_wrapper(
y_pred[0])
elif self.task == 'binary_classification':
output_loss = bayesian_binary_crossentropy_wrapper(y_pred[1])
variance_loss = bayesian_binary_crossentropy_var_wrapper(
y_pred[0])
else:
raise RuntimeError(
'Only "regression", "classification" and "binary_classification" are supported'
)
loss = output_loss(y['output'], y_pred[0]) + variance_loss(
y['variance_output'], y_pred[1])
# apply gradient here
if version.parse(tf.__version__) >= version.parse("2.4.0"):
self.keras_model.optimizer.minimize(
loss, self.keras_model.trainable_variables, tape=tape)
else:
tf.python.keras.engine.training._minimize(
self.keras_model.distribute_strategy, tape,
self.keras_model.optimizer, loss,
self.keras_model.trainable_variables)
self.keras_model.compiled_metrics.update_state(y, y_pred,
sample_weight)
return {m.name: m.result() for m in self.keras_model.metrics}
def model(self):
input_tensor = Input(shape=self._input_shape, name='input')
labels_err_tensor = Input(shape=(self._labels_shape,), name='labels_err')
cnn_layer_1 = Conv1D(kernel_initializer=self.initializer, padding="same", filters=self.num_filters[0],
kernel_size=self.filter_len, kernel_regularizer=regularizers.l2(self.l2))(input_tensor)
activation_1 = Activation(activation=self.activation)(cnn_layer_1)
dropout_1 = MCDropout(self.dropout_rate, disable=self.disable_dropout)(activation_1)
cnn_layer_2 = Conv1D(kernel_initializer=self.initializer, padding="same", filters=self.num_filters[1],
kernel_size=self.filter_len, kernel_regularizer=regularizers.l2(self.l2))(dropout_1)
activation_2 = Activation(activation=self.activation)(cnn_layer_2)
maxpool_1 = MaxPooling1D(pool_size=self.pool_length)(activation_2)
flattener = Flatten()(maxpool_1)
dropout_2 = MCDropout(self.dropout_rate, disable=self.disable_dropout)(flattener)
layer_3 = Dense(units=self.num_hidden[0], kernel_regularizer=regularizers.l2(self.l2),
kernel_initializer=self.initializer,
activation=self.activation)(dropout_2)
activation_3 = Activation(activation=self.activation)(layer_3)
dropout_3 = MCDropout(self.dropout_rate, disable=self.disable_dropout)(activation_3)
layer_4 = Dense(units=self.num_hidden[1], kernel_regularizer=regularizers.l2(self.l2),
kernel_initializer=self.initializer,
activation=self.activation)(dropout_3)
activation_4 = Activation(activation=self.activation)(layer_4)
output = Dense(units=self._labels_shape, name='output')(activation_4)
output_activated = Activation(activation=self._last_layer_activation)(output)
variance_output = Dense(units=self._labels_shape, activation='linear', name='variance_output')(activation_4)
model = Model(inputs=[input_tensor, labels_err_tensor], outputs=[output, variance_output])
# new astroNN high performance dropout variational inference on GPU expects single output
model_prediction = Model(inputs=[input_tensor], outputs=concatenate([output, variance_output]))
if self.task == 'regression':
variance_loss = mse_var_wrapper(output, labels_err_tensor)
output_loss = mse_lin_wrapper(variance_output, labels_err_tensor)
elif self.task == 'classification':
output_loss = bayesian_categorical_crossentropy_wrapper(variance_output)
variance_loss = bayesian_categorical_crossentropy_var_wrapper(output)
elif self.task == 'binary_classification':
output_loss = bayesian_binary_crossentropy_wrapper(variance_output)
variance_loss = bayesian_binary_crossentropy_var_wrapper(output)
else:
raise RuntimeError('Only "regression", "classification" and "binary_classification" are supported')
return model, model_prediction, output_loss, variance_loss
def model(self):
input_tensor = Input(shape=self._input_shape, name='input')
input_tensor_flattened = Flatten()(input_tensor)
labels_err_tensor = Input(shape=(self._labels_shape, ),
name='labels_err')
# slice spectra to censor out useless region for elements
censored_c_input = BoolMask(aspcap_mask("C", dr=14),
name='C_Mask')(input_tensor_flattened)
censored_c1_input = BoolMask(aspcap_mask("C1", dr=14),
name='C1_Mask')(input_tensor_flattened)
censored_n_input = BoolMask(aspcap_mask("N", dr=14),
name='N_Mask')(input_tensor_flattened)
censored_o_input = BoolMask(aspcap_mask("O", dr=14),
name='O_Mask')(input_tensor_flattened)
censored_na_input = BoolMask(aspcap_mask("Na", dr=14),
name='Na_Mask')(input_tensor_flattened)
censored_mg_input = BoolMask(aspcap_mask("Mg", dr=14),
name='Mg_Mask')(input_tensor_flattened)
censored_al_input = BoolMask(aspcap_mask("Al", dr=14),
name='Al_Mask')(input_tensor_flattened)
censored_si_input = BoolMask(aspcap_mask("Si", dr=14),
name='Si_Mask')(input_tensor_flattened)
censored_p_input = BoolMask(aspcap_mask("P", dr=14),
name='P_Mask')(input_tensor_flattened)
censored_s_input = BoolMask(aspcap_mask("S", dr=14),
name='S_Mask')(input_tensor_flattened)
censored_k_input = BoolMask(aspcap_mask("K", dr=14),
name='K_Mask')(input_tensor_flattened)
censored_ca_input = BoolMask(aspcap_mask("Ca", dr=14),
name='Ca_Mask')(input_tensor_flattened)
censored_ti_input = BoolMask(aspcap_mask("Ti", dr=14),
name='Ti_Mask')(input_tensor_flattened)
censored_ti2_input = BoolMask(aspcap_mask("Ti2", dr=14),
name='Ti2_Mask')(input_tensor_flattened)
censored_v_input = BoolMask(aspcap_mask("V", dr=14),
name='V_Mask')(input_tensor_flattened)
censored_cr_input = BoolMask(aspcap_mask("Cr", dr=14),
name='Cr_Mask')(input_tensor_flattened)
censored_mn_input = BoolMask(aspcap_mask("Mn", dr=14),
name='Mn_Mask')(input_tensor_flattened)
censored_co_input = BoolMask(aspcap_mask("Co", dr=14),
name='Co_Mask')(input_tensor_flattened)
censored_ni_input = BoolMask(aspcap_mask("Ni", dr=14),
name='Ni_Mask')(input_tensor_flattened)
# get neurones from each elements from censored spectra
c_dense = MCDropout(self.dropout_rate, disable=self.disable_dropout)(
Dense(units=self.num_hidden[2] * 8,
kernel_initializer=self.initializer,
name='c_dense',
activation=self.activation,
kernel_regularizer=regularizers.l2(
self.l2))(censored_c_input))
c1_dense = MCDropout(self.dropout_rate, disable=self.disable_dropout)(
Dense(units=self.num_hidden[2],
kernel_initializer=self.initializer,
name='c1_dense',
activation=self.activation,
kernel_regularizer=regularizers.l2(
self.l2))(censored_c1_input))
n_dense = MCDropout(self.dropout_rate, disable=self.disable_dropout)(
Dense(units=self.num_hidden[2] * 8,
kernel_initializer=self.initializer,
name='n_dense',
activation=self.activation,
kernel_regularizer=regularizers.l2(
self.l2))(censored_n_input))
o_dense = MCDropout(self.dropout_rate, disable=self.disable_dropout)(
Dense(units=self.num_hidden[2],
kernel_initializer=self.initializer,
name='o_dense',
activation=self.activation,
kernel_regularizer=regularizers.l2(
self.l2))(censored_o_input))
na_dense = MCDropout(self.dropout_rate, disable=self.disable_dropout)(
Dense(units=self.num_hidden[2],
kernel_initializer=self.initializer,
name='na_dense',
activation=self.activation,
kernel_regularizer=regularizers.l2(
self.l2))(censored_na_input))
mg_dense = MCDropout(self.dropout_rate, disable=self.disable_dropout)(
Dense(units=self.num_hidden[2],
kernel_initializer=self.initializer,
name='mg_dense',
activation=self.activation,
kernel_regularizer=regularizers.l2(
self.l2))(censored_mg_input))
al_dense = MCDropout(self.dropout_rate, disable=self.disable_dropout)(
Dense(units=self.num_hidden[2],
kernel_initializer=self.initializer,
name='al_dense',
activation=self.activation,
kernel_regularizer=regularizers.l2(
self.l2))(censored_al_input))
si_dense = MCDropout(self.dropout_rate, disable=self.disable_dropout)(
Dense(units=self.num_hidden[2],
kernel_initializer=self.initializer,
name='si_dense',
activation=self.activation,
kernel_regularizer=regularizers.l2(
self.l2))(censored_si_input))
p_dense = MCDropout(self.dropout_rate, disable=self.disable_dropout)(
Dense(units=self.num_hidden[2],
kernel_initializer=self.initializer,
name='p_dense',
activation=self.activation,
kernel_regularizer=regularizers.l2(
self.l2))(censored_p_input))
s_dense = MCDropout(self.dropout_rate, disable=self.disable_dropout)(
Dense(units=self.num_hidden[2],
kernel_initializer=self.initializer,
name='s_dense',
activation=self.activation,
kernel_regularizer=regularizers.l2(
self.l2))(censored_s_input))
k_dense = MCDropout(self.dropout_rate, disable=self.disable_dropout)(
Dense(units=self.num_hidden[2],
kernel_initializer=self.initializer,
name='k_dense',
activation=self.activation,
kernel_regularizer=regularizers.l2(
self.l2))(censored_k_input))
ca_dense = MCDropout(self.dropout_rate, disable=self.disable_dropout)(
Dense(units=self.num_hidden[2],
kernel_initializer=self.initializer,
name='ca_dense',
activation=self.activation,
kernel_regularizer=regularizers.l2(
self.l2))(censored_ca_input))
ti_dense = MCDropout(self.dropout_rate, disable=self.disable_dropout)(
Dense(units=self.num_hidden[2],
kernel_initializer=self.initializer,
name='ti_dense',
activation=self.activation,
kernel_regularizer=regularizers.l2(
self.l2))(censored_ti_input))
ti2_dense = MCDropout(self.dropout_rate, disable=self.disable_dropout)(
Dense(units=self.num_hidden[2],
kernel_initializer=self.initializer,
name='ti2_dense',
activation=self.activation,
kernel_regularizer=regularizers.l2(
self.l2))(censored_ti2_input))
v_dense = MCDropout(self.dropout_rate, disable=self.disable_dropout)(
Dense(units=self.num_hidden[2],
kernel_initializer=self.initializer,
name='v_dense',
activation=self.activation,
kernel_regularizer=regularizers.l2(
self.l2))(censored_v_input))
cr_dense = MCDropout(self.dropout_rate, disable=self.disable_dropout)(
Dense(units=self.num_hidden[2],
kernel_initializer=self.initializer,
name='cr_dense',
activation=self.activation,
kernel_regularizer=regularizers.l2(
self.l2))(censored_cr_input))
mn_dense = MCDropout(self.dropout_rate, disable=self.disable_dropout)(
Dense(units=self.num_hidden[2],
kernel_initializer=self.initializer,
name='mn_dense',
activation=self.activation,
kernel_regularizer=regularizers.l2(
self.l2))(censored_mn_input))
co_dense = MCDropout(self.dropout_rate, disable=self.disable_dropout)(
Dense(units=self.num_hidden[2],
kernel_initializer=self.initializer,
name='co_dense',
activation=self.activation,
kernel_regularizer=regularizers.l2(
self.l2))(censored_co_input))
ni_dense = MCDropout(self.dropout_rate, disable=self.disable_dropout)(
Dense(units=self.num_hidden[2],
kernel_initializer=self.initializer,
name='ni_dense',
activation=self.activation,
kernel_regularizer=regularizers.l2(
self.l2))(censored_ni_input))
# get neurones from each elements from censored spectra
c_dense_2 = MCDropout(self.dropout_rate, disable=self.disable_dropout)(
Dense(units=self.num_hidden[3] * 4,
kernel_initializer=self.initializer,
activation=self.activation,
name='c_dense_2')(c_dense))
c1_dense_2 = MCDropout(self.dropout_rate,
disable=self.disable_dropout)(Dense(
units=self.num_hidden[3],
kernel_initializer=self.initializer,
activation=self.activation,
name='c1_dense_2')(c1_dense))
n_dense_2 = MCDropout(self.dropout_rate, disable=self.disable_dropout)(
Dense(units=self.num_hidden[3] * 4,
kernel_initializer=self.initializer,
activation=self.activation,
name='n_dense_2')(n_dense))
o_dense_2 = MCDropout(self.dropout_rate, disable=self.disable_dropout)(
Dense(units=self.num_hidden[3],
kernel_initializer=self.initializer,
activation=self.activation,
name='o_dense_2')(o_dense))
na_dense_2 = MCDropout(self.dropout_rate,
disable=self.disable_dropout)(Dense(
units=self.num_hidden[3],
kernel_initializer=self.initializer,
activation=self.activation,
name='na_dense_2')(na_dense))
mg_dense_2 = MCDropout(self.dropout_rate,
disable=self.disable_dropout)(Dense(
units=self.num_hidden[3],
kernel_initializer=self.initializer,
activation=self.activation,
name='mg_dense_2')(mg_dense))
al_dense_2 = MCDropout(self.dropout_rate,
disable=self.disable_dropout)(Dense(
units=self.num_hidden[3],
kernel_initializer=self.initializer,
activation=self.activation,
name='al_dense_2')(al_dense))
si_dense_2 = MCDropout(self.dropout_rate,
disable=self.disable_dropout)(Dense(
units=self.num_hidden[3],
kernel_initializer=self.initializer,
activation=self.activation,
name='si_dense_2')(si_dense))
p_dense_2 = MCDropout(self.dropout_rate, disable=self.disable_dropout)(
Dense(units=self.num_hidden[3],
kernel_initializer=self.initializer,
activation=self.activation,
name='p_dense_2')(p_dense))
s_dense_2 = MCDropout(self.dropout_rate, disable=self.disable_dropout)(
Dense(units=self.num_hidden[3],
kernel_initializer=self.initializer,
activation=self.activation,
name='s_dense_2')(s_dense))
k_dense_2 = MCDropout(self.dropout_rate, disable=self.disable_dropout)(
Dense(units=self.num_hidden[3],
kernel_initializer=self.initializer,
activation=self.activation,
name='k_dense_2')(k_dense))
ca_dense_2 = MCDropout(self.dropout_rate,
disable=self.disable_dropout)(Dense(
units=self.num_hidden[3],
kernel_initializer=self.initializer,
activation=self.activation,
name='ca_dense_2')(ca_dense))
ti_dense_2 = MCDropout(self.dropout_rate,
disable=self.disable_dropout)(Dense(
units=self.num_hidden[3],
kernel_initializer=self.initializer,
activation=self.activation,
name='ti_dense_2')(ti_dense))
ti2_dense_2 = MCDropout(self.dropout_rate,
disable=self.disable_dropout)(Dense(
units=self.num_hidden[3],
kernel_initializer=self.initializer,
activation=self.activation,
name='ti2_dense_2')(ti2_dense))
v_dense_2 = MCDropout(self.dropout_rate, disable=self.disable_dropout)(
Dense(units=self.num_hidden[3],
kernel_initializer=self.initializer,
activation=self.activation,
name='v_dense_2')(v_dense))
cr_dense_2 = MCDropout(self.dropout_rate,
disable=self.disable_dropout)(Dense(
units=self.num_hidden[3],
kernel_initializer=self.initializer,
activation=self.activation,
name='cr_dense_2')(cr_dense))
mn_dense_2 = MCDropout(self.dropout_rate,
disable=self.disable_dropout)(Dense(
units=self.num_hidden[3],
kernel_initializer=self.initializer,
activation=self.activation,
name='mn_dense_2')(mn_dense))
co_dense_2 = MCDropout(self.dropout_rate,
disable=self.disable_dropout)(Dense(
units=self.num_hidden[3],
kernel_initializer=self.initializer,
activation=self.activation,
name='co_dense_2')(co_dense))
ni_dense_2 = MCDropout(self.dropout_rate,
disable=self.disable_dropout)(Dense(
units=self.num_hidden[3],
kernel_initializer=self.initializer,
activation=self.activation,
name='ni_dense_2')(ni_dense))
# Basically the same as ApogeeBCNN structure
cnn_layer_1 = Conv1D(kernel_initializer=self.initializer,
padding="same",
filters=self.num_filters[0],
kernel_size=self.filter_len,
kernel_regularizer=regularizers.l2(
self.l2))(input_tensor)
activation_1 = Activation(activation=self.activation)(cnn_layer_1)
dropout_1 = MCDropout(self.dropout_rate,
disable=self.disable_dropout)(activation_1)
cnn_layer_2 = Conv1D(kernel_initializer=self.initializer,
padding="same",
filters=self.num_filters[1],
kernel_size=self.filter_len,
kernel_regularizer=regularizers.l2(
self.l2))(dropout_1)
activation_2 = Activation(activation=self.activation)(cnn_layer_2)
maxpool_1 = MaxPooling1D(pool_size=self.pool_length)(activation_2)
flattener = Flatten()(maxpool_1)
dropout_2 = MCDropout(self.dropout_rate,
disable=self.disable_dropout)(flattener)
layer_3 = Dense(units=self.num_hidden[0],
kernel_regularizer=regularizers.l2(self.l2),
kernel_initializer=self.initializer)(dropout_2)
activation_3 = Activation(activation=self.activation)(layer_3)
dropout_3 = MCDropout(self.dropout_rate,
disable=self.disable_dropout)(activation_3)
layer_4 = Dense(units=self.num_hidden[1],
kernel_regularizer=regularizers.l2(self.l2),
kernel_initializer=self.initializer)(dropout_3)
activation_4 = Activation(activation=self.activation)(layer_4)
teff_output = Dense(units=1)(activation_4)
logg_output = Dense(units=1)(activation_4)
fe_output = Dense(units=1)(activation_4)
old_3_output_wo_grad = StopGrad()(concatenate(
[teff_output, logg_output, fe_output]))
teff_output_var = Dense(units=1)(activation_4)
logg_output_var = Dense(units=1)(activation_4)
fe_output_var = Dense(units=1)(activation_4)
aux_fullspec = Dense(units=self.num_hidden[4],
kernel_initializer=self.initializer,
kernel_constraint=MaxNorm(self.maxnorm),
name='aux_fullspec')(activation_4)
fullspec_hidden = concatenate([aux_fullspec, old_3_output_wo_grad])
# get the final answer
c_concat = Dense(units=1, name='c_concat')(concatenate(
[c_dense_2, fullspec_hidden]))
c1_concat = Dense(units=1, name='c1_concat')(concatenate(
[c1_dense_2, fullspec_hidden]))
n_concat = Dense(units=1, name='n_concat')(concatenate(
[n_dense_2, fullspec_hidden]))
o_concat = Dense(units=1, name='o_concat')(concatenate(
[o_dense_2, fullspec_hidden]))
na_concat = Dense(units=1, name='na_concat')(concatenate(
[na_dense_2, fullspec_hidden]))
mg_concat = Dense(units=1, name='mg_concat')(concatenate(
[mg_dense_2, fullspec_hidden]))
al_concat = Dense(units=1, name='al_concat')(concatenate(
[al_dense_2, fullspec_hidden]))
si_concat = Dense(units=1, name='si_concat')(concatenate(
[si_dense_2, fullspec_hidden]))
p_concat = Dense(units=1, name='p_concat')(concatenate(
[p_dense_2, fullspec_hidden]))
s_concat = Dense(units=1, name='s_concat')(concatenate(
[s_dense_2, fullspec_hidden]))
k_concat = Dense(units=1, name='k_concat')(concatenate(
[k_dense_2, fullspec_hidden]))
ca_concat = Dense(units=1, name='ca_concat')(concatenate(
[ca_dense_2, fullspec_hidden]))
ti_concat = Dense(units=1, name='ti_concat')(concatenate(
[ti_dense_2, fullspec_hidden]))
ti2_concat = Dense(units=1, name='ti2_concat')(concatenate(
[ti2_dense_2, fullspec_hidden]))
v_concat = Dense(units=1, name='v_concat')(concatenate(
[v_dense_2, fullspec_hidden]))
cr_concat = Dense(units=1, name='cr_concat')(concatenate(
[cr_dense_2, fullspec_hidden]))
mn_concat = Dense(units=1, name='mn_concat')(concatenate(
[mn_dense_2, fullspec_hidden]))
co_concat = Dense(units=1, name='co_concat')(concatenate(
[co_dense_2, fullspec_hidden]))
ni_concat = Dense(units=1, name='ni_concat')(concatenate(
[ni_dense_2, fullspec_hidden]))
# get the final predictive uncertainty
c_concat_var = Dense(units=1, name='c_concat_var')(concatenate(
[c_dense_2, fullspec_hidden]))
c1_concat_var = Dense(units=1, name='c1_concat_var')(concatenate(
[c1_dense_2, fullspec_hidden]))
n_concat_var = Dense(units=1, name='n_concat_var')(concatenate(
[n_dense_2, fullspec_hidden]))
o_concat_var = Dense(units=1, name='o_concat_var')(concatenate(
[o_dense_2, fullspec_hidden]))
na_concat_var = Dense(units=1, name='na_concat_var')(concatenate(
[na_dense_2, fullspec_hidden]))
mg_concat_var = Dense(units=1, name='mg_concat_var')(concatenate(
[mg_dense_2, fullspec_hidden]))
al_concat_var = Dense(units=1, name='al_concat_var')(concatenate(
[al_dense_2, fullspec_hidden]))
si_concat_var = Dense(units=1, name='si_concat_var')(concatenate(
[si_dense_2, fullspec_hidden]))
p_concat_var = Dense(units=1, name='p_concat_var')(concatenate(
[p_dense_2, fullspec_hidden]))
s_concat_var = Dense(units=1, name='s_concat_var')(concatenate(
[s_dense_2, fullspec_hidden]))
k_concat_var = Dense(units=1, name='k_concat_var')(concatenate(
[k_dense_2, fullspec_hidden]))
ca_concat_var = Dense(units=1, name='ca_concat_var')(concatenate(
[ca_dense_2, fullspec_hidden]))
ti_concat_var = Dense(units=1, name='ti_concat_var')(concatenate(
[ti_dense_2, fullspec_hidden]))
ti2_concat_var = Dense(units=1, name='ti2_concat_var')(concatenate(
[ti2_dense_2, fullspec_hidden]))
v_concat_var = Dense(units=1, name='v_concat_var')(concatenate(
[v_dense_2, fullspec_hidden]))
cr_concat_var = Dense(units=1, name='cr_concat_var')(concatenate(
[cr_dense_2, fullspec_hidden]))
mn_concat_var = Dense(units=1, name='mn_concat_var')(concatenate(
[mn_dense_2, fullspec_hidden]))
co_concat_var = Dense(units=1, name='co_concat_var')(concatenate(
[co_dense_2, fullspec_hidden]))
ni_concat_var = Dense(units=1, name='ni_concat_var')(concatenate(
[ni_dense_2, fullspec_hidden]))
# concatenate answer
output = concatenate([
teff_output, logg_output, c_concat, c1_concat, n_concat, o_concat,
na_concat, mg_concat, al_concat, si_concat, p_concat, s_concat,
k_concat, ca_concat, ti_concat, ti2_concat, v_concat, cr_concat,
mn_concat, fe_output, co_concat, ni_concat
],
name='output')
# concatenate predictive uncertainty
variance_output = concatenate([
teff_output_var, logg_output_var, c_concat_var, c1_concat_var,
n_concat_var, o_concat_var, na_concat_var, mg_concat_var,
al_concat_var, si_concat_var, p_concat_var, s_concat_var,
k_concat_var, ca_concat_var, ti_concat_var, ti2_concat_var,
v_concat_var, cr_concat_var, mn_concat_var, fe_output_var,
co_concat_var, ni_concat_var
],
name='variance_output')
model = Model(inputs=[input_tensor, labels_err_tensor],
outputs=[output, variance_output])
# new astroNN high performance dropout variational inference on GPU expects single output
model_prediction = Model(inputs=input_tensor,
outputs=concatenate([output,
variance_output]))
variance_loss = mse_var_wrapper(output, labels_err_tensor)
output_loss = mse_lin_wrapper(variance_output, labels_err_tensor)
return model, model_prediction, output_loss, variance_loss
def model(self):
input_tensor = Input(shape=self._input_shape,
name='input') # training data
labels_err_tensor = Input(shape=(self._labels_shape, ),
name='labels_err')
# extract spectra from input data and expand_dims for convolution
spectra = Lambda(lambda x: tf.expand_dims(x, axis=-1))(BoolMask(
self.specmask())(input_tensor))
# value to denorm magnitude
app_mag = BoolMask(self.magmask())(input_tensor)
# tf.convert_to_tensor(self.input_mean[self.magmask()])
denorm_mag = DeNormAdd(self.input_mean[self.magmask()])(app_mag)
inv_pow_mag = Lambda(lambda mag: tf.pow(10., tf.multiply(-0.2, mag)))(
denorm_mag)
# data to infer Gia DR2 offset
gaia_aux_data = BoolMask(self.gaia_aux_mask())(input_tensor)
gaia_aux_hidden = MCDropout(self.dropout_rate,
disable=self.disable_dropout)(Dense(
units=18,
kernel_regularizer=regularizers.l2(
self.l2),
kernel_initializer=self.initializer,
activation='tanh')(gaia_aux_data))
offset = Dense(units=1,
kernel_initializer=self.initializer,
activation='tanh',
name='offset_output')(gaia_aux_hidden)
# good old NN takes spectra and output fakemag
cnn_layer_1 = Conv1D(kernel_initializer=self.initializer,
padding="same",
filters=self.num_filters[0],
kernel_size=self.filter_len,
kernel_regularizer=regularizers.l2(
self.l2))(spectra)
activation_1 = Activation(activation=self.activation)(cnn_layer_1)
dropout_1 = MCDropout(self.dropout_rate,
disable=self.disable_dropout)(activation_1)
cnn_layer_2 = Conv1D(kernel_initializer=self.initializer,
padding="same",
filters=self.num_filters[1],
kernel_size=self.filter_len,
kernel_regularizer=regularizers.l2(
self.l2))(dropout_1)
activation_2 = Activation(activation=self.activation)(cnn_layer_2)
maxpool_1 = MaxPooling1D(pool_size=self.pool_length)(activation_2)
flattener = Flatten()(maxpool_1)
dropout_2 = MCDropout(self.dropout_rate,
disable=self.disable_dropout)(flattener)
layer_3 = Dense(units=self.num_hidden[0],
kernel_regularizer=regularizers.l2(self.l2),
kernel_initializer=self.initializer)(dropout_2)
activation_3 = Activation(activation=self.activation)(layer_3)
dropout_3 = MCDropout(self.dropout_rate,
disable=self.disable_dropout)(activation_3)
layer_4 = Dense(units=self.num_hidden[1],
kernel_regularizer=regularizers.l2(self.l2),
kernel_initializer=self.initializer)(dropout_3)
activation_4 = Activation(activation=self.activation)(layer_4)
fakemag_output = Dense(units=self._labels_shape,
activation='softplus',
name='fakemag_output')(activation_4)
fakemag_variance_output = Dense(
units=self._labels_shape,
activation='linear',
name='fakemag_variance_output')(activation_4)
# multiplt a pre-determined de-normalization factor, such that fakemag std approx. 1 for Sloan APOGEE population
_fakemag_denorm = Lambda(lambda x: tf.multiply(x, 68.))(fakemag_output)
_fakemag_var_denorm = Lambda(lambda x: tf.add(x, tf.log(68.)))(
fakemag_variance_output)
_fakemag_parallax = Multiply()([_fakemag_denorm, inv_pow_mag])
# output parallax
output = Add(name='output')([_fakemag_parallax, offset])
variance_output = Lambda(lambda x: tf.log(
tf.abs(tf.multiply(x[2], tf.divide(tf.exp(x[0]), x[1])))),
name='variance_output')([
fakemag_variance_output, fakemag_output,
_fakemag_parallax
])
model = Model(inputs=[input_tensor, labels_err_tensor],
outputs=[output, variance_output])
# new astroNN high performance dropout variational inference on GPU expects single output
# while training with parallax, we want testing output fakemag
model_prediction = Model(inputs=[input_tensor],
outputs=concatenate(
[_fakemag_denorm, _fakemag_var_denorm]))
variance_loss = mse_var_wrapper(output, labels_err_tensor)
output_loss = mse_lin_wrapper(variance_output, labels_err_tensor)
return model, model_prediction, output_loss, variance_loss
def compile(self,
optimizer=None,
loss=None,
metrics=None,
weighted_metrics=None,
loss_weights=None,
sample_weight_mode=None):
if optimizer is not None:
self.optimizer = optimizer
elif self.optimizer is None or self.optimizer == 'adam':
self.optimizer = Adam(learning_rate=self.lr,
beta_1=self.beta_1,
beta_2=self.beta_2,
epsilon=self.optimizer_epsilon,
decay=0.0)
if metrics is not None:
self.metrics = metrics
if self.task == 'regression':
if self._last_layer_activation is None:
self._last_layer_activation = 'linear'
elif self.task == 'classification':
if self._last_layer_activation is None:
self._last_layer_activation = 'softmax'
elif self.task == 'binary_classification':
if self._last_layer_activation is None:
self._last_layer_activation = 'sigmoid'
else:
raise RuntimeError(
'Only "regression", "classification" and "binary_classification" are supported'
)
self.keras_model, self.keras_model_predict, self.output_loss, self.variance_loss = self.model(
)
if self.task == 'regression':
self._output_loss = lambda predictive, labelerr: mse_lin_wrapper(
predictive, labelerr)
elif self.task == 'classification':
self._output_loss = lambda predictive, labelerr: bayesian_categorical_crossentropy_wrapper(
predictive)
elif self.task == 'binary_classification':
self._output_loss = lambda predictive, labelerr: bayesian_binary_crossentropy_wrapper(
predictive)
else:
raise RuntimeError(
'Only "regression", "classification" and "binary_classification" are supported'
)
# all zero losss as dummy lose
if self.task == 'regression':
self.metrics = [mean_absolute_error, mean_error
] if not self.metrics else self.metrics
self.keras_model.compile(optimizer=self.optimizer,
loss=zeros_loss,
metrics=self.metrics,
weighted_metrics=weighted_metrics,
sample_weight_mode=sample_weight_mode)
elif self.task == 'classification':
self.metrics = [categorical_accuracy
] if not self.metrics else self.metrics
self.keras_model.compile(optimizer=self.optimizer,
loss=zeros_loss,
metrics={'output': self.metrics},
weighted_metrics=weighted_metrics,
sample_weight_mode=sample_weight_mode)
elif self.task == 'binary_classification':
self.metrics = [binary_accuracy
] if not self.metrics else self.metrics
self.keras_model.compile(optimizer=self.optimizer,
loss=zeros_loss,
metrics={'output': self.metrics},
weighted_metrics=weighted_metrics,
sample_weight_mode=sample_weight_mode)
# inject custom training step if needed
try:
self.custom_train_step()
except NotImplementedError:
pass
except TypeError:
self.keras_model.train_step = self.custom_train_step
# inject custom testing step if needed
try:
self.custom_test_step()
except NotImplementedError:
pass
except TypeError:
self.keras_model.test_step = self.custom_test_step
return None
