def loss_(self, X_, Y_):
Y_hat_ = self.transform_(X_)
return mse(Y_hat_, Y_)
def train(self,
sess,
X,
Y,
R,
S,
scale=1.0,
scale_width=4.0,
n_scales=5,
alpha_pair=.1,
pre_train_steps=500,
pre_train_optimizer=None,
steps=10000,
paired_minibatch_size=20,
unpaired_minibatch_size=200,
optimizer=tf.train.RMSPropOptimizer(learning_rate=.001)):
X_ = tf.placeholder(tf.float32, (None, self.sizes[0]))
Y_ = tf.placeholder(tf.float32, (None, self.sizes[-1]))
R_ = tf.placeholder(tf.float32, (None, self.sizes[0]))
S_ = tf.placeholder(tf.float32, (None, self.sizes[-1]))
Y_hat_ = self.transform_(X_)
S_hat_ = self.transform_(R_)
mse_ = mse(Y_hat_, Y_)
if pre_train_optimizer is None:
pre_train_optimizer = optimizer
train_mse_ = self.train_(mse_, pre_train_optimizer)
if n_scales > 1:
scales = list(scale * (10**np.linspace(
-scale_width / 2, scale_width / 2, num=n_scales)))
weigths = [1] * len(scales)
kernel = lambda x, y: multiscale_gaussian(x, y, scales, weigths)
else:
kernel = lambda x, y: gaussian(x, y, scale)
mmd_ = mmd_squared(S_hat_, S_, kernel)
loss_ = alpha_pair * mse_ + (1 - alpha_pair) * mmd_
train_ = self.train_(loss_, optimizer)
sess.run(tf.global_variables_initializer())
# Fit all training data
losses = []
for step in range(pre_train_steps):
XY_indicies = self._gen_paired_minibatch(paired_minibatch_size,
step, X.shape[0])
batch_X = X[XY_indicies]
batch_Y = Y[XY_indicies]
_, loss_val = sess.run([train_mse_, mse_],
feed_dict={
X_: batch_X,
Y_: batch_Y
})
if step % 100 == 0:
mse_val = sess.run(mse_, feed_dict={X_: batch_X, Y_: batch_Y})
print("Pre-Training Step {} of {}, loss {}, mse {}".format(
step, pre_train_steps, loss_val, mse_val))
for step in range(steps):
XY_indicies = self._gen_paired_minibatch(paired_minibatch_size,
step, X.shape[0])
R_indicies = self._gen_unpaired_minibatch(unpaired_minibatch_size,
R.shape[0])
S_indicies = self._gen_unpaired_minibatch(unpaired_minibatch_size,
S.shape[0])
batch_X = X[XY_indicies]
batch_Y = Y[XY_indicies]
batch_R = R[R_indicies]
batch_S = S[S_indicies]
_, loss_val = sess.run([train_, loss_],
feed_dict={
X_: batch_X,
Y_: batch_Y,
R_: batch_R,
S_: batch_S
})
losses.append(loss_val)
if step % 500 == 0 or step in {0, 1, 2, 5, 10}:
mmd_val, mse_val = sess.run([mmd_, mse_],
feed_dict={
X_: batch_X,
Y_: batch_Y,
R_: batch_R,
S_: batch_S
})
print("Training Step {} of {}, loss {}, mmd {}, mse {}".format(
step, steps, loss_val, mmd_val, mse_val))
return losses