2martens/masterthesis
1
0
Fork 0
Code Issues Pull Requests Releases Wiki Activity

Switched to keras binary_crossentropy loss

Browse Source 
Signed-off-by: Jim Martens <github@2martens.de>
This commit is contained in:
Jim Martens
2019-02-08 05:42:54 +01:00
parent c52a3cdfb9
commit bb76a51d76
1 changed files with 8 additions and 7 deletions
Download Patch File Download Diff File Expand all files Collapse all files

15
src/twomartens/masterthesis/aae/train.py
View File

@ -31,6 +31,7 @@ from .util import save_image
k = tf.keras.backend k = tf.keras.backend
AdamOptimizer = tf.train.AdamOptimizer AdamOptimizer = tf.train.AdamOptimizer
tfe = tf.contrib.eager tfe = tf.contrib.eager
binary_crossentropy = tf.keras.losses.binary_crossentropy
def train_mnist(folding_id: int, inlier_classes: Sequence[int], total_classes: int, def train_mnist(folding_id: int, inlier_classes: Sequence[int], total_classes: int,
@ -138,13 +139,13 @@ def train_mnist(folding_id: int, inlier_classes: Sequence[int], total_classes: i
# x discriminator # x discriminator
with tf.GradientTape() as tape: with tf.GradientTape() as tape:
xd_result = tf.squeeze(x_discriminator(x)) xd_result = tf.squeeze(x_discriminator(x))
xd_real_loss = k.mean(k.binary_crossentropy(y_real, xd_result), axis=0) xd_real_loss = binary_crossentropy(y_real, xd_result)
z = k.reshape(k.random_normal((batch_size, zsize)), (-1, 1, 1, zsize)) z = k.reshape(k.random_normal((batch_size, zsize)), (-1, 1, 1, zsize))
z = k.variable(z) z = k.variable(z)
x_fake = decoder(z) x_fake = decoder(z)
xd_result = tf.squeeze(x_discriminator(x_fake)) xd_result = tf.squeeze(x_discriminator(x_fake))
xd_fake_loss = k.mean(k.binary_crossentropy(y_fake, xd_result), axis=0) xd_fake_loss = binary_crossentropy(y_fake, xd_result)
_xd_train_loss = xd_real_loss + xd_fake_loss _xd_train_loss = xd_real_loss + xd_fake_loss
@ -161,7 +162,7 @@ def train_mnist(folding_id: int, inlier_classes: Sequence[int], total_classes: i
x_fake = decoder(z) x_fake = decoder(z)
xd_result = tf.squeeze(x_discriminator(x_fake)) xd_result = tf.squeeze(x_discriminator(x_fake))
_decoder_train_loss = k.mean(k.binary_crossentropy(y_real, xd_result), axis=0) _decoder_train_loss = binary_crossentropy(y_real, xd_result)
decoder_grads = tape.gradient(_decoder_train_loss, decoder.trainable_variables) decoder_grads = tape.gradient(_decoder_train_loss, decoder.trainable_variables)
decoder_optimizer.apply_gradients(zip(decoder_grads, decoder.trainable_variables), decoder_optimizer.apply_gradients(zip(decoder_grads, decoder.trainable_variables),
@ -175,11 +176,11 @@ def train_mnist(folding_id: int, inlier_classes: Sequence[int], total_classes: i
z = k.variable(z) z = k.variable(z)
zd_result = tf.squeeze(z_discriminator(z)) zd_result = tf.squeeze(z_discriminator(z))
zd_real_loss = k.mean(k.binary_crossentropy(y_real_z, zd_result), axis=0) zd_real_loss = binary_crossentropy(y_real_z, zd_result)
z = tf.squeeze(encoder(x)) z = tf.squeeze(encoder(x))
zd_result = tf.squeeze(z_discriminator(z)) zd_result = tf.squeeze(z_discriminator(z))
zd_fake_loss = k.mean(k.binary_crossentropy(y_fake_z, zd_result), axis=0) zd_fake_loss = binary_crossentropy(y_fake_z, zd_result)
_zd_train_loss = zd_real_loss + zd_fake_loss _zd_train_loss = zd_real_loss + zd_fake_loss
@ -195,8 +196,8 @@ def train_mnist(folding_id: int, inlier_classes: Sequence[int], total_classes: i
x_decoded = decoder(z) x_decoded = decoder(z)
zd_result = tf.squeeze(z_discriminator(tf.squeeze(z))) zd_result = tf.squeeze(z_discriminator(tf.squeeze(z)))
encoder_loss = k.mean(k.binary_crossentropy(y_real_z, zd_result), axis=0) * 2.0 encoder_loss = binary_crossentropy(y_real_z, zd_result) * 2.0
recovery_loss = k.mean(k.binary_crossentropy(x, x_decoded)) recovery_loss = binary_crossentropy(x, x_decoded)
_enc_dec_train_loss = encoder_loss + recovery_loss _enc_dec_train_loss = encoder_loss + recovery_loss
enc_dec_grads = tape.gradient(_enc_dec_train_loss, enc_dec_grads = tape.gradient(_enc_dec_train_loss,