from collections import defaultdict
from math import inf
from tensorflow.python.ops.losses.losses_impl import Reduction
from trainers import trainer_utils
from trainers.AEMODEL import AEMODEL, Phase, indicate_early_stopping, update_log_dicts
from trainers.DLMODEL import *
class AnoVAEGAN(AEMODEL):
class Config(AEMODEL.Config):
def __init__(self):
super().__init__('AnoVAEGAN')
self.scale = 10.0
self.kappa = 1.0
self.kl_weight = 1.0
def __init__(self, sess, config, network=None):
super().__init__(sess, config, network)
self.x = tf.placeholder(tf.float32, [None, self.config.outputHeight, self.config.outputWidth, self.config.numChannels], name='x')
self.z = tf.placeholder(tf.float32, [None, self.config.zDim], name='z')
self.outputs = self.network(self.x, dropout_rate=self.dropout_rate, dropout=self.dropout, config=self.config)
self.reconstruction = self.outputs['out']
self.z_mu = self.outputs['z_mu']
self.z_sigma = self.outputs['z_sigma']
self.d_fake_features = self.outputs['d_fake_features']
self.d_ = self.outputs['d_']
self.d_features = self.outputs['d_features']
self.d = self.outputs['d']
self.x_hat = self.outputs['x_hat']
self.d_hat = self.outputs['d_hat']
self.kappa = self.config.kappa
self.kl_weight = self.config.kl_weight
self.scale = self.config.scale
# Print Stats
self.get_number_of_trainable_params()
# Instantiate Saver
self.saver = tf.train.Saver()
def train(self, dataset):
# Determine trainable variables
self.variables = tf.get_collection(tf.GraphKeys.TRAINABLE_VARIABLES)
# Build losses
self.losses['disc_fake'] = disc_fake = tf.reduce_mean(self.d_)
self.losses['disc_real'] = disc_real = tf.reduce_mean(self.d)
disc_loss = disc_fake - disc_real
ddx = tf.gradients(self.d_hat, self.x_hat)[0] # gradient
ddx = tf.sqrt(tf.reduce_sum(tf.square(ddx), axis=1)) # slopes
ddx = tf.reduce_mean(tf.square(ddx - 1.0)) * self.scale # gradient penalty
self.losses['disc_loss'] = disc_loss = disc_loss + ddx
# Build losses
kl = 0.5 * tf.reduce_sum(tf.square(self.z_mu) + tf.square(self.z_sigma) - tf.log(tf.square(self.z_sigma)) - 1,
axis=1)
self.losses['kl'] = loss_kl = tf.reduce_mean(kl)
self.losses['loss_img'] = tf.reduce_mean(
tf.reduce_mean(tf.losses.mean_squared_error(self.x, self.reconstruction, reduction=Reduction.NONE), axis=[1, 2, 3]))
self.losses['loss_fts'] = tf.reduce_mean(
tf.reduce_mean(tf.losses.mean_squared_error(self.d_fake_features, self.d_features, reduction=Reduction.NONE), axis=[1, 2, 3]))
self.losses['L1'] = tf.losses.absolute_difference(self.x, self.reconstruction, reduction=Reduction.NONE)
self.losses['reconstructionLoss'] = self.losses['loss'] = tf.reduce_mean(tf.reduce_sum(self.losses['L1'], axis=[1, 2, 3]))
self.losses['gen_loss'] = gen_loss = - disc_fake
self.losses['enc_loss'] = enc_loss = self.losses['reconstructionLoss'] + self.kl_weight * loss_kl
with tf.control_dependencies(tf.get_collection(tf.GraphKeys.UPDATE_OPS)):
# Set the optimizer
t_vars = tf.trainable_variables()
dis_vars = [var for var in t_vars if 'Discriminator' in var.name]
gen_vars = [var for var in t_vars if 'Generator' in var.name]
enc_vars = [var for var in t_vars if 'Encoder' in var.name]
optim_dis = tf.train.AdamOptimizer(learning_rate=self.config.learningrate, beta1=0.5, beta2=0.9).minimize(disc_loss, var_list=dis_vars)
optim_gen = tf.train.AdamOptimizer(learning_rate=self.config.learningrate, beta1=0.5, beta2=0.9).minimize(gen_loss, var_list=gen_vars)
optim_vae = tf.train.AdamOptimizer(learning_rate=self.config.learningrate, beta1=0.5, beta2=0.9).minimize(enc_loss, var_list=enc_vars + gen_vars)
# initialize all variables
tf.global_variables_initializer().run(session=self.sess)
best_cost = inf
last_improvement = 0
last_epoch = self.load_checkpoint()
# Go go go!
for epoch in range(last_epoch, self.config.numEpochs):
#################
# TRAINING WGAN #
#################
phase = Phase.TRAIN
scalars = defaultdict(list)
visuals = []
d_iters = 5
num_batches = dataset.num_batches(self.config.batchsize, set=phase.value)
for idx in range(0, num_batches):
batch, _, _ = dataset.next_batch(self.config.batchsize, set=phase.value)
# Encoder optimization
fetches = {
# 'generated': self.generated,
'reconstruction': self.reconstruction,
'reconstructionLoss': self.losses['reconstructionLoss'],
'L1': self.losses['L1'],
'enc_loss': self.losses['enc_loss'],
'optimizer_e': optim_vae,
}
feed_dict = {
self.x: batch,
self.dropout: phase == Phase.TRAIN,
self.dropout_rate: self.config.dropout_rate
}
run = self.sess.run(fetches, feed_dict=feed_dict)
# Generator optimization
fetches = {
'gen_loss': self.losses['gen_loss'],
'optimizer_g': optim_gen,
}
feed_dict = {
self.x: batch,
self.dropout: phase == Phase.TRAIN,
self.dropout_rate: self.config.dropout_rate
}
run = {**run, **self.sess.run(fetches, feed_dict=feed_dict)}
for _ in range(0, d_iters):
# Discriminator optimization
fetches = {
'disc_loss': self.losses['disc_loss'],
'disc_fake': self.losses['disc_fake'],
'disc_real': self.losses['disc_real'],
'optimizer_d': optim_dis,
}
feed_dict = {
self.x: batch,
self.dropout: phase == Phase.TRAIN,
self.dropout_rate: self.config.dropout_rate
}
run = {**run, **self.sess.run(fetches, feed_dict=feed_dict)}
# Print to console
print(f'Epoch ({phase.value}): [{epoch:2d}] [{idx:4d}/{num_batches:4d}]'
f' gen_loss: {run["gen_loss"]:.8f}, disc_loss: {run["disc_loss"]:.8f}, reconstructionLoss: {run["reconstructionLoss"]:.8f}')
update_log_dicts(*trainer_utils.get_summary_dict(batch, run), scalars, visuals)
self.log_to_tensorboard(epoch, scalars, visuals, phase)
# Increment last_epoch counter and save model
last_epoch += 1
self.save(self.checkpointDir, last_epoch)
##############
# VALIDATION #
##############
phase = Phase.VAL
scalars = defaultdict(list)
visuals = []
num_batches = dataset.num_batches(self.config.batchsize, set=phase.value)
for idx in range(0, num_batches):
batch, _, _ = dataset.next_batch(self.config.batchsize, set=phase.value)
# Encoder optimization
fetches = {
'reconstruction': self.reconstruction,
'reconstructionLoss': self.losses['reconstructionLoss'],
'L1': self.losses['L1'],
'enc_loss': self.losses['enc_loss'],
}
feed_dict = {
self.x: batch,
self.dropout: phase == Phase.TRAIN,
self.dropout_rate: self.config.dropout_rate
}
run = self.sess.run(fetches, feed_dict=feed_dict)
# Print to console
print(f'Epoch ({phase.value}): [{epoch:2d}] [{idx:4d}/{num_batches:4d}] reconstructionLoss: {run["reconstructionLoss"]:.8f}')
update_log_dicts(*trainer_utils.get_summary_dict(batch, run), scalars, visuals)
self.log_to_tensorboard(epoch, scalars, visuals, phase)
best_cost, last_improvement, stop = indicate_early_stopping(scalars['reconstructionLoss'], best_cost, last_improvement)
if stop:
print('Early stopping was triggered due to no improvement over the last 5 epochs')
break
def reconstruct(self, x, dropout=False):
if x.ndim < 4:
x = np.expand_dims(x, 0)
fetches = {
'reconstruction': self.reconstruction
}
feed_dict = {
self.x: x,
self.dropout: dropout, # apply only during MC sampling.
self.dropout_rate: self.config.dropout_rate
}
results = self.sess.run(fetches, feed_dict=feed_dict)
results['l1err'] = np.sum(np.abs(x - results['reconstruction']))
results['l2err'] = np.sum(np.sqrt((x - results['reconstruction']) ** 2))
return results
