def optimizer(self):
lr = tf.get_variable('learning_rate', initializer=0.1, trainable=False)
opt = tf.train.MomentumOptimizer(lr, 0.9, use_nesterov=True)
gradprocs = [
gradproc.ScaleGradient([('conv0.*', 0.1), ('group[0-2].*', 0.1)])
]
return optimizer.apply_grad_processors(opt, gradprocs)
def optimizer(self):
lr = tf.get_variable('learning_rate', initializer=2e-4, dtype=tf.float32, trainable=False)
# lr = tf.get_variable('learning_rate', initializer=0.0, dtype=tf.float32, trainable=False)
opt = tf.train.AdamOptimizer(lr, beta1=0.5, epsilon=1e-6)
# generator learns 5 times faster
return optimizer.apply_grad_processors(
opt, [gradproc.ScaleGradient(('gen/.*', 5))])
def optimizer(self):
lr = tf.get_variable('learning_rate', initializer=5e-4, trainable=False)
opt = tf.train.AdamOptimizer(lr, epsilon=1e-3)
return optimizer.apply_grad_processors(
opt, [
gradproc.ScaleGradient(('STN.*', 0.1)),
gradproc.SummaryGradient()])
def _get_optimizer(self):
lr = tf.get_variable('learning_rate',
initializer=3e-5,
trainable=False)
opt = tf.train.AdamOptimizer(lr, epsilon=1e-3)
return optimizer.apply_grad_processors(opt, [
gradproc.ScaleGradient([('convfcweight.*', 0.1), ('conv5_.*', 5)])
])
def optimizer(self):
lr = tf.get_variable('lr', initializer=3e-3, trainable=False)
# opt = tf.train.GradientDescentOptimizer(lr)
opt = tf.train.AdamOptimizer(lr)
# freeze all variables in network
opt = optimizer.apply_grad_processors(
opt, [gradproc.ScaleGradient([('.*/b', 0.), ('.*/W', 0.)])])
return opt
def optimizer(self):
lr = tf.get_variable('learning_rate',
initializer=args.base_lr,
trainable=False)
opt = tf.train.MomentumOptimizer(lr, 0.9, use_nesterov=True)
add_moving_summary(lr)
if args.load:
gradprocs = [
gradproc.ScaleGradient([('conv.*', 0.1), ('fc.*', 0.1)])
]
return optimizer.apply_grad_processors(opt, gradprocs)
else:
return opt
def _get_optimizer(self):
lr = symbf.get_scalar_var('learning_rate', 0.003, summary=True)
factor = get_batch_factor()
if factor != 1:
lr = lr / float(factor)
opt = tf.train.MomentumOptimizer(lr, 0.9)
opt = optimizer.AccumGradOptimizer(opt, factor)
else:
opt = tf.train.MomentumOptimizer(lr, 0.9)
return opt
return optimizer.apply_grad_processors(
opt, [gradproc.ScaleGradient(('.*/b', 2))])
def _get_optimizer(self):
lr = tf.get_variable('learning_rate',
initializer=0.003,
trainable=False)
tf.summary.scalar('learning_rate', lr)
factor = get_batch_factor()
if factor != 1:
lr = lr / float(factor)
opt = tf.train.MomentumOptimizer(lr, 0.9)
opt = optimizer.AccumGradOptimizer(opt, factor)
else:
opt = tf.train.MomentumOptimizer(lr, 0.9)
return opt
return optimizer.apply_grad_processors(
opt, [gradproc.ScaleGradient(('.*/b', 2))])
def _get_optimizer(self):
lr = get_scalar_var('learning_rate', cfg.base_lr, summary=True)
opt = tf.train.MomentumOptimizer(learning_rate=lr,
momentum=cfg.momentum)
gradprocs = [
gradproc.ScaleGradient([('conv.*/W', 1),
('conv.*/b', cfg.bias_lr_mult),
('bottleneck.*/W', 1),
('bottleneck.*/b', cfg.bias_lr_mult),
('stage_1.*/W', 1),
('stage_1.*/b', cfg.bias_lr_mult),
('stage_[2-6].*/W', cfg.lr_mult),
('stage_[2-6].*/b',
cfg.lr_mult * cfg.bias_lr_mult)])
]
return optimizer.apply_grad_processors(opt, gradprocs)
def _get_optimizer(self):
lr = symbf.get_scalar_var('learning_rate', 0.003, summary=True)
opt = tf.train.MomentumOptimizer(lr, 0.9)
return optimizer.apply_grad_processors(
opt, [gradproc.ScaleGradient(('.*/b', 2))])
def optimizer(self):
lr = tf.get_variable('learning_rate', initializer=1e-3, trainable=False)
# opt = tf.train.MomentumOptimizer(lr, 0.9)
opt = tf.train.AdamOptimizer(learning_rate=lr)
# return opt
return optimizer.apply_grad_processors(opt, [gradproc.ScaleGradient(('5_Tr_Cv.*', 0.1)), gradproc.SummaryGradient()])