def __init__(self, num_classes, optimizer, seed=1):
# params
# print("**** Numb classes:", num_classes)
self.num_classes = num_classes
self.optimizer = optimizer
#self.create_model(optimizer)
# create computation graph
self.graph = tf.Graph()
with self.graph.as_default():
tf.set_random_seed(123 + seed)
self.features, self.labels, self.train_op, self.grads, self.eval_metric_ops, self.loss = self.create_model(
optimizer)
self.saver = tf.train.Saver()
self.sess = tf.Session(graph=self.graph)
# find memory footprint and compute cost of the model
self.size = graph_size(self.graph)
with self.graph.as_default():
self.sess.run(tf.global_variables_initializer())
metadata = tf.RunMetadata()
opts = tf.profiler.ProfileOptionBuilder.float_operation()
self.flops = tf.profiler.profile(self.graph,
run_meta=metadata,
cmd='scope',
options=opts).total_float_ops
def __init__(self, seq_len, num_classes, n_hidden, optimizer, seed):
self.seq_len = seq_len
self.num_classes = num_classes
self.n_hidden = n_hidden
self.optimizer = optimizer
self.graph = tf.Graph()
with self.graph.as_default():
tf.set_random_seed(123 + seed)
self.features, self.labels, self.train_op, self.grads, self.eval_metric_ops, self.loss = self.create_model(
optimizer)
self.saver = tf.train.Saver()
self.sess = tf.Session(graph=self.graph)
self.size = graph_size(self.graph)
with self.graph.as_default():
self.sess.run(tf.global_variables_initializer())
metadata = tf.RunMetadata()
opts = tf.profiler.ProfileOptionBuilder.float_operation()
self.flops = tf.profiler.profile(self.graph,
run_meta=metadata,
cmd='scope',
options=opts).total_float_ops
def __init__(self, num_classes, optimizer, seed=1):
"""
定义 Omniglot 的 CNN 模型
:param num_classes:
:param optimizer:
:param seed:
"""
# params
self.num_classes = num_classes
self.train_lr = 1e-3
self.meta_lr = 1e-2
# create computation graph
self.graph = tf.Graph()
with self.graph.as_default():
# TODO 参数; 每次运行一个 task/client
self.train_op = self.build(K=5,
task_num=1,
query_shots=10,
query_ways=5,
sprt_shots=10,
sprt_ways=5)
self.saver = tf.train.Saver()
self.sess = tf.Session(graph=self.graph)
# find memory footprint and compute cost of the model
self.size = graph_size(self.graph)
with self.graph.as_default():
self.sess.run(tf.global_variables_initializer())
metadata = tf.RunMetadata()
opts = tf.profiler.ProfileOptionBuilder.float_operation()
self.flops = tf.profiler.profile(self.graph,
run_meta=metadata,
cmd='scope',
options=opts).total_float_ops
def __init__(self, num_classes, q, optimizer, seed=1):
# params
self.num_classes = num_classes
# create computation graph
self.graph = tf.Graph()
with self.graph.as_default():
tf.set_random_seed(123 + seed)
self.features, self.labels, self.output2, self.train_op, self.grads, self.kl_grads, self.eval_metric_ops, \
self.loss, self.kl_loss, self.soft_max, self.predictions = self.create_model(q, optimizer)
self.saver = tf.train.Saver()
config = tf.ConfigProto()
config.gpu_options.allow_growth = True
self.sess = tf.Session(graph=self.graph, config=config)
# find memory footprint and compute cost of the model
self.size = graph_size(self.graph)
with self.graph.as_default():
self.sess.run(tf.global_variables_initializer())
metadata = tf.RunMetadata()
opts = tf.profiler.ProfileOptionBuilder.float_operation()
self.flops = tf.profiler.profile(self.graph,
run_meta=metadata,
cmd='scope',
options=opts).total_float_ops
def __init__(self, num_classes, optimizer, seed=1):
# params
self.num_classes = num_classes
# select optimizer
if optimizer == 'SGD':
self.lr = tf.placeholder(tf.float32, shape=(), name='lr')
selected_optimizer = tf.train.GradientDescentOptimizer(
learning_rate=self.lr)
# create computation graph
self.graph = tf.Graph()
with self.graph.as_default():
tf.set_random_seed(123 + seed)
self.features, self.labels, self.train_op, self.grads, self.eval_metric_ops, self.loss, self.pred = self.create_model(
selected_optimizer)
self.saver = tf.train.Saver()
self.sess = tf.Session(graph=self.graph)
# find memory footprint and compute cost of the model
self.size = graph_size(self.graph)
with self.graph.as_default():
self.sess.run(tf.global_variables_initializer())
metadata = tf.RunMetadata()
opts = tf.profiler.ProfileOptionBuilder.float_operation()
self.flops = tf.profiler.profile(self.graph,
run_meta=metadata,
cmd='scope',
options=opts).total_float_ops
