def make_nn(model_path):
nn = Model(model_path)
nn.build_graph(
image_shape=(32, 32, 1),
n_classes=10,
layers=[
Flatten(),
Dense(32 * 32, activation='relu'),
Dense(10),
],
alpha=1e-4,
)
return nn
def make_simplenet(model_path):
cnn = Model(model_path)
cnn.build_graph(
image_shape=(32, 32, 1),
n_classes=10,
layers=[
Conv2D(64, (3, 3), activation='relu', batch_normal=0.95), #1
Conv2D(32, (3, 3), activation='relu', batch_normal=0.95), #2
Conv2D(32, (3, 3), activation='relu', batch_normal=0.95), #3
Conv2D(32, (3, 3), activation='relu', batch_normal=0.95), #4
MaxPooling2D((2, 2)),
Conv2D(32, (3, 3), activation='relu', batch_normal=0.95), #5
Conv2D(32, (3, 3), activation='relu', batch_normal=0.95), #6
Conv2D(64, (3, 3), activation='relu', batch_normal=0.95), #7
MaxPooling2D((2, 2)),
Conv2D(64, (3, 3), activation='relu', batch_normal=0.95), #8
Conv2D(64, (3, 3), activation='relu', batch_normal=0.95), #9
MaxPooling2D((2, 2)),
Conv2D(128, (3, 3), activation='relu', batch_normal=0.95), #10
Conv2D(256, (1, 1), activation='relu', batch_normal=0.95), #11
Conv2D(64, (1, 1), activation='relu', batch_normal=0.95), #12
MaxPooling2D((2, 2)),
Conv2D(64, (3, 3), activation='relu', batch_normal=0.95), #13
MaxPooling2D((2, 2)),
Flatten(),
Dense(10),
],
alpha=1e-3,
)
return cnn
def make_cnn(model_path):
cnn = Model(model_path)
cnn.build_graph(
image_shape=(32, 32, 1),
n_classes=10,
layers=[
Conv2D(32, (3, 3), activation='relu'),
BatchNorm(),
MaxPooling2D((2, 2)),
Flatten(),
Dense(128, activation='relu'),
Dropout(0.5),
Dense(10),
],
alpha=1e-3,
)
return cnn
def __init__(self,
input_dim,
output_dim,
model_size="small",
neigh_input_dim=None,
dropout=0.,
bias=False,
act=tf.nn.relu,
name=None,
concat=False,
**kwargs):
super(MeanPoolingAggregator, self).__init__(**kwargs)
self.dropout = dropout
self.bias = bias
self.act = act
self.concat = concat
if neigh_input_dim is None:
neigh_input_dim = input_dim
if name is not None:
name = '/' + name
else:
name = ''
if model_size == "small":
hidden_dim = self.hidden_dim = 512
elif model_size == "big":
hidden_dim = self.hidden_dim = 1024
self.mlp_layers = []
self.mlp_layers.append(
Dense(input_dim=neigh_input_dim,
output_dim=hidden_dim,
act=tf.nn.relu,
dropout=dropout,
sparse_inputs=False,
logging=self.logging))
with tf.variable_scope(self.name + name + '_vars'):
self.vars['neigh_weights'] = glorot([hidden_dim, output_dim],
name='neigh_weights')
self.vars['self_weights'] = glorot([input_dim, output_dim],
name='self_weights')
if self.bias:
self.vars['bias'] = zeros([self.output_dim], name='bias')
if self.logging:
self._log_vars()
self.input_dim = input_dim
self.output_dim = output_dim
self.neigh_input_dim = neigh_input_dim
def make_lr(model_path):
lr = Model(model_path)
lr.build_graph(image_shape=(32, 32, 1),
n_classes=10,
layers=[
Flatten(),
Dense(10),
],
alpha=1e-3)
return lr
def make_simplenet_dropout(model_path):
cnn = Model(model_path)
cnn.build_graph(
image_shape=(32, 32, 1),
n_classes=10,
layers=[
Conv2D(64, (3, 3), activation='relu', batch_normal=0.95), #1
Dropout(0.8),
Conv2D(32, (3, 3), activation='relu', batch_normal=0.95), #2
Dropout(0.8),
Conv2D(32, (3, 3), activation='relu', batch_normal=0.95), #3
Dropout(0.8),
Conv2D(32, (3, 3), activation='relu', batch_normal=0.95), #4
Dropout(0.8),
MaxPooling2D((2, 2)),
Conv2D(32, (3, 3), activation='relu', batch_normal=0.95), #5
Dropout(0.8),
Conv2D(32, (3, 3), activation='relu', batch_normal=0.95), #6
Dropout(0.8),
Conv2D(64, (3, 3), activation='relu', batch_normal=0.95), #7
Dropout(0.8),
MaxPooling2D((2, 2)),
Conv2D(64, (3, 3), activation='relu', batch_normal=0.95), #8
Dropout(0.8),
Conv2D(64, (3, 3), activation='relu', batch_normal=0.95), #9
Dropout(0.8),
MaxPooling2D((2, 2)),
Conv2D(128, (3, 3), activation='relu', batch_normal=0.95), #10
Dropout(0.8),
Conv2D(256, (1, 1), activation='relu', batch_normal=0.95), #11
Dropout(0.8),
Conv2D(64, (1, 1), activation='relu', batch_normal=0.95), #12
Dropout(0.8),
MaxPooling2D((2, 2)),
Conv2D(64, (3, 3), activation='relu', batch_normal=0.95), #13
Dropout(0.8),
MaxPooling2D((2, 2)),
Flatten(),
Dense(10),
],
alpha=1e-3,
)
return cnn