def generate_model(self, **kwargs):
nfilters_embeded_last = int(kwargs.get("nfilters_embeded_last", 1))
nfilters_embeded = int(kwargs.get("nfilters_embeded", 1))
nlayers_embeded = int(kwargs.get("nlayers_embeded", 1))
input_model = Sequential()
for i in range(nlayers_embeded):
if i == 0:
nfilters = self.input_dims
input_model.add(
ConvPoolBNBST(nfilters, nfilters_embeded, 3, 1, 1, 3, 1,
1))
elif i == nlayers_embeded - 1:
nfilters = nfilters_embeded
input_model.add(
BinaryConvPoolBNBST(nfilters, nfilters_embeded_last, 3, 1,
1, 3, 1, 1))
else:
nfilters = nfilters_embeded
input_model.add(
BinaryConvPoolBNBST(nfilters, nfilters_embeded, 3, 1, 1, 3,
1, 1))
input_model.add(BinaryLinearBNSoftmax(None, self.output_dims))
input_model.build()
return input_model
def generate_model(self, **kwargs):
nfilters_embeded = int(kwargs.get("nfilters_embeded", 1))
nlayers_embeded = int(kwargs.get("nlayers_embeded", 1))
model = Sequential()
for i in range(nlayers_embeded):
if i == 0:
nfilters = self.input_dims
else:
nfilters = nfilters_embeded
model.add(Convolution2D(nfilters, nfilters_embeded, 3, 1, 1))
model.add(max_pooling_2d(3, 1, 1))
model.add(BatchNormalization(nfilters_embeded))
model.add(Activation('bst'))
model.add(Linear(None, self.output_dims))
model.build()
return model
def generate_model(self, **kwargs):
nfilters_embeded = int(kwargs.get("nfilters_embeded", 1))
nlayers_embeded = int(kwargs.get("nlayers_embeded", 1))
nfilters_cloud = int(kwargs.get("nfilters_cloud", 1))
nlayers_cloud = int(kwargs.get("nlayers_cloud", 1))
for key, value in kwargs.iteritems():
print key, value
input_model = Sequential()
for i in range(nlayers_embeded):
if i == 0:
nfilters = self.input_dims
input_model.add(
ConvPoolBNBST(nfilters, nfilters_embeded, 3, 1, 1, 3, 1,
1))
else:
nfilters = nfilters_embeded
input_model.add(
BinaryConvPoolBNBST(nfilters, nfilters_embeded, 3, 1, 1, 3,
1, 1))
branch = Sequential()
branch.add(BinaryLinearBNSoftmax(None, self.output_dims))
input_model.add(branch)
model = MultiInputSequential(self.ninputs)
for i in range(self.ninputs):
model.add_input(input_model)
# float branch
for i in range(nlayers_cloud):
if i == 0:
nfilters = self.ninputs * nfilters_embeded
else:
nfilters = nfilters_cloud
model.add(Convolution2D(nfilters, nfilters_cloud, 3, 1, 1))
model.add(max_pooling_2d(3, 1, 1))
model.add(BatchNormalization(nfilters_cloud))
model.add(Activation('bst'))
# Note: should we move pool to before batch norm like in binary?
model.add(Linear(None, self.output_dims))
model.add(BatchNormalization(self.output_dims))
model.build()
return model
def generate_model(self, **kwargs):
nfilters_embeded = int(kwargs.get("nfilters_embeded", 1))
nlayers_embeded = int(kwargs.get("nlayers_embeded", 1))
#print('test print kwargs.get(nfilters_embeded) ',nfilters_embeded)
#print('test print kwargs.get(nlayers_embeded) ',nlayers_embeded)
model = Sequential()
for i in range(nlayers_embeded):
if i == 0:
nfilters = self.input_dims
model.add(
ConvPoolBNBST(nfilters, nfilters_embeded, 3, 1, 1, 3, 1,
1))
# model.add(ConvBNBST(nfilters, nfilters_embeded, 3, 1, 1))
else:
nfilters = nfilters_embeded
model.add(
BinaryConvPoolBNBST(nfilters, nfilters_embeded, 3, 1, 1, 3,
1, 1))
# model.add(BinaryConvBNBST(nfilters, nfilters_embeded, 3, 1, 1))
model.add(BinaryLinearBNSoftmax(None, self.output_dims))
model.build()
return model
def generate_model(self, **kwargs):
nfilters_embeded_last = int(kwargs.get("nfilters_embeded_last", 1))
nfilters_embeded = int(kwargs.get("nfilters_embeded", 1))
nlayers_embeded = int(kwargs.get("nlayers_embeded", 1))
nfilters_cloud = int(kwargs.get("nfilters_cloud", 1))
nlayers_cloud = int(kwargs.get("nlayers_cloud", 1))
nfilters_edge = int(kwargs.get("nfilters_edge", 1))
nlayers_edge = int(kwargs.get("nlayers_edge", 1))
input_model = Sequential()
if nlayers_embeded == 1:
nfilters_embeded_last = nfilters_embeded
for i in range(nlayers_embeded):
if i == 0:
nfilters = self.input_dims
input_model.add(
ConvPoolBNBST(nfilters, nfilters_embeded, 3, 1, 1, 3, 2,
1))
elif i == nlayers_embeded - 1:
nfilters = nfilters_embeded
input_model.add(
BinaryConvPoolBNBST(nfilters, nfilters_embeded_last, 3, 1,
1, 3, 2, 1))
else:
nfilters = nfilters_embeded
input_model.add(
BinaryConvPoolBNBST(nfilters, nfilters_embeded, 3, 1, 1, 3,
2, 1))
branch = Sequential()
branch.add(BinaryLinearBNSoftmax(None, self.output_dims))
input_model.add(branch)
model = MultiInputSequential(self.ninputs,
merge_function=self.merge_function)
for i in range(self.ninputs):
model.add_input(input_model)
# Local branch
local_branch = Sequential()
#local_branch.add(Linear(None, self.output_dims))
#local_branch.add(BatchNormalization(self.output_dims))
if self.merge_function in [
'concat', 'concat_avg_pool', 'concat_max_pool'
]:
local_branch.add(BinaryLinearBNSoftmax(None, self.output_dims))
model.add_local(local_branch)
# Edge branches
#for i in range(nlayers_edge):
# if i == 0:
# if 'concat' in self.merge_function:
# nfilters = self.ninputs*nfilters_embeded_last
# else:
# nfilters = nfilters_embeded_last
# else:
# nfilters = nfilters_edge
# model.add(BinaryConvPoolBNBST(nfilters, nfilters_edge, 3, 1, 1, 3, 1, 1))
#
# #model.add(Convolution2D(nfilters, nfilters_edge, 3, 1, 1))
# #model.add(Activation('relu'))
# #model.add(max_pooling_2d(3,1,1))
# #model.add(BatchNormalization(nfilters_edge))
# #model.add(Activation('relu'))
# # Note: should we move pool to before batch norm like in binary?
#
#edge_branch = Sequential()
##edge_branch.add(Linear(None, self.output_dims))
##edge_branch.add(BatchNormalization(self.output_dims))
#edge_branch.add(BinaryLinearBNSoftmax(None, self.output_dims))
#model.add(edge_branch)
# Cloud branches
for i in range(nlayers_cloud):
if i == 0:
if self.merge_function in [
'concat', 'avg_pool_concat', 'max_pool_concat'
]:
nfilters = self.ninputs * nfilters_embeded_last
else:
nfilters = nfilters_embeded_last
else:
nfilters = nfilters_cloud
model.add(
BinaryConvPoolBNBST(nfilters, nfilters_cloud, 3, 1, 1, 3, 1,
1))
#model.add(Convolution2D(nfilters, nfilters_cloud, 3, 1, 1))
#model.add(Activation('relu'))
#model.add(max_pooling_2d(3,1,1))
#model.add(BatchNormalization(nfilters_cloud))
#model.add(Activation('relu'))
# Note: should we move pool to before batch norm like in binary?
#model.add(Linear(None, self.output_dims))
#model.add(BatchNormalization(self.output_dims))
model.add(BinaryLinearBNSoftmax(None, self.output_dims))
model.build()
return model
#nlayers_cloud = int(2)
branchweight = numpy.float64(0.1)
ent_T = numpy.int64(100)
lr = numpy.float64(0.001)
nepochs = int(60)
name = str("cifar10_10f_end")
input_dims = 3
output_dims = 10
model = Sequential()
for i in range(nlayers_embeded):
if i == 0:
nfilters = input_dims
model.add(ConvPoolBNBST(nfilters, nfilters_embeded, 3, 1, 1, 3, 1, 1))
else:
nfilters = nfilters_embeded
model.add(
BinaryConvPoolBNBST(nfilters, nfilters_embeded, 3, 1, 1, 3, 1, 1))
model.add(BinaryLinearBNSoftmax(None, output_dims))
# float branch
#for i in range(nlayers_cloud):
# if i == 0:
# nfilters = nfilters_embeded
# else:
# nfilters = nfilters_cloud
# model.add(Convolution2D(nfilters, nfilters_cloud, 3, 1, 1))
# model.add(max_pooling_2d(3,1,1))
# seq.add(function.Activation("relu"))
# seq.add(link.Deconvolution2D(32, 16, ksize=4, stride=2, pad=paddings.pop(0), use_weightnorm=True))
# seq.add(link.BatchNormalization(16))
# seq.add(function.Activation("relu"))
# seq.add(link.Deconvolution2D(16, 3, ksize=4, stride=2, pad=paddings.pop(0), use_weightnorm=True))
# seq.build()
# y = seq(x)
# print y.data.shape
# train test
x = np.random.normal(scale=1, size=(128, 28 * 28)).astype(np.float32)
x = Variable(x)
seq = Sequential(weight_initializer="GlorotNormal", weight_init_std=0.05)
seq.add(Linear(28 * 28, 500, use_weightnorm=True))
seq.add(BatchNormalization(500))
seq.add(Activation("relu"))
seq.add(Linear(None, 500, use_weightnorm=True))
seq.add(BatchNormalization(500))
seq.add(Activation("relu"))
seq.add(Linear(500, 28 * 28, use_weightnorm=True))
seq.build()
chain = Chain()
chain.add_sequence(seq)
chain.setup_optimizers("adam",
0.001,
momentum=0.9,
weight_decay=0.000001,
gradient_clipping=10)
def generate_model(self, **kwargs):
nfilters_embeded = int(kwargs.get("nfilters_embeded", 1))
nlayers_embeded = int(kwargs.get("nlayers_embeded", 1))
nfilters_cloud = int(kwargs.get("nfilters_cloud", 1))
nlayers_cloud = int(kwargs.get("nlayers_cloud", 1))
model = Sequential()
for i in range(nlayers_embeded):
if i == 0:
nfilters = self.input_dims
else:
nfilters = nfilters_embeded
model.add(Convolution2D(nfilters, nfilters_embeded, 3, 1, 1))
model.add(max_pooling_2d(3, 1, 1))
model.add(BatchNormalization(nfilters_embeded))
model.add(Activation('bst'))
branch = Sequential()
branch.add(Linear(None, self.output_dims))
branch.add(BatchNormalization(self.output_dims))
model.add(branch)
# float branch
for i in range(nlayers_cloud):
if i == 0:
nfilters = nfilters_embeded
else:
nfilters = nfilters_cloud
model.add(Convolution2D(nfilters, nfilters_cloud, 3, 1, 1))
model.add(max_pooling_2d(3, 1, 1))
model.add(BatchNormalization(nfilters_cloud))
model.add(Activation('bst'))
# Note: should we move pool to before batch norm like in binary?
model.add(Linear(None, self.output_dims))
model.add(BatchNormalization(self.output_dims))
model.build()
return model
folder = "_models/test_seq"
nfilters_embeded = int(2)
nlayers_embeded = int(1)
nfilters_cloud = int(2)
nlayers_cloud = int(1)
branchweight = int(3)
lr = numpy.float64(0.001)
nepochs = int(10)
name = str("_lr_0.001_nfilters_embeded_2_nlayers_embeded_1.0")
model = Sequential()
nfilters = 1
model.add(ConvPoolBNBST(nfilters, nfilters_embeded, 3, 1, 1, 3, 1, 1))
#model.add(ConvBNBST(nfilters, nfilters_embeded, 3, 1, 1))
nfilters = nfilters_embeded
model.add(BinaryConvPoolBNBST(nfilters, nfilters_embeded, 3, 1, 1, 3, 1, 1))
#model.add(BinaryConvBNBST(nfilters, nfilters_embeded, 3, 1, 1))
branch = Sequential()
branch.add(BinaryLinearBNSoftmax(None, 10))
model.add(branch)
############################ Embeded network ##############################
nfilters = nfilters_embeded
model.add(Convolution2D(nfilters, nfilters_embeded, 3, 1, 1))
model.add(BatchNormalization(nfilters_embeded))
def generate_model(self, **kwargs):
nfilters_embeded = int(kwargs.get("nfilters_embeded", 1))
nlayers_embeded = int(kwargs.get("nlayers_embeded", 1))
nfilters_cloud = int(kwargs.get("nfilters_cloud", 1))
nlayers_cloud = int(kwargs.get("nlayers_cloud", 1))
model = Sequential()
for i in range(nlayers_embeded):
if i == 0:
nfilters = self.input_dims
model.add(
ConvPoolBNBST(nfilters, nfilters_embeded, 3, 1, 1, 3, 1,
1))
else:
nfilters = nfilters_embeded
model.add(
BinaryConvPoolBNBST(nfilters, nfilters_embeded, 3, 1, 1, 3,
1, 1))
branch = Sequential()
branch.add(BinaryLinearBNSoftmax(None, self.output_dims))
model.add(branch)
for i in range(nlayers_cloud):
if i == 0:
nfilters = nfilters_embeded
else:
nfilters = nfilters_cloud
model.add(Convolution2D(nfilters, nfilters_cloud, 3, 1, 1))
model.add(BatchNormalization(nfilters_cloud))
model.add(Activation('relu'))
model.add(max_pooling_2d(3, 1, 1))
model.add(Linear(None, self.output_dims))
model.build()
return model
