def get_symbol(num_classes=1000, num_layers=92, **kwargs):
if num_layers == 68:
k_R = 128
G = 32
k_sec = {2: 3, \
3: 4, \
4: 12, \
5: 3}
inc_sec = {2: 16, \
3: 32, \
4: 32, \
5: 64}
elif num_layers == 92:
k_R = 96
G = 32
k_sec = {2: 3, \
3: 4, \
4: 20, \
5: 3}
inc_sec = {2: 16, \
3: 32, \
4: 24, \
5: 128}
elif num_layers == 107:
k_R = 200
G = 50
k_sec = {2: 4, \
3: 8, \
4: 20, \
5: 3}
inc_sec = {2: 20, \
3: 64, \
4: 64, \
5: 128}
elif num_layers == 131:
k_R = 160
G = 40
k_sec = {2: 4, \
3: 8, \
4: 28, \
5: 3}
inc_sec = {2: 16, \
3: 32, \
4: 32, \
5: 128}
else:
raise ValueError(
"no experiments done on dpn num_layers {}, you can do it yourself".
format(num_layers))
version_se = kwargs.get('version_se', 1)
version_input = kwargs.get('version_input', 1)
assert version_input >= 0
version_output = kwargs.get('version_output', 'E')
fc_type = version_output
version_unit = kwargs.get('version_unit', 3)
print(version_se, version_input, version_output, version_unit)
## define Dual Path Network
data = mx.symbol.Variable(name="data")
# data = data-127.5
# data = data*0.0078125
# if version_input==0:
# conv1_x_1 = Conv(data=data, num_filter=128, kernel=(7, 7), name='conv1_x_1', pad=(3,3), stride=(2,2))
# else:
# conv1_x_1 = Conv(data=data, num_filter=128, kernel=(3, 3), name='conv1_x_1', pad=(3,3), stride=(1,1))
# conv1_x_1 = BN_AC(conv1_x_1, name='conv1_x_1__relu-sp')
# conv1_x_x = mx.symbol.Pooling(data=conv1_x_1, pool_type="max", kernel=(3, 3), pad=(1,1), stride=(2,2), name="pool1")
conv1_x_x = symbol_utils.get_head(data, version_input, 128)
# conv2
bw = 256
inc = inc_sec[2]
R = (k_R * bw) / 256
conv2_x_x = DualPathFactory(conv1_x_x, R, R, bw, 'conv2_x__1', inc, G,
'proj')
for i_ly in range(2, k_sec[2] + 1):
conv2_x_x = DualPathFactory(conv2_x_x, R, R, bw,
('conv2_x__%d' % i_ly), inc, G, 'normal')
# conv3
bw = 512
inc = inc_sec[3]
R = (k_R * bw) / 256
conv3_x_x = DualPathFactory(conv2_x_x, R, R, bw, 'conv3_x__1', inc, G,
'down')
for i_ly in range(2, k_sec[3] + 1):
conv3_x_x = DualPathFactory(conv3_x_x, R, R, bw,
('conv3_x__%d' % i_ly), inc, G, 'normal')
# conv4
bw = 1024
inc = inc_sec[4]
R = (k_R * bw) / 256
conv4_x_x = DualPathFactory(conv3_x_x, R, R, bw, 'conv4_x__1', inc, G,
'down')
for i_ly in range(2, k_sec[4] + 1):
conv4_x_x = DualPathFactory(conv4_x_x, R, R, bw,
('conv4_x__%d' % i_ly), inc, G, 'normal')
# conv5
bw = 2048
inc = inc_sec[5]
R = (k_R * bw) / 256
conv5_x_x = DualPathFactory(conv4_x_x, R, R, bw, 'conv5_x__1', inc, G,
'down')
for i_ly in range(2, k_sec[5] + 1):
conv5_x_x = DualPathFactory(conv5_x_x, R, R, bw,
('conv5_x__%d' % i_ly), inc, G, 'normal')
# output: concat
conv5_x_x = mx.symbol.Concat(*[conv5_x_x[0], conv5_x_x[1]],
name='conv5_x_x_cat-final')
# conv5_x_x = BN_AC(conv5_x_x, name='conv5_x_x__relu-sp')
before_pool = conv5_x_x
fc1 = symbol_utils.get_fc1(before_pool, num_classes, fc_type)
return fc1
def get_symbol(num_classes, num_layers, **kwargs):
"""Return DenseNet symbol of imagenet
Parameters
----------
units : list
Number of units in each stage
num_stage : int
Number of stage
growth_rate : int
Number of output channels
num_class : int
Ouput size of symbol
data_type : str
the type of dataset
reduction : float
Compression ratio. Default = 0.5
drop_out : float
Probability of an element to be zeroed. Default = 0.2
workspace : int
Workspace used in convolution operator
"""
version_input = kwargs.get('version_input', 1)
assert version_input>=0
version_output = kwargs.get('version_output', 'E')
fc_type = version_output
version_unit = kwargs.get('version_unit', 3)
print(version_input, version_output, version_unit)
num_stage = 4
reduction = 0.5
drop_out = 0.2
bottle_neck = True
bn_mom = 0.9
workspace = 256
growth_rate = 32
if num_layers == 121:
units = [6, 12, 24, 16]
elif num_layers == 169:
units = [6, 12, 32, 32]
elif num_layers == 201:
units = [6, 12, 48, 32]
elif num_layers == 161:
units = [6, 12, 36, 24]
growth_rate = 48
else:
raise ValueError("no experiments done on num_layers {}, you can do it yourself".format(num_layers))
num_unit = len(units)
assert(num_unit == num_stage)
init_channels = 2 * growth_rate
n_channels = init_channels
data = mx.sym.Variable(name='data')
#data = data-127.5
#data = data*0.0078125
#if version_input==0:
# body = mx.sym.Convolution(data=data, num_filter=growth_rate*2, kernel=(7, 7), stride=(2,2), pad=(3, 3),
# no_bias=True, name="conv0", workspace=workspace)
#else:
# body = mx.sym.Convolution(data=data, num_filter=growth_rate*2, kernel=(3,3), stride=(1,1), pad=(1,1),
# no_bias=True, name="conv0", workspace=workspace)
#body = mx.sym.BatchNorm(data=body, fix_gamma=False, eps=2e-5, momentum=bn_mom, name='bn0')
#body = mx.sym.Activation(data=body, act_type='relu', name='relu0')
#body = mx.symbol.Pooling(data=body, kernel=(3, 3), stride=(2,2), pad=(1,1), pool_type='max')
body = symbol_utils.get_head(data, version_input, growth_rate*2)
for i in range(num_stage-1):
body = DenseBlock(units[i], body, growth_rate=growth_rate, name='DBstage%d' % (i + 1), bottle_neck=bottle_neck, drop_out=drop_out, bn_mom=bn_mom, workspace=workspace)
n_channels += units[i]*growth_rate
n_channels = int(math.floor(n_channels*reduction))
body = TransitionBlock(i, body, n_channels, stride=(1,1), name='TBstage%d' % (i + 1), drop_out=drop_out, bn_mom=bn_mom, workspace=workspace)
body = DenseBlock(units[num_stage-1], body, growth_rate=growth_rate, name='DBstage%d' % (num_stage), bottle_neck=bottle_neck, drop_out=drop_out, bn_mom=bn_mom, workspace=workspace)
fc1 = symbol_utils.get_fc1(body, num_classes, fc_type)
return fc1
def get_symbol(num_classes, num_layers, **kwargs):
"""Return DenseNet symbol of imagenet
Parameters
----------
units : list
Number of units in each stage
num_stage : int
Number of stage
growth_rate : int
Number of output channels
num_class : int
Ouput size of symbol
data_type : str
the type of dataset
reduction : float
Compression ratio. Default = 0.5
drop_out : float
Probability of an element to be zeroed. Default = 0.2
workspace : int
Workspace used in convolution operator
"""
version_input = kwargs.get('version_input', 1)
assert version_input >= 0
version_output = kwargs.get('version_output', 'E')
fc_type = version_output
version_unit = kwargs.get('version_unit', 3)
print(version_input, version_output, version_unit)
num_stage = 4
reduction = 0.5
drop_out = 0.2
bottle_neck = True
bn_mom = 0.9
workspace = 256
growth_rate = 32
if num_layers == 121:
units = [6, 12, 24, 16]
elif num_layers == 169:
units = [6, 12, 32, 32]
elif num_layers == 201:
units = [6, 12, 48, 32]
elif num_layers == 161:
units = [6, 12, 36, 24]
growth_rate = 48
else:
raise ValueError(
"no experiments done on num_layers {}, you can do it yourself".
format(num_layers))
num_unit = len(units)
assert (num_unit == num_stage)
init_channels = 2 * growth_rate
n_channels = init_channels
data = mx.sym.Variable(name='data')
#data = data-127.5
#data = data*0.0078125
#if version_input==0:
# body = mx.sym.Convolution(data=data, num_filter=growth_rate*2, kernel=(7, 7), stride=(2,2), pad=(3, 3),
# no_bias=True, name="conv0", workspace=workspace)
#else:
# body = mx.sym.Convolution(data=data, num_filter=growth_rate*2, kernel=(3,3), stride=(1,1), pad=(1,1),
# no_bias=True, name="conv0", workspace=workspace)
#body = mx.sym.BatchNorm(data=body, fix_gamma=False, eps=2e-5, momentum=bn_mom, name='bn0')
#body = mx.sym.Activation(data=body, act_type='relu', name='relu0')
#body = mx.symbol.Pooling(data=body, kernel=(3, 3), stride=(2,2), pad=(1,1), pool_type='max')
body = symbol_utils.get_head(data, version_input, growth_rate * 2)
for i in range(num_stage - 1):
body = DenseBlock(units[i],
body,
growth_rate=growth_rate,
name='DBstage%d' % (i + 1),
bottle_neck=bottle_neck,
drop_out=drop_out,
bn_mom=bn_mom,
workspace=workspace)
n_channels += units[i] * growth_rate
n_channels = int(math.floor(n_channels * reduction))
body = TransitionBlock(i,
body,
n_channels,
stride=(1, 1),
name='TBstage%d' % (i + 1),
drop_out=drop_out,
bn_mom=bn_mom,
workspace=workspace)
body = DenseBlock(units[num_stage - 1],
body,
growth_rate=growth_rate,
name='DBstage%d' % (num_stage),
bottle_neck=bottle_neck,
drop_out=drop_out,
bn_mom=bn_mom,
workspace=workspace)
fc1 = symbol_utils.get_fc1(body, num_classes, fc_type)
return fc1
def get_symbol(num_classes = 1000, num_layers=92, **kwargs):
if num_layers==68:
k_R = 128
G = 32
k_sec = { 2: 3, \
3: 4, \
4: 12, \
5: 3 }
inc_sec= { 2: 16, \
3: 32, \
4: 32, \
5: 64 }
elif num_layers==92:
k_R = 96
G = 32
k_sec = { 2: 3, \
3: 4, \
4: 20, \
5: 3 }
inc_sec= { 2: 16, \
3: 32, \
4: 24, \
5: 128 }
elif num_layers==107:
k_R = 200
G = 50
k_sec = { 2: 4, \
3: 8, \
4: 20, \
5: 3 }
inc_sec= { 2: 20, \
3: 64, \
4: 64, \
5: 128 }
elif num_layers==131:
k_R = 160
G = 40
k_sec = { 2: 4, \
3: 8, \
4: 28, \
5: 3 }
inc_sec= { 2: 16, \
3: 32, \
4: 32, \
5: 128 }
else:
raise ValueError("no experiments done on dpn num_layers {}, you can do it yourself".format(num_layers))
version_se = kwargs.get('version_se', 1)
version_input = kwargs.get('version_input', 1)
assert version_input>=0
version_output = kwargs.get('version_output', 'E')
fc_type = version_output
version_unit = kwargs.get('version_unit', 3)
print(version_se, version_input, version_output, version_unit)
## define Dual Path Network
data = mx.symbol.Variable(name="data")
#data = data-127.5
#data = data*0.0078125
#if version_input==0:
# conv1_x_1 = Conv(data=data, num_filter=128, kernel=(7, 7), name='conv1_x_1', pad=(3,3), stride=(2,2))
#else:
# conv1_x_1 = Conv(data=data, num_filter=128, kernel=(3, 3), name='conv1_x_1', pad=(3,3), stride=(1,1))
#conv1_x_1 = BN_AC(conv1_x_1, name='conv1_x_1__relu-sp')
#conv1_x_x = mx.symbol.Pooling(data=conv1_x_1, pool_type="max", kernel=(3, 3), pad=(1,1), stride=(2,2), name="pool1")
conv1_x_x = symbol_utils.get_head(data, version_input, 128)
# conv2
bw = 256
inc= inc_sec[2]
R = (k_R*bw)/256
conv2_x_x = DualPathFactory( conv1_x_x, R, R, bw, 'conv2_x__1', inc, G, 'proj' )
for i_ly in range(2, k_sec[2]+1):
conv2_x_x = DualPathFactory( conv2_x_x, R, R, bw, ('conv2_x__%d'% i_ly), inc, G, 'normal')
# conv3
bw = 512
inc= inc_sec[3]
R = (k_R*bw)/256
conv3_x_x = DualPathFactory( conv2_x_x, R, R, bw, 'conv3_x__1', inc, G, 'down' )
for i_ly in range(2, k_sec[3]+1):
conv3_x_x = DualPathFactory( conv3_x_x, R, R, bw, ('conv3_x__%d'% i_ly), inc, G, 'normal')
# conv4
bw = 1024
inc= inc_sec[4]
R = (k_R*bw)/256
conv4_x_x = DualPathFactory( conv3_x_x, R, R, bw, 'conv4_x__1', inc, G, 'down' )
for i_ly in range(2, k_sec[4]+1):
conv4_x_x = DualPathFactory( conv4_x_x, R, R, bw, ('conv4_x__%d'% i_ly), inc, G, 'normal')
# conv5
bw = 2048
inc= inc_sec[5]
R = (k_R*bw)/256
conv5_x_x = DualPathFactory( conv4_x_x, R, R, bw, 'conv5_x__1', inc, G, 'down' )
for i_ly in range(2, k_sec[5]+1):
conv5_x_x = DualPathFactory( conv5_x_x, R, R, bw, ('conv5_x__%d'% i_ly), inc, G, 'normal')
# output: concat
conv5_x_x = mx.symbol.Concat(*[conv5_x_x[0], conv5_x_x[1]], name='conv5_x_x_cat-final')
#conv5_x_x = BN_AC(conv5_x_x, name='conv5_x_x__relu-sp')
before_pool = conv5_x_x
fc1 = symbol_utils.get_fc1(before_pool, num_classes, fc_type)
return fc1