def SE_block(x,nb_filter,block,stage):
y = Pooling_v1(x,name='conv'+str(stage)+'_'+str(block)+'_global_pool',kernel=(2,2),pool_type='avg',global_pool=True)
y = FullyConnected(y, name='conv'+str(stage)+'_'+str(block)+'_1x1_down', num_hidden=int(nb_filter/16))
y = Activation(y, name='conv'+str(stage)+'_'+str(block)+'_1x1_down_relu', act_type='relu')
y = FullyConnected(y, name='conv'+str(stage)+'_'+str(block)+'_1x1_up', num_hidden=nb_filter)
y = Activation(y, name='conv'+str(stage)+'_'+str(block)+'_prob', act_type='sigmoid')
y = broadcast_mul(Reshape(y,shape=(-1,nb_filter,1,1)),x)
return y
def detai_net(inputs,classes=1000,use_att=False,model_name='resnetv20'):
"""Instantiates the ResNet50 architecture.
"""
if use_att:
x = inputs[0]
else:
x = inputs
x = Norm(x,fix_gamma=True,use_global_stats=False,eps=1e-5,name=model_name+'_batchnorm0')
x = Convolution(x,
num_filter=64, kernel=(7, 7), stride=(2, 2), pad=(3,3), no_bias=True, name=model_name+'_conv0')
x = Norm(x,fix_gamma=False,use_global_stats=False,eps=1e-5,name=model_name+'_batchnorm1')
x = Activation(x,name=model_name+'_relu0',act_type='relu')
x = Pooling(x,kernel=(3,3),stride=(2,2),pad=(1,1),pool_type='max',name=model_name+'_pool0')
x = conv_block(x, 3, [64, 64, 256], stage=1, block=0, strides=(1, 1),model_name=model_name)
x = identity_block(x, 3, [64, 64, 256], stage=1, block=3,model_name=model_name)
x = identity_block(x, 3, [64, 64, 256], stage=1, block=6,model_name=model_name)
x = conv_block(x, 3, [128, 128, 512], stage=2, block=0,model_name=model_name)
x = identity_block(x, 3, [128, 128, 512], stage=2, block=3,model_name=model_name)
x = identity_block(x, 3, [128, 128, 512], stage=2, block=6,model_name=model_name)
x = identity_block(x, 3, [128, 128, 512], stage=2, block=9,model_name=model_name)
x = conv_block(x, 3, [256, 256, 1024], stage=3, block=0,model_name=model_name)
x = identity_block(x, 3, [256, 256, 1024], stage=3, block=3,model_name=model_name)
x = identity_block(x, 3, [256, 256, 1024], stage=3, block=6,model_name=model_name)
x = identity_block(x, 3, [256, 256, 1024], stage=3, block=9,model_name=model_name)
x = identity_block(x, 3, [256, 256, 1024], stage=3, block=12,model_name=model_name)
x = identity_block(x, 3, [256, 256, 1024], stage=3, block=15,model_name=model_name)
x = conv_block(x, 3, [512, 512, 2048], stage=4, block=0,model_name=model_name)
x = identity_block(x, 3, [512, 512, 2048], stage=4, block=3,model_name=model_name)
x = identity_block(x, 3, [512, 512, 2048], stage=4, block=6,model_name=model_name)
x = Norm(x,fix_gamma=False,use_global_stats=False,eps=1e-5,name=model_name+'_batchnorm2')
x = Activation(x,name=model_name+'_relu1',act_type='relu')
x = Pooling(x,kernel=(3,3),stride=(2, 2),pool_type='avg',global_pool=True,name=model_name+'_pool1')
x = Flatten(x,name=model_name+'_flatten0')
if use_att:
x = mx.sym.Concat(x,inputs[1],dim=-1)
weight = mx.sym.Variable(model_name+'_concat_dense_weight',shape=(classes,4096))
x = FullyConnected(x,num_hidden=classes,weight=weight,no_bias=True,name=model_name+'_dense1')
else:
x = FullyConnected(x,num_hidden=classes,name=model_name+'_dense1')
return x
def ResNet18_V2(inputs,classes=1000,model_name='resnetv20'):
"""Instantiates the ResNet18 architecture.
"""
x = inputs
x = Norm(x,fix_gamma=True,use_global_stats=False,eps=1e-5,name=model_name+'_batchnorm0')
x = Convolution(x,
num_filter=64, kernel=(7, 7), stride=(2, 2), pad=(3,3), no_bias=True, name=model_name+'_conv0')
x = Norm(x,fix_gamma=False,use_global_stats=False,eps=1e-5,name=model_name+'_batchnorm1')
x = Activation(x,name=model_name+'_relu0',act_type='relu')
x = Pooling(x,kernel=(3,3),stride=(2,2),pad=(1,1),pool_type='max',name=model_name+'_pool0')
x = conv_block(x, 3, [64, 64], stage=1, block=0, strides=(1, 1),short_connect=False,model_name=model_name)
x = identity_block(x, 3, [64, 64], stage=1, block=2,model_name=model_name)
x = conv_block(x, 3, [128, 128], stage=2, block=0,model_name=model_name)
x = identity_block(x, 3, [128, 128], stage=2, block=3,model_name=model_name)
x = conv_block(x, 3, [256, 256], stage=3, block=0,model_name=model_name)
x = identity_block(x, 3, [256, 256], stage=3, block=3,model_name=model_name)
x = conv_block(x, 3, [512, 512], stage=4, block=0, model_name=model_name)
x = identity_block(x, 3, [512, 512], stage=4, block=3, model_name=model_name)
x = Norm(x,fix_gamma=False,use_global_stats=False,eps=1e-5,name=model_name+'_batchnorm2')
x = Activation(x,name=model_name+'_relu1',act_type='relu')
x = Pooling(x,kernel=(3,3),stride=(2, 2),pool_type='avg',global_pool=True,name=model_name+'_pool1')
x = Flatten(x,name=model_name+'_flatten0')
xl2 = mx.sym.norm(x,axis=1,keepdims=True)
xl2 = mx.sym.clip(xl2,0,8)
x1 = mx.sym.L2Normalization(x)
x1 = mx.sym.broadcast_mul(x1,xl2)
weight = mx.sym.Variable(model_name+'_dense1_weight')
bias = mx.sym.Variable(model_name+'_dense1_bias')
output = FullyConnected(x,num_hidden=classes,weight=weight,bias=bias,name=model_name+'_dense1')
# weight1 = mx.sym.L2Normalization(weight)
output1 = FullyConnected(x1,num_hidden=classes,weight=weight,bias=bias,name=model_name+'_dense2')
return output,output1
#
def atten_net(inputs, classes=200, model_name='resnet50_v10'):
"""Instantiates the ResNet50 architecture.
"""
x = inputs
x = Convolution(x,
num_filter=64,
kernel=(7, 7),
stride=(2, 2),
pad=(3, 3),
no_bias=True,
name=model_name + '_conv0')
x = Norm(x,
fix_gamma=False,
use_global_stats=False,
eps=1e-5,
name=model_name + '_batchnorm0')
x = Activation(x, name=model_name + '_relu0', act_type='relu')
x = Pooling(x,
kernel=(3, 3),
stride=(2, 2),
pad=(1, 1),
pool_type='max',
name=model_name + '_pool0')
x = conv_block(x,
3, [64, 64, 256],
stage=1,
block=0,
strides=(1, 1),
model_name=model_name)
x = identity_block(x,
3, [64, 64, 256],
stage=1,
block=4,
model_name=model_name)
x = identity_block(x,
3, [64, 64, 256],
stage=1,
block=7,
model_name=model_name)
x = conv_block(x,
3, [128, 128, 512],
stage=2,
block=0,
model_name=model_name)
x = identity_block(x,
3, [128, 128, 512],
stage=2,
block=4,
model_name=model_name)
x = identity_block(x,
3, [128, 128, 512],
stage=2,
block=7,
model_name=model_name)
x = identity_block(x,
3, [128, 128, 512],
stage=2,
block=10,
model_name=model_name)
x = conv_block(x,
3, [256, 256, 1024],
stage=3,
block=0,
strides=(1, 1),
dilate=(2, 2),
model_name=model_name)
x = identity_block(x,
3, [256, 256, 1024],
stage=3,
block=4,
dilate=(2, 2),
model_name=model_name)
x = identity_block(x,
3, [256, 256, 1024],
stage=3,
block=7,
dilate=(2, 2),
model_name=model_name)
x = identity_block(x,
3, [256, 256, 1024],
stage=3,
block=10,
dilate=(2, 2),
model_name=model_name)
x = identity_block(x,
3, [256, 256, 1024],
stage=3,
block=13,
dilate=(2, 2),
model_name=model_name)
x = identity_block(x,
3, [256, 256, 1024],
stage=3,
block=16,
dilate=(2, 2),
model_name=model_name)
x = conv_block(x,
3, [512, 512, 2048],
stage=4,
block=0,
strides=(1, 1),
dilate=(2, 2),
model_name=model_name)
x = identity_block(x,
3, [512, 512, 2048],
stage=4,
block=4,
dilate=(1, 1),
model_name=model_name)
x = identity_block(x,
3, [512, 512, 2048],
stage=4,
block=7,
dilate=(1, 1),
model_name=model_name)
att_fea = x
x = Pooling(x,
kernel=(3, 3),
stride=(2, 2),
pool_type='avg',
global_pool=True,
name=model_name + '_pool1')
x = flatten(x)
weight = mx.sym.Variable(model_name + '_dense1_weight')
x = FullyConnected(x,
num_hidden=classes,
weight=weight,
name=model_name + '_dense1')
# x = SoftmaxOutput(x, name = 'softmax')
return x, att_fea, weight
def ResNet50_V2(inputs, classes=1000, batch_size=144, model_name='resnetv20'):
"""Instantiates the ResNet50 architecture.
"""
x = inputs[0]
lam = inputs[1]
x = Norm(x,
fix_gamma=True,
use_global_stats=False,
eps=1e-5,
name=model_name + '_batchnorm0')
x = Convolution(x,
num_filter=64,
kernel=(7, 7),
stride=(2, 2),
pad=(3, 3),
no_bias=True,
name=model_name + '_conv0')
x = Norm(x,
fix_gamma=False,
use_global_stats=False,
eps=1e-5,
name=model_name + '_batchnorm1')
x = Activation(x, name=model_name + '_relu0', act_type='relu')
x = Pooling(x,
kernel=(3, 3),
stride=(2, 2),
pad=(1, 1),
pool_type='max',
name=model_name + '_pool0')
x = conv_block(x,
3, [64, 64, 256],
stage=1,
block=0,
strides=(1, 1),
model_name=model_name)
x = identity_block(x,
3, [64, 64, 256],
stage=1,
block=3,
model_name=model_name)
x = identity_block(x,
3, [64, 64, 256],
stage=1,
block=6,
model_name=model_name)
x_f = mx.sym.slice(x,
begin=(None, None, None, None),
end=(None, None, None, None),
step=(-1, 1, 1, 1))
lam1 = mx.sym.slice(lam, begin=(0), end=(1))
lam2 = mx.sym.slice(lam, begin=(1), end=(2))
x_mix = mx.sym.broadcast_mul(x, lam1) + broadcast_mul(x_f, lam2)
x = mx.sym.concat(x, x_mix, dim=0)
x = conv_block(x,
3, [128, 128, 512],
stage=2,
block=0,
model_name=model_name)
x = identity_block(x,
3, [128, 128, 512],
stage=2,
block=3,
model_name=model_name)
x = identity_block(x,
3, [128, 128, 512],
stage=2,
block=6,
model_name=model_name)
x = identity_block(x,
3, [128, 128, 512],
stage=2,
block=9,
model_name=model_name)
x = conv_block(x,
3, [256, 256, 1024],
stage=3,
block=0,
model_name=model_name)
x = identity_block(x,
3, [256, 256, 1024],
stage=3,
block=3,
model_name=model_name)
x = identity_block(x,
3, [256, 256, 1024],
stage=3,
block=6,
model_name=model_name)
x = identity_block(x,
3, [256, 256, 1024],
stage=3,
block=9,
model_name=model_name)
x = identity_block(x,
3, [256, 256, 1024],
stage=3,
block=12,
model_name=model_name)
x = identity_block(x,
3, [256, 256, 1024],
stage=3,
block=15,
model_name=model_name)
x = conv_block(x,
3, [512, 512, 2048],
stage=4,
block=0,
model_name=model_name)
x = identity_block(x,
3, [512, 512, 2048],
stage=4,
block=3,
model_name=model_name)
x = identity_block(x,
3, [512, 512, 2048],
stage=4,
block=6,
model_name=model_name)
x = Norm(x,
fix_gamma=False,
use_global_stats=False,
eps=1e-5,
name=model_name + '_batchnorm2')
x = Activation(x, name=model_name + '_relu1', act_type='relu')
x = Pooling(x,
kernel=(3, 3),
stride=(2, 2),
pool_type='avg',
global_pool=True,
name=model_name + '_pool1')
x = Flatten(x, name=model_name + '_flatten0')
# x = mx.sym.L2Normalization(x)
weight = mx.sym.Variable(model_name + '_dense_weight')
# weight = mx.sym.L2Normalization(weight)
output = FullyConnected(x,
num_hidden=classes,
weight=weight,
no_bias=True,
name=model_name + '_dense1')
#
output1 = mx.sym.slice_axis(output, axis=0, begin=0, end=batch_size)
mix_output = mx.sym.slice_axis(output,
axis=0,
begin=batch_size,
end=2 * batch_size)
output1_f = mx.sym.slice(output1,
begin=(None, None),
end=(None, None),
step=(-1, 1))
output_mix = mx.sym.broadcast_mul(output1, lam1) + broadcast_mul(
output1_f, lam2)
# x = mx.sym.L2Normalization(x)
# weight = mx.sym.Variable(model_name+'_dense1_weight',shape=(classes,2048))
# weight = mx.sym.L2Normalization(weight)
# alpha_w = mx.sym.Variable(model_name+'_alpha',shape=(1,1),init=mx.init.Constant(3))
# alpha_w = mx.sym.clip(alpha_w,2,6)
# weight = mx.sym.broadcast_mul(weight, alpha_w)
# x = mx.sym.broadcast_mul(x, alpha_w+1)
# weight = weight * alpha
# x = x * (alpha+2)
# weight = mx.sym.expand_dims(weight,0)
# x = mx.sym.expand_dims(x,1)
# x = mx.sym.tile(x,(1,classes,1))
# dis = mx.sym.broadcast_minus(x,weight)
# x = -mx.sym.sum(dis*dis,axis=2)
# x = x / alpha
return output1, output_mix, mix_output
#inputs = mx.sym.Variable('data')
#lam = mx.sym.Variable('lam')
#outputs = ResNet50_V2([inputs,lam],200,batch_size=144)
#net = mx.gluon.SymbolBlock(outputs,[inputs,lam])
##import gluoncv
##gluoncv.utils.viz.plot_network(net,shape=)
#
#a = mx.viz.plot_network(outputs[1], shape={'data':(144,3,224,224),'lam':(2,)},save_format='png')
#a.view('12')