def prepare_model(self):
l1_dense = sym.FullyConnected(data=sym.Variable("data"),
num_hidden=80,
no_bias=False,
name="l1_dense")
l2_dense = sym.FullyConnected(data=l1_dense,
num_hidden=120,
no_bias=True,
name="l2_dense")
l2_activation = sym.Activation(data=l2_dense,
act_type="relu",
name="l2_activation")
l3_dense = sym.FullyConnected(data=l2_activation,
num_hidden=60,
no_bias=True,
name="l3_dense")
l3_activation = sym.Activation(data=l3_dense,
act_type="relu",
name="l3_activation")
l4_dense = sym.FullyConnected(data=l3_activation,
num_hidden=20,
no_bias=True,
name="l4_dense")
l4_activation = sym.Activation(data=l4_dense,
act_type="relu",
name="l4_activation")
self.l5_dense = sym.FullyConnected(data=l4_activation,
num_hidden=1,
no_bias=False,
name="l5_dense")
output = sym.MAERegressionOutput(data=self.l5_dense,
label=sym.Variable("target"))
self.train_module = mod.Module(symbol=output,
data_names=["data"],
label_names=["target"])
def get_symbol(num_classes=10, flag='training', add_stn=False, **kwargs):
data = sym.Variable('data')
if add_stn:
data = sym.SpatialTransformer(data=data,
loc=get_loc(data),
target_shape=(28, 28),
transform_type="affine",
sampler_type="bilinear")
# first conv
conv1 = sym.Convolution(data=data, kernel=(5, 5), num_filter=10)
relu1 = sym.Activation(data=conv1, act_type="relu")
pool1 = sym.Pooling(data=relu1,
pool_type="max",
kernel=(2, 2),
stride=(2, 2))
# second conv
conv2 = sym.Convolution(data=pool1, kernel=(5, 5), num_filter=20)
relu2 = sym.Activation(data=conv2, act_type="relu")
pool2 = sym.Pooling(data=relu2,
pool_type="max",
kernel=(2, 2),
stride=(2, 2))
drop1 = mx.sym.Dropout(data=pool2)
# first fullc
flatten = sym.Flatten(data=drop1)
fc1 = sym.FullyConnected(data=flatten, num_hidden=50)
relu3 = sym.Activation(data=fc1, act_type="relu")
# second fullc
drop2 = mx.sym.Dropout(data=relu3, mode=flag)
fc2 = sym.FullyConnected(data=drop2, num_hidden=num_classes)
# loss
lenet = sym.SoftmaxOutput(data=fc2, name='softmax')
return lenet
def get_loc(data, attr={'lr_mult': '0.01'}):
"""
the localisation network in lenet-stn, it will increase acc about more than 1%,
when num-epoch >=15
"""
loc = sym.Convolution(data=data,
num_filter=8,
kernel=(7, 7),
stride=(1, 1))
loc = sym.Activation(data=loc, act_type='relu')
loc = sym.Pooling(data=loc, kernel=(2, 2), stride=(2, 2), pool_type='max')
loc = sym.Convolution(data=loc,
num_filter=10,
kernel=(5, 5),
stride=(1, 1))
loc = sym.Activation(data=loc, act_type='relu')
loc = sym.Pooling(data=loc, kernel=(2, 2), stride=(2, 2), pool_type='max')
loc = sym.FullyConnected(data=loc,
num_hidden=32,
name="stn_loc_fc1",
attr=attr)
loc = sym.Activation(data=loc, act_type='relu')
# loc = sym.Flatten(data=loc)
loc = sym.FullyConnected(data=loc,
num_hidden=6,
name="stn_loc_fc2",
attr=attr)
return loc
def hybrid_fusionFMaps(lMap, sMap, neighbor_scale=2, method='upconv'):
# lMap/sMap stand for large/small feature maps
# methods: 'upconv', 'lin_interpol'
if method == 'upconv':
raise Exception("NOT IMPLEMENTED YET.")
upconver = sym.Deconvolution(data=sMap,
kernel=upconv_ksize,
num_filter=512,
stride=(1, 1),
weight=sym.random.normal(
0, 1, (512, 512, 3, 3)))
upconver = sym.Activation(data=upconver, act_type='relu')
upconv_sMap = sym.BatchNorm(data=upconver,
gamma=sym.random_gamma(alpha=9,
beta=0.5,
shape=(2, 2)))
# upconver.initialize(ctx=mx.gpu()) # how to init? should I make the params trainable?
# TODO: Modify this. Figure out a way to deal with size problem brought by pooling
upconv_sMap = sym.UpSampling(sMap, scale=2, sample_type="nearest")
elif method == 'bilinear':
upconv_sMap = sym.UpSampling(sMap,
scale=neighbor_scale,
sample_type="nearest")
else:
raise Exception(
"ERROR! [jcy checkpoint]: Unexpected enlarging method.")
res = (lMap + upconv_sMap) / 2 # add large fmap with the smaller one
# res = sym.broadcast_div(res, sym.max(res))
return res
def _mish(x, name):
"""
mish(x) = x * tanh(softplus(x)) = x * tanh(ln(1 + e^{x}))
:param x:
:param name:
:return:
"""
return sym.elemwise_mul(x,
sym.tanh(sym.Activation(x, act_type='softrelu')),
name=f'{name}_mish')
def _activate(x, act_type: Optional[str] = None, name: Optional[str] = None):
"""
激活 x
:param x:
:param act_type: {None, 'relu', 'sigmoid', 'softrelu', 'softsign', 'tanh', 'swish', 'mish'}
:param name:
:return:
"""
if name is None:
name = 'activation'
if act_type is None:
return x
elif act_type == 'swish':
return _swish(x, name)
elif act_type == 'mish':
return _mish(x, name)
else:
return sym.Activation(x, act_type=act_type, name=f'{name}_{act_type}')
def getConvLayer(data,
name,
num_filter,
kernel_size,
pad=(0, 0),
stride=(1, 1),
activation='relu',
bn=True):
"""
return a conv layer with act, batchnorm, and pooling layers, if any
:return:
"""
convunit = sym.Convolution(data=data,
num_filter=num_filter,
pad=pad,
stride=stride,
kernel=kernel_size,
name=name + "conv")
if bn:
convunit = sym.BatchNorm(data=convunit, name=name + "bn")
convunit = sym.Activation(data=convunit,
act_type=activation,
name=name + "act")
return convunit
def _swish(x, name):
return sym.elemwise_mul(x,
sym.Activation(x, act_type='sigmoid'),
name=f'{name}_swish')
