def forward(self, x):
print(x.shape, "is the shape of the input")
x = self.first_cn(x)
x = M.ConvLayer(1,
int(x.shape[-1] / 2),
activation=M.PARAM_LRELU,
usebias=False,
batch_norm=True)(x)
x = self.dense_1(x)
print(x.shape, "is the shape after the dense_1")
x = M.ConvLayer(1,
int(x.shape[-1] / 2),
activation=M.PARAM_LRELU,
usebias=False,
batch_norm=True)(x)
print(x.shape, "is the shape of the transition shape")
x = self.dense_2(x)
x = M.ConvLayer(1,
int(x.shape[-1] / 2),
activation=M.PARAM_LRELU,
usebias=False,
batch_norm=True)(x)
x = self.dense_3(x)
x = M.ConvLayer(1,
int(x.shape[-1] / 2),
activation=M.PARAM_LRELU,
usebias=False,
batch_norm=True)(x)
x = self.dense_4(x)
x = self.first_cn(x)
x = self.bn(x)
x = M.flatten(x)
x = tf.nn.dropout(x, 0.4)
x = self.embedding(x)
return x
def initialize(self):
self.branch3x3 = M.ConvLayer(
3,
384,
activation=M.PARAM_LRELU,
usebias=False,
batch_norm=True,
stride=2
) #M.ConvLayer(3, 384,activation=M.PARAM_LRELU, usebias=False, batch_norm=True)
self.branch3x3dbl_1 = M.ConvLayer(1,
64,
activation=M.PARAM_LRELU,
usebias=False,
batch_norm=True)
self.branch3x3dbl_2 = M.ConvLayer(3,
96,
activation=M.PARAM_LRELU,
usebias=False,
batch_norm=True)
self.branch3x3dbl_3 = M.ConvLayer(
3,
96,
activation=M.PARAM_LRELU,
usebias=False,
batch_norm=True,
stride=2) #__init__(self, size, stride, pad='SAME')
self.pool = M.MaxPool(3, 2)
def initialize(self, channels_7x7):
self.branch1x1 = M.ConvLayer(
1, 192, activation=M.PARAM_LRELU, usebias=False, batch_norm=True
) #M.ConvLayer(1, 192,activation=M.PARAM_LRELU, usebias=False, batch_norm=True)
c7 = channels_7x7
self.branch7x7_1 = M.ConvLayer(1,
c7,
activation=M.PARAM_LRELU,
usebias=False,
batch_norm=True)
self.branch7x7_2 = M.ConvLayer([1, 7],
c7,
activation=M.PARAM_LRELU,
usebias=False,
batch_norm=True)
self.branch7x7_3 = M.ConvLayer([7, 1],
192,
activation=M.PARAM_LRELU,
usebias=False,
batch_norm=True)
self.branch7x7dbl_1 = M.ConvLayer(1,
c7,
activation=M.PARAM_LRELU,
usebias=False,
batch_norm=True)
self.branch7x7dbl_2 = M.ConvLayer([7, 1],
c7,
activation=M.PARAM_LRELU,
usebias=False,
batch_norm=True)
self.branch7x7dbl_3 = M.ConvLayer([1, 7],
c7,
activation=M.PARAM_LRELU,
usebias=False,
batch_norm=True)
self.branch7x7dbl_4 = M.ConvLayer([7, 1],
c7,
activation=M.PARAM_LRELU,
usebias=False,
batch_norm=True)
self.branch7x7dbl_5 = M.ConvLayer([1, 7],
192,
activation=M.PARAM_LRELU,
usebias=False,
batch_norm=True)
self.branch_pool = M.ConvLayer(1,
192,
activation=M.PARAM_LRELU,
usebias=False,
batch_norm=True)
self.pool = M.MaxPool(3, 1) #__init__(self, size, stride, pad='SAME')
def initialize(self, outchn):
self.c1 = M.ConvLayer(1,
4 * outchn,
activation=M.PARAM_LRELU,
usebias=False,
batch_norm=True)
self.c2 = M.ConvLayer(3,
outchn,
activation=M.PARAM_LRELU,
usebias=False,
batch_norm=True)
def initialize(self, embedding_size, embedding_bn=True, outchn=0):
self.first_cn = M.ConvLayer(7,
32,
activation=M.PARAM_LRELU,
usebias=False,
batch_norm=True)
self.dense_1 = Dense_block(4, 32)
self.dense_2 = Dense_block(4, 64)
self.dense_3 = Dense_block(4, 128)
self.dense_4 = Dense_block(4, 256)
self.bn = M.BatchNorm()
self.transition = M.ConvLayer(1,
256,
activation=M.PARAM_LRELU,
usebias=False,
batch_norm=True)
#self.transition=M.ConvLayer(6, 2*outchn, activation=M.PARAM_LRELU, usebias=False, batch_norm=True)
self.embedding = M.Dense(embedding_size, batch_norm=embedding_bn)
def initialize(self, embedding_size, embedding_bn=True, outchn=0):
#super(Inception3, self).__init__()
#self.aux_logits = aux_logits
#self.transform_input = transform_input M.ConvLayer(7, 32, stride=2, activation=M.PARAM_LRELU, usebias=False, batch_norm=True)
self.Conv2d_1a_3x3 = M.ConvLayer(3,
32,
activation=M.PARAM_LRELU,
usebias=False,
batch_norm=True,
stride=2)
self.Conv2d_2a_3x3 = M.ConvLayer(3,
32,
activation=M.PARAM_LRELU,
usebias=False,
batch_norm=True)
self.Conv2d_2b_3x3 = M.ConvLayer(3,
64,
activation=M.PARAM_LRELU,
usebias=False,
batch_norm=True)
self.Conv2d_3b_1x1 = M.ConvLayer(1,
80,
activation=M.PARAM_LRELU,
usebias=False,
batch_norm=True)
self.Conv2d_4a_3x3 = M.ConvLayer(3,
192,
activation=M.PARAM_LRELU,
usebias=False,
batch_norm=True)
self.Mixed_5b = InceptionA(pool_features=32) #192
self.Mixed_5c = InceptionA(pool_features=64) #256
self.Mixed_5d = InceptionA(pool_features=64) #288
self.Mixed_6a = InceptionB() #288
self.Mixed_6b = InceptionC(channels_7x7=128) #768
self.Mixed_6c = InceptionC(channels_7x7=160)
self.Mixed_6d = InceptionC(channels_7x7=160)
self.Mixed_6e = InceptionC(channels_7x7=192)
#if aux_logits:
#self.AuxLogits = InceptionAux(768, num_classes)
self.Mixed_7a = InceptionD(768)
self.Mixed_7b = InceptionE()
self.Mixed_7c = InceptionE()
self.fc = M.Dense(embedding_size, batch_norm=embedding_bn)
def initialize(self, out, stride, shortcut=False):
self.shortcut = shortcut
self.c1 = M.ConvLayer(3,
out,
pad='SAME_LEFT',
usebias=False,
activation=M.PARAM_RELU,
batch_norm=True)
self.c2 = M.ConvLayer(3,
out,
pad='SAME_LEFT',
usebias=False,
batch_norm=True)
if shortcut:
self.sc = M.ConvLayer(1,
out,
usebias=False,
stride=stride,
batch_norm=True)
def initialize(self):
self.branch1x1 = M.ConvLayer(
1, 320, activation=M.PARAM_LRELU, usebias=False, batch_norm=True
) #M.ConvLayer(1,320,activation=M.PARAM_LRELU, usebias=False, batch_norm=True)
self.branch3x3_1 = M.ConvLayer(1,
384,
activation=M.PARAM_LRELU,
usebias=False,
batch_norm=True)
self.branch3x3_2a = M.ConvLayer([1, 3],
384,
activation=M.PARAM_LRELU,
usebias=False,
batch_norm=True)
self.branch3x3_2b = M.ConvLayer([3, 1],
384,
activation=M.PARAM_LRELU,
usebias=False,
batch_norm=True)
self.branch3x3dbl_1 = M.ConvLayer(1,
448,
activation=M.PARAM_LRELU,
usebias=False,
batch_norm=True)
self.branch3x3dbl_2 = M.ConvLayer(3,
384,
activation=M.PARAM_LRELU,
usebias=False,
batch_norm=True)
self.branch3x3dbl_3a = M.ConvLayer([1, 3],
384,
activation=M.PARAM_LRELU,
usebias=False,
batch_norm=True)
self.branch3x3dbl_3b = M.ConvLayer([3, 1],
384,
activation=M.PARAM_LRELU,
usebias=False,
batch_norm=True)
self.branch_pool = M.ConvLayer(1,
192,
activation=M.PARAM_LRELU,
usebias=False,
batch_norm=True)
self.avg_pool = M.AvgPool(3, 1)
def initialize(self, steps, inp, o):
self.mods = []
for i in range(steps):
if i == (steps - 1):
self.mods.append(
M.ConvLayer(3,
o,
stride=2,
pad='SAME_LEFT',
batch_norm=True,
usebias=False))
else:
self.mods.append(
M.ConvLayer(3,
inp,
stride=2,
pad='SAME_LEFT',
activation=M.PARAM_RELU,
batch_norm=True,
usebias=False))
def initialize(self, outchns, strides):
self.trans = []
for i, (o, s) in enumerate(zip(outchns, strides)):
if o is None or s is None:
self.trans.append(None)
elif s == 1:
self.trans.append(
M.ConvLayer(3,
o,
stride=s,
pad='SAME_LEFT',
activation=M.PARAM_RELU,
usebias=False,
batch_norm=True))
else:
self.trans.append(
M.ConvLayer(3,
o,
stride=s,
pad='SAME_LEFT',
activation=M.PARAM_RELU,
usebias=False,
batch_norm=True))
def initialize(self):
self.c1 = M.ConvLayer(3,
64,
pad='SAME_LEFT',
stride=2,
activation=M.PARAM_RELU,
usebias=False,
batch_norm=True)
self.c2 = M.ConvLayer(3,
64,
pad='SAME_LEFT',
stride=2,
activation=M.PARAM_RELU,
usebias=False,
batch_norm=True)
self.layer1 = ResBlock(256, 1, 4)
self.stage1 = Stage([32, 64], [1, 2], 1, 4, 1)
self.stage2 = Stage([32, 64, 128], [None, None, 2], 2, 4, 4)
self.stage3 = Stage([32, 64, 128, 256], [None, None, None, 2],
3,
4,
3,
d=True)
self.lastfuse = FuseLast([32, 64, 128, 256])
def initialize(self, pool_features):
self.branch1x1 = M.ConvLayer(1,
64,
activation=M.PARAM_LRELU,
usebias=False,
batch_norm=True)
self.branch5x5_1 = M.ConvLayer(1,
48,
activation=M.PARAM_LRELU,
usebias=False,
batch_norm=True)
self.branch5x5_2 = M.ConvLayer(5,
64,
activation=M.PARAM_LRELU,
usebias=False,
batch_norm=True)
self.branch3x3dbl_1 = M.ConvLayer(1,
64,
activation=M.PARAM_LRELU,
usebias=False,
batch_norm=True)
self.branch3x3dbl_2 = M.ConvLayer(1,
96,
activation=M.PARAM_LRELU,
usebias=False,
batch_norm=True)
self.branch3x3dbl_3 = M.ConvLayer(3,
96,
activation=M.PARAM_LRELU,
usebias=False,
batch_norm=True)
self.branch_pool = M.ConvLayer(
1,
pool_features,
activation=M.PARAM_LRELU,
usebias=False,
batch_norm=True
) #__init__(self, size, stride, pad='SAME') __init__(self, size, stride, pad='SAME')
self.avg_pool = M.AvgPool(3, 1)
def initialize(
self, channels_7x7
): #M.ConvLayer(1, 192,activation=M.PARAM_LRELU, usebias=False, batch_norm=True)
self.branch3x3_1 = M.ConvLayer(1,
192,
activation=M.PARAM_LRELU,
usebias=False,
batch_norm=True)
self.branch3x3_2 = M.ConvLayer(3,
320,
stride=2,
activation=M.PARAM_LRELU,
usebias=False,
batch_norm=True)
self.branch7x7x3_1 = M.ConvLayer(1,
192,
activation=M.PARAM_LRELU,
usebias=False,
batch_norm=True)
self.branch7x7x3_2 = M.ConvLayer([1, 7],
192,
activation=M.PARAM_LRELU,
usebias=False,
batch_norm=True)
self.branch7x7x3_3 = M.ConvLayer([7, 1],
192,
activation=M.PARAM_LRELU,
usebias=False,
batch_norm=True)
self.branch7x7x3_4 = M.ConvLayer(3,
192,
stride=2,
activation=M.PARAM_LRELU,
usebias=False,
batch_norm=True)
self.max_pool = M.MaxPool(3,
2) #__init__(self, size, stride, pad='SAME')
def initialize(self, num_pts):
self.backbone = hrnet.ResNet()
self.lastconv = M.ConvLayer(1, num_pts)
def initialize(self, o):
self.c1 = M.ConvLayer(1, o, batch_norm=True, usebias=False)
