def model(t):
global pi
input = Input((3, 192, 192))
l = ZeroPadding2D(padding=3, data_format='channels_first')(input)
l = Conv(64, 7, strides=2, padding='valid', tobj=get(t, 0))(l)
l = BatchNorm(tobj=get(t, 1))(l)
l = Activation('relu')(l)
l = add_conv(64, get(t, 3), l)
l = MaxPool2D(data_format='channels_first')(l)
l = add_id(128, get(t, 19), l)
l = add_conv(128, get(t, 33), l)
pi = 0
print("Half 1")
l = half(get(t, 48), l)
pi = 0
print("Half 2")
l = half(get(t, 689), l)
l = Conv(256, 1, tobj=get(t, 1333))(l)
l = BatchNorm(tobj=get(t, 1334))(l)
l = Activation('relu')(l)
l = Conv(200, 1, tobj=get(t, 1336))(l)
l = UpSamplingBilinear()(l)
l = Activation('sigmoid')(l)
model = Model(input, l)
model.compile(loss='mean_squared_error', optimizer='sgd')
return model
def half(tobj, input):
top = tobj['modules'][1]
seq = part(top, input, 1)
c, d = top['modules'][-2:]
for s in (c, d):
seq = Conv(256, 1, tobj=get(s, 0))(seq)
seq = BatchNorm(tobj=get(s, 1))(seq)
seq = Activation('relu')(seq)
return Add()([input, seq])
def conv131(input_dim, inputs, tobj=None, first_layer=False):
tbn1 = get(tobj, 0)
tbn2 = get(tobj, 3)
tbn3 = get(tobj, 6)
tconv1 = get(tobj, 2)
tconv2 = get(tobj, 5)
tconv3 = get(tobj, 8)
batch_norm1 = BatchNorm(tobj=tbn1)(inputs)
relu1 = Activation('relu')(batch_norm1)
output_dim1 = input_dim if first_layer else input_dim / 2
conv1 = Conv(output_dim1, 1, tobj=tconv1)(relu1)
batch_norm2 = BatchNorm(tobj=tbn2)(conv1)
relu2 = Activation('relu')(batch_norm2)
conv2 = Conv(output_dim1, 3, tobj=tconv2)(relu2)
batch_norm3 = BatchNorm(tobj=tbn3)(conv2)
relu3 = Activation('relu')(batch_norm3)
conv3 = Conv(output_dim1 * 2, 1, tobj=tconv3)(relu3)
return conv3
def add_conv(input_dim, tobj, l):
return Add()([
conv131(input_dim, l, tobj=get(tobj, 1), first_layer=True),
Conv(input_dim * 2, 1, tobj=get(tobj, 12))(l)
])