def stem(self, inp):
'''
inception v4 stem
input: 256 x 256 x 3
output: 32 x 32 x 576
'''
xi = Input(shape=inp.get_shape().as_list()[1:]) # 256 x 256 x 3
x = layers.conv_bn_act(xi, 32, (3, 3), strides=(2, 2))
x = layers.conv_bn_act(x, 32, (3, 3))
x = layers.conv_bn_act(x, 64, (3, 3))
a = layers.conv_bn_act(x, 96, (3, 3), strides=(2, 2))
b = MaxPooling2D((3, 3), strides=(2, 2), padding='same')(x)
x = concatenate([a, b])
a = layers.conv_bn_act(x, 64, (1, 1))
a = layers.conv_bn(a, 96, (3, 3))
b = layers.conv_bn_act(x, 64, (1, 1))
b = layers.conv_bn_act(b, 64, (5, 1))
b = layers.conv_bn_act(b, 64, (1, 5))
b = layers.conv_bn(b, 96, (3, 3))
x = concatenate([a, b])
a = layers.act_conv_bn(x, 192, (3, 3), strides=(2, 2))
b = MaxPooling2D((2, 2), strides=(2, 2))(x)
x = concatenate([a, b])
x = blocks.sepconv_residual(x, 3 * 192, name='sepconv1')
model = Model(xi, x, name='Stem')
x = model(inp)
return x
def fremap_block(self, inp, num_filters, name=None):
# input_shape = inp.get_shape().as_list()[1:]
# xi = Input(shape=input_shape)
x = layers.act_conv_bn(inp, num_filters, (1, 1))
# model = Model(inputs=inp, outputs=x, name=name)
return x
def reception_block(self, inp, name):
'''
each pose block starts with a reception block
it is u-shaped and relies on separable convolutions
inp ------------------------- a (SR 576) -------------------- + -- out
| |
| |
MP --- C 288 -- SR 288 ---- b (SR 288) ---- + -- SR 576 -- US
| |
| |
MP --- SR -- SR -- SR --- US <- all 288 channels
SR: Sepconv Residual (all 5x5)
C: Conv (1x1)
MP: Max Pooling (2x2 with stride 2x2)
US: UpSampling (2x2)
input: 32 x 32 x 576
output: 32 x 32 x 576
'''
ksize = self.kernel_size
input_shape = inp.get_shape().as_list()[1:]
size = int(input_shape[-1])
# first branch
# xi = Input(shape=input_shape)
a = blocks.sepconv_residual(inp, size, name='sepconv_l1', kernel_size=ksize)
# second branch
low1 = MaxPooling2D((2, 2))(inp)
low1 = layers.act_conv_bn(low1, int(size/2), (1, 1))
low1 = blocks.sepconv_residual(low1, int(size/2), name='sepconv_l2_1', kernel_size=ksize)
b = blocks.sepconv_residual(low1, int(size/2), name='sepconv_l2_2', kernel_size=ksize)
# third branch
c = MaxPooling2D((2, 2))(low1)
c = blocks.sepconv_residual(c, int(size/2), name='sepconv_l3_1', kernel_size=ksize)
c = blocks.sepconv_residual(c, int(size/2), name='sepconv_l3_2', kernel_size=ksize)
c = blocks.sepconv_residual(c, int(size/2), name='sepconv_l3_3', kernel_size=ksize)
c = UpSampling2D((2, 2))(c)
# merge second and third branches
b = add([b, c])
b = blocks.sepconv_residual(b, size, name='sepconv_l2_3', kernel_size=ksize)
b = UpSampling2D((2, 2))(b)
# merge first and second branches
x = add([a, b])
# model = Model(inputs=xi, outputs=x, name=name)
return x
def sepconv_residual(x, out_size, name, kernel_size=(3, 3)):
'''
Separable convolution with residual
TODO: Schema
'''
shortcut_name = name + '_shortcut'
reduce_name = name + '_reduce'
num_filters = K.int_shape(x)[-1]
if num_filters == out_size:
ident = x
else:
ident = layers.act_conv_bn(x, out_size, (1, 1), name=shortcut_name)
if out_size < num_filters:
x = layers.act_conv_bn(x, out_size, (1, 1), name=reduce_name)
x = layers.separable_act_conv_bn(x, out_size, kernel_size, name=name)
x = add([ident, x])
return x
def stem(self, inp):
'''
inception v4 stem
input: 256 x 256 x 3
output: 32 x 32 x 576
'''
print("BUILDING STEM")
# xi = Input(shape=inp.get_shape().as_list()[1:]) # 256 x 256 x 3
x = layers.conv_bn_act(inp, 32, (3, 3), strides=(2, 2), name="stem1")
x = layers.conv_bn_act(x, 32, (3, 3), name="stem2")
x = layers.conv_bn_act(x, 64, (3, 3), name="stem3")
a = layers.conv_bn_act(x, 96, (3, 3), strides=(2, 2), name="stem4")
b = MaxPooling2D((3, 3), strides=(2, 2), padding='same', name="stem5")(x)
x = concatenate([a, b], name="stem6")
a = layers.conv_bn_act(x, 64, (1, 1), name="stem7")
a = layers.conv_bn(a, 96, (3, 3), name="stem8")
b = layers.conv_bn_act(x, 64, (1, 1), name="stem9")
b = layers.conv_bn_act(b, 64, (5, 1), name="stem10")
b = layers.conv_bn_act(b, 64, (1, 5), name="stem11")
b = layers.conv_bn(b, 96, (3, 3), name="stem12")
x = concatenate([a, b], name="stem13")
a = layers.act_conv_bn(x, 192, (3, 3), strides=(2, 2), name="stem14")
b = MaxPooling2D((2, 2), strides=(2, 2), name="stem15")(x)
x = concatenate([a, b], name="stem16")
x = blocks.sepconv_residual(x, 3 * 192, name='sepconv1')
# model = Model(xi, x, name='Stem')
# print("Stem model summary")
# model.summary()
# x = model(inp)
return x