def CapsNetR3(input_shape, n_class=2):
x = layers.Input(shape=input_shape)
# Layer 1: Just a conventional Conv2D layer
conv1 = layers.Conv2D(filters=16,
kernel_size=5,
strides=1,
padding='same',
activation='relu',
name='conv1')(x)
# Reshape layer to be 1 capsule x [filters] atoms
_, H, W, C = conv1.get_shape()
# print("conv1 params",conv1.get_shape())
conv1_reshaped = layers.Reshape((H.value, W.value, 1, C.value))(conv1)
# Layer 1: Primary Capsule: Conv cap with routing 1
primary_caps = ConvCapsuleLayer(kernel_size=5,
num_capsule=2,
num_atoms=16,
strides=2,
padding='same',
routings=1,
name='primarycaps')(conv1_reshaped)
# Layer 2: Convolutional Capsule
conv_cap_2_1 = ConvCapsuleLayer(kernel_size=5,
num_capsule=4,
num_atoms=16,
strides=1,
padding='same',
routings=3,
name='conv_cap_2_1')(primary_caps)
# Layer 2: Convolutional Capsule
conv_cap_2_2 = ConvCapsuleLayer(kernel_size=5,
num_capsule=4,
num_atoms=32,
strides=2,
padding='same',
routings=3,
name='conv_cap_2_2')(conv_cap_2_1)
# Layer 3: Convolutional Capsule
conv_cap_3_1 = ConvCapsuleLayer(kernel_size=5,
num_capsule=8,
num_atoms=32,
strides=1,
padding='same',
routings=3,
name='conv_cap_3_1')(conv_cap_2_2)
# Layer 3: Convolutional Capsule
conv_cap_3_2 = ConvCapsuleLayer(kernel_size=5,
num_capsule=8,
num_atoms=64,
strides=2,
padding='same',
routings=3,
name='conv_cap_3_2')(conv_cap_3_1)
# Layer 4: Convolutional Capsule
conv_cap_4_1 = ConvCapsuleLayer(kernel_size=5,
num_capsule=8,
num_atoms=32,
strides=1,
padding='same',
routings=3,
name='conv_cap_4_1')(conv_cap_3_2)
# Layer 1 Up: Deconvolutional Capsule
deconv_cap_1_1 = DeconvCapsuleLayer(kernel_size=4,
num_capsule=8,
num_atoms=32,
upsamp_type='deconv',
scaling=2,
padding='same',
routings=3,
name='deconv_cap_1_1')(conv_cap_4_1)
# Skip connection
up_1 = layers.Concatenate(axis=-2,
name='up_1')([deconv_cap_1_1, conv_cap_3_1])
# Layer 1 Up: Deconvolutional Capsule
deconv_cap_1_2 = ConvCapsuleLayer(kernel_size=5,
num_capsule=4,
num_atoms=32,
strides=1,
padding='same',
routings=3,
name='deconv_cap_1_2')(up_1)
# Layer 2 Up: Deconvolutional Capsule
deconv_cap_2_1 = DeconvCapsuleLayer(kernel_size=4,
num_capsule=4,
num_atoms=16,
upsamp_type='deconv',
scaling=2,
padding='same',
routings=3,
name='deconv_cap_2_1')(deconv_cap_1_2)
# Skip connection
up_2 = layers.Concatenate(axis=-2,
name='up_2')([deconv_cap_2_1, conv_cap_2_1])
# Layer 2 Up: Deconvolutional Capsule
deconv_cap_2_2 = ConvCapsuleLayer(kernel_size=5,
num_capsule=4,
num_atoms=16,
strides=1,
padding='same',
routings=3,
name='deconv_cap_2_2')(up_2)
# Layer 3 Up: Deconvolutional Capsule
deconv_cap_3_1 = DeconvCapsuleLayer(kernel_size=4,
num_capsule=2,
num_atoms=16,
upsamp_type='deconv',
scaling=2,
padding='same',
routings=3,
name='deconv_cap_3_1')(deconv_cap_2_2)
# Skip connection
up_3 = layers.Concatenate(axis=-2,
name='up_3')([deconv_cap_3_1, conv1_reshaped])
# Layer 4: Convolutional Capsule: 1x1
seg_caps = ConvCapsuleLayer(kernel_size=1,
num_capsule=1,
num_atoms=16,
strides=1,
padding='same',
routings=3,
name='seg_caps')(up_3)
# Layer 4: This is an auxiliary layer to replace each capsule with its length. Just to match the true label's shape.
out_seg = Length(num_classes=n_class, seg=True, name='out_seg')(seg_caps)
# Decoder network.
_, H, W, C, A = seg_caps.get_shape()
y = layers.Input(shape=input_shape[:-1] + (1, ))
# print(y)
masked_by_y = Mask()(
[seg_caps, y]
) # The true label is used to mask the output of capsule layer. For training
masked = Mask()(
seg_caps) # Mask using the capsule with maximal length. For prediction
# print(masked_by_y, masked)
def shared_decoder(mask_layer):
recon_remove_dim = layers.Reshape(
(H.value, W.value, A.value))(mask_layer)
recon_1 = layers.Conv2D(filters=64,
kernel_size=1,
padding='same',
kernel_initializer='he_normal',
activation='relu',
name='recon_1')(recon_remove_dim)
recon_2 = layers.Conv2D(filters=128,
kernel_size=1,
padding='same',
kernel_initializer='he_normal',
activation='relu',
name='recon_2')(recon_1)
out_recon = layers.Conv2D(filters=1,
kernel_size=1,
padding='same',
kernel_initializer='he_normal',
activation='sigmoid',
name='out_recon')(recon_2)
return out_recon
# Models for training and evaluation (prediction)
train_model = models.Model(inputs=[x, y],
outputs=[out_seg,
shared_decoder(masked_by_y)])
eval_model = models.Model(inputs=x,
outputs=[out_seg,
shared_decoder(masked)])
# manipulate model
noise = layers.Input(shape=((H.value, W.value, C.value, A.value)))
noised_seg_caps = layers.Add()([seg_caps, noise])
masked_noised_y = Mask()([noised_seg_caps, y])
manipulate_model = models.Model(inputs=[x, y, noise],
outputs=shared_decoder(masked_noised_y))
train_model.compile(optimizer='rmsprop',
loss='binary_crossentropy',
metrics=[mean_iou])
return train_model
def CapsNetR3(input_shape, modalities=1, n_class=2):
capsules_base = 2
filter_multiplier = 1
atoms_base = 16 * filter_multiplier
x = layers.Input(shape=input_shape)
# Layer 1: Just a conventional Conv2D layer
conv1 = layers.Conv2D(filters=16 * filter_multiplier,
kernel_size=5,
strides=1,
padding='same',
activation='relu',
name='conv1')(x)
# Reshape layer to be 1 capsule x [filters] atoms
_, H, W, C = conv1.get_shape()
print(conv1.shape)
conv1_reshaped = layers.Reshape((H.value, W.value, 1, C.value))(conv1)
print(conv1_reshaped.shape)
# Layer 1: Primary Capsule: Conv cap with routing 1
primary_caps = ConvCapsuleLayer(kernel_size=5,
num_capsule=capsules_base,
num_atoms=atoms_base,
strides=2,
padding='same',
routings=1,
name='primarycaps')(conv1_reshaped)
# Layer 2: Convolutional Capsule
conv_cap_2_1 = ConvCapsuleLayer(kernel_size=5,
num_capsule=capsules_base * 2,
num_atoms=atoms_base,
strides=1,
padding='same',
routings=3,
name='conv_cap_2_1')(primary_caps)
# Layer 2: Convolutional Capsule
conv_cap_2_2 = ConvCapsuleLayer(kernel_size=5,
num_capsule=capsules_base * 2,
num_atoms=atoms_base * 2,
strides=2,
padding='same',
routings=3,
name='conv_cap_2_2')(conv_cap_2_1)
# Layer 3: Convolutional Capsule
conv_cap_3_1 = ConvCapsuleLayer(kernel_size=5,
num_capsule=capsules_base * 4,
num_atoms=atoms_base * 2,
strides=1,
padding='same',
routings=3,
name='conv_cap_3_1')(conv_cap_2_2)
# Layer 3: Convolutional Capsule
conv_cap_3_2 = ConvCapsuleLayer(kernel_size=5,
num_capsule=capsules_base * 4,
num_atoms=atoms_base * 4,
strides=2,
padding='same',
routings=3,
name='conv_cap_3_2')(conv_cap_3_1)
# Layer 4: Convolutional Capsule
conv_cap_4_1 = ConvCapsuleLayer(kernel_size=5,
num_capsule=capsules_base * 4,
num_atoms=atoms_base * 2,
strides=1,
padding='same',
routings=3,
name='conv_cap_4_1')(conv_cap_3_2)
# Layer 1 Up: Deconvolutional Capsule
deconv_cap_1_1 = DeconvCapsuleLayer(kernel_size=4,
num_capsule=capsules_base * 4,
num_atoms=atoms_base * 2,
upsamp_type='deconv',
scaling=2,
padding='same',
routings=3,
name='deconv_cap_1_1')(conv_cap_4_1)
# Skip connection
up_1 = layers.Concatenate(axis=-2,
name='up_1')([deconv_cap_1_1, conv_cap_3_1])
# Layer 1 Up: Deconvolutional Capsule
deconv_cap_1_2 = ConvCapsuleLayer(kernel_size=5,
num_capsule=capsules_base * 2,
num_atoms=atoms_base * 2,
strides=1,
padding='same',
routings=3,
name='deconv_cap_1_2')(up_1)
# Layer 2 Up: Deconvolutional Capsule
deconv_cap_2_1 = DeconvCapsuleLayer(kernel_size=4,
num_capsule=capsules_base * 2,
num_atoms=atoms_base,
upsamp_type='deconv',
scaling=2,
padding='same',
routings=3,
name='deconv_cap_2_1')(deconv_cap_1_2)
# Skip connection
up_2 = layers.Concatenate(axis=-2,
name='up_2')([deconv_cap_2_1, conv_cap_2_1])
# Layer 2 Up: Deconvolutional Capsule
deconv_cap_2_2 = ConvCapsuleLayer(kernel_size=5,
num_capsule=capsules_base * 2,
num_atoms=atoms_base,
strides=1,
padding='same',
routings=3,
name='deconv_cap_2_2')(up_2)
# Layer 3 Up: Deconvolutional Capsule
deconv_cap_3_1 = DeconvCapsuleLayer(kernel_size=4,
num_capsule=capsules_base * 2,
num_atoms=atoms_base,
upsamp_type='deconv',
scaling=2,
padding='same',
routings=3,
name='deconv_cap_3_1')(deconv_cap_2_2)
# Skip connection
up_3 = layers.Concatenate(axis=-2,
name='up_3')([deconv_cap_3_1, conv1_reshaped])
seg_caps = ConvCapsuleLayer(kernel_size=1,
num_capsule=n_class,
num_atoms=atoms_base,
strides=1,
padding='same',
routings=3,
name='seg_caps')(up_3)
# Layer 4: This is an auxiliary layer to replace each capsule with its length. Just to match the true label's shape.
out_seg = Length(num_classes=n_class, seg=True, name='out_seg')(seg_caps)
print(out_seg.shape)
#assert False, "Out seg shape"
#out_seg = seg_caps
# Decoder network.
_, H, W, C, A = seg_caps.get_shape()
print(H.value, W.value, C.value, A.value)
y = layers.Input(shape=input_shape[:-1] + (1, ), name='recon_input')
masked_by_y = Mask()(
[seg_caps, y]
) # The true label is used to mask the output of capsule layer. For training
print('masked by y ' + str(masked_by_y.shape))
print('Y ' + str(y.shape))
masked = Mask()(
seg_caps) # Mask using the capsule with maximal length. For prediction
def shared_decoder(mask_layer):
recon_remove_dim = layers.Reshape(
(input_shape[0], input_shape[1], n_class * atoms_base))(mask_layer)
recon_1 = layers.Conv2D(filters=64,
kernel_size=1,
padding='same',
kernel_initializer='he_normal',
activation='relu',
name='recon_1')(recon_remove_dim)
recon_2 = layers.Conv2D(filters=128,
kernel_size=1,
padding='same',
kernel_initializer='he_normal',
activation='relu',
name='recon_2')(recon_1)
out_recon = layers.Conv2D(filters=modalities,
kernel_size=1,
padding='same',
kernel_initializer='he_normal',
activation='sigmoid',
name='out_recon')(recon_2)
return out_recon
# Models for training and evaluation (prediction)
train_outputs = [out_seg, shared_decoder(masked_by_y)]
print("Train outputs")
print(train_outputs[0].shape)
print(train_outputs[1].shape)
#assert False
train_model = models.Model(inputs=[x, y], outputs=train_outputs)
eval_model = models.Model(inputs=x,
outputs=[
out_seg, shared_decoder(masked)
]) #TODO: Check masked by y for testing!
# manipulate model
noise = layers.Input(shape=((H.value, W.value, C.value, A.value)))
noised_seg_caps = layers.Add()([seg_caps, noise])
masked_noised_y = Mask()([noised_seg_caps, y])
manipulate_model = models.Model(inputs=[x, y, noise],
outputs=shared_decoder(masked_noised_y))
return train_model, eval_model, manipulate_model
