def __init__(self, selected_model):
self.n_classes = 1
self.LeNet = Sequential([
Lambda(lambda x: (x / 255.0) - 0.5, input_shape=(160, 320, 3)),
Cropping2D(cropping=((50, 20), (0, 0)), input_shape=(160, 320, 3)),
Conv2D(filters=24,
kernel_size=(5, 5),
activation='relu',
strides=(1, 1),
padding='valid'),
MaxPooling2D((2, 2)),
Conv2D(filters=36, kernel_size=(5, 5), activation='relu'),
MaxPooling2D((2, 2)),
Conv2D(filters=48, kernel_size=(5, 5), activation='relu'),
MaxPooling2D((2, 2)),
Conv2D(filters=64, kernel_size=(1, 1), activation='relu'),
MaxPooling2D((2, 2)),
Conv2D(filters=64, kernel_size=(1, 1), activation='relu'),
MaxPooling2D((2, 2)),
Flatten(),
Dense(120, activation='relu', kernel_regularizer='l2'),
Dense(84, activation='relu'),
Dense(self.n_classes, activation='sigmoid')
])
self.nvidia = Sequential([
Lambda(lambda x: (x / 255.0) - 0.5, input_shape=(160, 320, 3)),
Cropping2D(cropping=((50, 20), (0, 0)), input_shape=(160, 320, 3)),
LayerNormalization(epsilon=0.001),
Conv2D(filters=24,
kernel_size=(5, 5),
activation='relu',
strides=(3, 3),
padding='valid'),
Conv2D(filters=36, kernel_size=(5, 5), activation='relu'),
Conv2D(filters=48, kernel_size=(5, 5), activation='relu'),
Conv2D(filters=64, kernel_size=(3, 3), activation='relu'),
Conv2D(filters=64, kernel_size=(3, 3), activation='relu'),
# Extra Layer
Conv2D(filters=64, kernel_size=(3, 3), activation='relu'),
Flatten(),
Dense(1164, activation='relu'),
Dense(100, activation='relu'),
Dense(50, activation='relu'),
Dense(10, activation='relu'),
Dense(self.n_classes)
])
self.inception = self.prepare_inception()
if selected_model.lower() == 'lenet':
self.current_model = self.LeNet
elif selected_model.lower(
) == 'nvidia' or selected_model.lower() is None:
self.current_model = self.nvidia
elif selected_model.lower() == 'inception':
self.selected_model = self.inception
else:
raise Exception(
'Please select a valid model: LeNet, Nvidia or Inception')
def yolo_conv(x_in):
if isinstance(x_in, tuple):
inputs = Input(x_in[0].shape[1:]), Input(x_in[1].shape[1:])
x, x_skip = inputs
# concat with skip connection
x = Conv2D_BN_Relu(x, filters, 1)
x = UpSampling2D(2)(x)
if (x.shape[1] != x_skip.shape[1] and x.shape[2] != x_skip.shape[2]):
x = Cropping2D(cropping=((1,0), (1,0)), input_shape = x.shape[1:])(x)
elif (x.shape[1] != x_skip.shape[1] and x.shape[2] == x_skip.shape[2]):
x = Cropping2D(cropping=((1,0), (0,0)), input_shape = x.shape[1:])(x)
elif (x.shape[1] == x_skip.shape[1] and x.shape[2] != x_skip.shape[2]):
x = Cropping2D(cropping=((0,0), (1,0)), input_shape = x.shape[1:])(x)
x = Concatenate()([x, x_skip])
else:
x = inputs = Input(x_in.shape[1:])
x = Conv2D_BN_Relu(x, filters, 1)
x = Conv2D_BN_Relu(x, filters * 2, 3)
x = Conv2D_BN_Relu(x, filters, 1)
x = Conv2D_BN_Relu(x, filters * 2, 3)
x = Conv2D_BN_Relu(x, filters, 1)
return Model(inputs, x, name=name)(x_in)
def _pad_or_crop_to_shape_2D(x, in_shape, tgt_shape):
'''
in_shape, tgt_shape are both 2x1 numpy arrays
'''
in_shape = np.asarray(in_shape)
tgt_shape = np.asarray(tgt_shape)
print('Padding input from {} to {}'.format(in_shape, tgt_shape))
im_diff = in_shape - tgt_shape
if im_diff[0] < 0:
pad_amt = (int(np.ceil(abs(im_diff[0]) / 2.0)),
int(np.floor(abs(im_diff[0]) / 2.0)))
x = ZeroPadding2D((pad_amt, (0, 0)))(x)
if im_diff[1] < 0:
pad_amt = (int(np.ceil(abs(im_diff[1]) / 2.0)),
int(np.floor(abs(im_diff[1]) / 2.0)))
x = ZeroPadding2D(((0, 0), pad_amt))(x)
if im_diff[0] > 0:
crop_amt = (int(np.ceil(im_diff[0] / 2.0)),
int(np.floor(im_diff[0] / 2.0)))
x = Cropping2D((crop_amt, (0, 0)))(x)
if im_diff[1] > 0:
crop_amt = (int(np.ceil(im_diff[1] / 2.0)),
int(np.floor(im_diff[1] / 2.0)))
x = Cropping2D(((0, 0), crop_amt))(x)
return x
def __init__(self, direction, size=1, **kwargs):
self.size = size
self.direction = direction
super(Shift, self).__init__(**kwargs)
if self.direction == "down":
self.pad = ZeroPadding2D(padding=((self.size, 0), (0, 0)), data_format="channels_last")
self.crop = Cropping2D(((0, self.size), (0, 0)))
elif self.direction == "right":
self.pad = ZeroPadding2D(padding=((0, 0), (self.size, 0)), data_format="channels_last")
self.crop = Cropping2D(((0, 0), (0, self.size)))
def resize_layer(conv, deconv):
if tf.keras.backend.image_data_format() == "channels_first":
if deconv.get_shape().as_list()[2] > conv.get_shape().as_list()[2]:
deconv = Cropping2D(cropping=((0, 1), (0, 0)))(deconv)
if deconv.get_shape().as_list()[3] > conv.get_shape().as_list()[3]:
deconv = Cropping2D(cropping=((0, 0), (0, 1)))(deconv)
else:
if deconv.get_shape().as_list()[1] > conv.get_shape().as_list()[1]:
deconv = Cropping2D(cropping=((0, 1), (0, 0)))(deconv)
if deconv.get_shape().as_list()[2] > conv.get_shape().as_list()[2]:
deconv = Cropping2D(cropping=((0, 0), (0, 1)))(deconv)
return deconv
def crop(tensor, size):
"""
Crops the given tensor to the size of the second tensor
"""
dx = int(tensor.shape[2] - size.shape[2])
dy = int(tensor.shape[1] - size.shape[1])
crop = Cropping2D(cropping=((0, dy), (0, 0)))(tensor)
crop = Cropping2D(cropping=((0, 0), (0, dx)))(crop)
return crop
def FashionMnist_classifier_full_bn():
"""
The architecture of the single-output model
"""
input_shape = (28, 28, 1)
input_img = Input(shape=input_shape, name="Input", dtype='float32')
conv1 = Conv2D(2, (3, 3), padding='same', name="conv2d_1",
trainable=True)(input_img)
conv1 = (BatchNormalization(name='batch_normalization'))(conv1)
conv1 = (Activation('relu', name='activation'))(conv1)
conv2 = Conv2D(4, (3, 3), padding='same', name="conv2d_2",
trainable=True)(conv1)
conv2 = (BatchNormalization(name='batch_normalization_1'))(conv2)
conv2 = (Activation('relu', name='activation_1'))(conv2)
conv2bis = MaxPooling2D(pool_size=(2, 2), name="max_pooling2d_1")(conv2)
conv3 = Conv2D(8, (3, 3), padding='same', name="conv2d_3",
trainable=True)(conv2bis)
conv3 = (BatchNormalization(name='batch_normalization_2'))(conv3)
conv3 = (Activation('relu', name='activation_2'))(conv3)
conv3bis = MaxPooling2D(pool_size=(2, 2), name="max_pooling2d_2")(conv3)
conv4 = Conv2D(16, (3, 3), padding='same', name="conv2d_4",
trainable=True)(conv3bis)
conv4 = (BatchNormalization(name='batch_normalization_3'))(conv4)
conv4 = (Activation('relu', name='activation_3'))(conv4)
conv4bis = MaxPooling2D(pool_size=(2, 2), name="max_pooling2d_3")(conv4)
conv5 = Conv2D(32, (3, 3), padding='same', name="conv2d_5",
trainable=True)(conv4bis)
conv5 = (BatchNormalization(name='batch_normalization_4'))(conv5)
conv5 = (Activation('relu', name='activation_4'))(conv5)
conv5bis = UpSampling2D(size=(2, 2), name='up_sampling2d_1')(conv5)
conv4tris = Cropping2D(cropping=((1, 0), (1, 0)))(conv4)
conv6 = Concatenate(name='concatenate_1', axis=3)([conv5bis, conv4tris])
conv7 = Conv2D(16, (3, 3), padding='same', name="conv2d_6",
trainable=True)(conv6)
conv7 = (BatchNormalization(name='batch_normalization_5',
trainable=True))(conv7)
conv7 = (Activation('relu', name='activation_5'))(conv7)
conv7bis = UpSampling2D(size=(2, 2), name='up_sampling2d_2')(conv7)
conv3tris = Cropping2D(cropping=((1, 1), (1, 1)))(conv3)
conv8 = Concatenate(name='concatenate_2', axis=3)([conv7bis, conv3tris])
conv9 = Conv2D(16, (3, 3), padding='same', name="conv2d_7",
trainable=True)(conv8)
conv9 = (BatchNormalization(name='batch_normalization_6',
trainable=True))(conv9)
conv9 = (Activation('relu', name='activation_6'))(conv9)
res3 = GlobalAveragePooling2D()(conv9)
res3 = Dense(10, name="fc3", trainable=True)(res3)
res3 = Activation('softmax', name='fine')(res3)
final_result = res3
model = Model(inputs=input_img, outputs=final_result)
return (model)
def createModel_Unet(row, col, depth):
# channel last , tensorflow
inputShape = (row, col, depth)
X_input = Input(inputShape)
c1 = unet_convblock(X_input, 32, (3, 3))
p1 = MaxPooling2D(pool_size=(2, 2))(c1)
c2 = unet_convblock(p1, 64, (3, 3))
p2 = MaxPooling2D(pool_size=(2, 2))(c2)
c3 = unet_convblock(p2, 128, (3, 3))
p3 = MaxPooling2D(pool_size=(2, 2))(c3)
c4 = unet_convblock(p3, 256, (3, 3))
p4 = MaxPooling2D(pool_size=(2, 2))(c4)
c5 = unet_convblock(p4, 512, (3, 3))
d6 = Conv2DTranspose(256, (3, 3), strides=(2, 2), padding='same')(c5)
crop_c4 = Cropping2D(cropping=(get_crop_size(c4, d6)))(c4)
d6 = Concatenate()([d6, crop_c4])
c6 = unet_convblock(d6, 256, (3, 3))
d7 = Conv2DTranspose(128, (3, 3), strides=(2, 2), padding='same')(c6)
crop_c3 = Cropping2D(cropping=(get_crop_size(c3, d7)))(c3)
d7 = Concatenate()([d7, crop_c3])
c7 = unet_convblock(d7, 128, (3, 3))
d8 = Conv2DTranspose(64, (3, 3), strides=(2, 2), padding='same')(c7)
crop_c2 = Cropping2D(cropping=(get_crop_size(c2, d8)))(c2)
d8 = Concatenate()([d8, crop_c2])
c8 = unet_convblock(d8, 64, (3, 3))
d9 = Conv2DTranspose(32, (3, 3), strides=(2, 2), padding='same')(c8)
crop_c1 = Cropping2D(cropping=(get_crop_size(c1, d9)))(c1)
d9 = Concatenate()([d9, crop_c1])
c9 = unet_convblock(d9, 32, (3, 3))
ch, cw = get_crop_size(c1, c9)
c9 = ZeroPadding2D(padding=((ch[0], ch[1]), (cw[0], cw[1])))(c9)
output = Conv2D(1, (1, 1), padding='same')(c9)
output = Activation('sigmoid')(output)
model = models.Model(inputs=X_input, outputs=output, name='Unet')
return model
def deep_cascade_unet(depth_str='ki',
H=218,
W=170,
Hpad=3,
Wpad=3,
kshape=(3, 3),
channels=22):
inputs = Input(shape=(H, W, channels))
mask = Input(shape=(H, W, channels))
layers = [inputs]
kspace_flag = True
for ii in depth_str:
if ii == 'i':
# Add IFFT
layers.append(Lambda(ifft_layer)(layers[-1]))
kspace_flag = False
# Add CNN block
layers.append(ZeroPadding2D(padding=(Hpad, Wpad))(layers[-1]))
layers.append(unet_block(layers[-1], kshape, channels))
layers.append(Cropping2D(cropping=(Hpad, Wpad))(layers[-1]))
# Add DC block
layers.append(
DC_block(layers[-1], mask, inputs, channels, kspace=kspace_flag))
kspace_flag = True
out = Lambda(ifft_layer)(layers[-1])
model = Model(inputs=[inputs, mask], outputs=out)
return model
def _decodeBlock(x,
shortcut,
rows_odd,
cols_odd,
cweights,
bns,
activation=LeakyReLU(alpha=ALPHA)):
#Add zero padding on bottom and right if odd dimension required at output,
#giving an output of one greater than required
x = ZeroPadding2D(padding=((0, rows_odd), (0, cols_odd)))(x)
# x = UpSampling2D(size=(2,2), interpolation=UPSAMPLE_INTERP)(x)
# up_size = np.array(x.shape)
# up_size[1] *= 2
# up_size[2] *= 2
# x = bicubic_interp_2d(x,(up_size[1],up_size[2]))
x = upsample_helper(x)
#If padding was added, crop the output to match the target shape
#print(rows_odd)
#print(cols_odd)
x = Cropping2D(cropping=((0, rows_odd), (0, cols_odd)))(x)
x = Concatenate()([shortcut, x])
x = res_Block(x, cweights, bns, activation=LeakyReLU(alpha=ALPHA))
return x
def Nvidia_model():
input_shape = (160, 320, 3)
model = Sequential()
model.add(Lambda(lambda x: x/127.5 - 1., input_shape=input_shape))
model.add(Cropping2D(cropping=((50, 20), (0, 0))))
model.add(Conv2D(24, (5, 5), strides=(2, 2), activation='elu'))
model.add(Dropout(0.5))
model.add(Conv2D(36, (5, 5), strides=(2, 2), activation='elu'))
model.add(Dropout(0.5))
model.add(Conv2D(48, (3, 3), strides=(1, 1), activation='elu'))
model.add(Dropout(0.5))
model.add(Conv2D(64, (3, 3), strides=(1, 1), activation='elu'))
model.add(Dropout(0.5))
model.add(Conv2D(64, (3, 3), strides=(1, 1), activation='elu'))
model.add(Dropout(0.5))
model.add(Flatten())
model.add(Dense(100, activation='elu'))
model.add(Dropout(0.5))
model.add(Dense(50, activation='elu'))
model.add(Dropout(0.5))
model.add(Dense(10, activation='elu'))
model.add(Dropout(0.5))
model.add(Dense(1))
model.summary()
return model
def _merger(self, net: Tensor, item: Tensor) -> Tensor:
""""Combine feature maps"""
# crop feature maps
crop_size = int(item.shape[1] - net.shape[1]) / 2
item_cropped = Cropping2D(int(crop_size))(item)
# adapt number of filters via 1x1 convolutional to allow merge
current_filters = int(net.shape[-1])
item_cropped = Conv2D(current_filters,
1,
activation=self._activation,
padding=self._padding)(item_cropped)
# Combine feature maps by adding
if self._merge_type == "add":
return Add()([item_cropped, net])
# Combine feature maps by subtracting
if self._merge_type == "subtract":
return Subtract()([item_cropped, net])
# Combine feature maps by multiplication
if self._merge_type == "multiply":
return Multiply()([item_cropped, net])
# Raise ValueError if merge type is unsupported
raise ValueError(f"unsupported merge type: {self._merge_type}")
def InceptionModel(x_train):
X_input = Input(x_train.shape)
X = Cropping2D(cropping=((60, 25), (0, 0)))(X_input)
# Re-sizes the input with Kera's Lambda layer & attach to cifar_input
X = Lambda(lambda image: tf.image.resize(image, (input_size, input_size)))(
X)
# Feeds the re-sized input into Inception model
inp = inception(X)
model = GlobalAveragePooling2D(data_format=None)(
inception.get_output_at(-1))
model = Dense(240)(model)
model = Dense(64)(model)
predictions = Dense(1, activation='relu')(model)
# Creates the model, assuming your final layer is named "predictions"
model = Model(inputs=X_input, outputs=predictions)
# Compile the model
model.compile(optimizer='Adam', loss='mse', metrics=['mse'])
# Check the summary of this new model to confirm the architecture
model.summary()
return model
def refunit(divider, ch):
image_input = Input(shape=(int(img_y / divider), int(img_x / divider), ch))
x = Conv2D(64, (7, 7), strides=(2, 2), padding='same',
name='conv1')(image_input)
x = BatchNormalization(axis=3, name='bn_conv1')(x)
x = Activation('relu')(x)
x = MaxPooling2D((3, 3))(x)
x = encoding_conv_block(x,
3, [64, 64, 256],
stage=2,
block='a',
strides=(1, 1))
x = encoding_conv_block(x, 3, [128, 128, 512], stage=3, block='a')
x = decoding_conv_block(x, 3, [512, 512, 128], stage=6, block='a')
x = decoding_conv_block(x, 3, [256, 256, 64], stage=7, block='a')
x = ZeroPadding2D(padding=(0, 1), data_format=None)(x)
x = UpSampling2D(size=(3, 3))(x)
x = Cropping2D(cropping=((2, 2), (1, 1)), data_format=None)(x)
x = Conv2DTranspose(1, (3, 3), padding='same', name='c8o')(x)
x = Activation('sigmoid')(x)
modelo = Model(inputs=image_input, outputs=x)
modelo.summary()
return modelo
def ar_model(t_skip, pred_steps, dim_c, dim_z):
"""Build CPC auto-regressive model (i.e. g_ar).
Parameters
----------
t_skip : int
number of c_t steps to crop from beginning of output seqence
pred_steps : int
number of future steps to predict
dim_c : int
dimension of output context vectors (c_t from fig. 1 of CPC paper)
dim_z : int
dimension of input latent vectors (z_t from fig. 1 of CPC paper)
Returns
-------
model : keras.Model
Model that expects time sequence input of shape (N, t_steps, z_dim) and
returns encoded sequence of shape (N, t_steps, pred_steps, z_dim). N is
the batch dimension.
"""
model = keras.Sequential(name='ar')
model.add(GRU(dim_c, return_sequences=True))
model.add(Lambda(lambda x: K.expand_dims(x, axis=2)))
model.add(Conv2DTranspose(filters=dim_z, kernel_size=(1, pred_steps)))
model.add(Cropping2D(((t_skip, pred_steps + 1), (0, 0))))
return model
def padded_unet(config):
unet_start_neurons = config['unet_start_neurons']
unet_dropout_ratio = config['unet_dropout_ratio']
input_shape, num_output_channels = utils.get_input_output_shapes(config)
inputs = [Input(input_shape, dtype='float32', name='input_array')]
x = inputs[0]
# Pad the inputs with zeros
x = ZeroPadding2D()(x)
x = Cropping2D(((1, 0), (1, 0)))(x)
# Unet
x = unet_core(x, unet_start_neurons, unet_dropout_ratio)
# MLP layers
mlp_layers = config['action_mlp_layers'] + [num_output_channels]
for i, layer_size in enumerate(mlp_layers):
if i < (len(mlp_layers) - 1):
# Non final fully connected layers
x = Dense(layer_size, activation='linear')(x)
x = Activation('relu')(x)
else:
# Sigmoid activation on the final Q-value function output layer
x = Dense(layer_size, activation='linear')(x)
outputs = [Activation('sigmoid')(x)]
return inputs, outputs
def cond2d_arch_decoder(encoded,
decoded_dim_x,
decoded_dim_y,
n_conv,
x_shape,
activation = 'relu',
filters = 4,
filter_factor = 2,
last_filters = 1,
kernel_size = (5,5),
strides = (1,1),
upsample_size = (2,2),
activity_regularizer = None,
transpose = False):
x = Dense(x_shape[1]*x_shape[2]*x_shape[3],
activation=activation)(encoded)
x = Reshape(x_shape[1:])(x)
for _ in range(n_conv-1):
x = decoder_cnn_2d(x,
filters = filters,
kernel_size = kernel_size,
strides = strides,
activation = activation,
upsample_size = upsample_size,
activity_regularizer = activity_regularizer,
transpose = transpose)
filters = int(filters//filter_factor)
decoded = decoder_cnn_2d(x,
filters = last_filters,
kernel_size = kernel_size,
strides = (1,1),
activation = activation,
upsample_size = 0,
activity_regularizer = activity_regularizer,
transpose = transpose)
d_shape = decoded.get_shape().as_list()
delta_x = d_shape[1] - decoded_dim_x
delta_y = d_shape[2] - decoded_dim_y
d_crop_x = int(delta_x/2)
d_crop_y = int(delta_y/2)
if delta_x%2==0:
d_crop_x = (d_crop_x,d_crop_x)
else:
d_crop_x = (d_crop_x,d_crop_x+1)
if delta_y%2==0:
d_crop_y = (d_crop_y,d_crop_y)
else:
d_crop_y = (d_crop_y,d_crop_y+1)
decoded = Cropping2D(cropping=(d_crop_x,d_crop_y))(decoded)
return decoded
def unet(H, W, Hpad=3, Wpad=3, kshape=(3, 3), channels=24):
"""
U-net reconstruction model. It receives as input the channel-wise zero-filled reconstruction.
Reference: Jin et al., "Deep Convolutional Neural Network for Inverse Problems in Imaging", IEEE Tran Img Proc, 2017.
` """
inputs = Input(shape=(H, W, channels))
input_padded = ZeroPadding2D(padding=(Hpad, Wpad))(
inputs) # Pad to compensate for the max-poolings
conv1 = Conv2D(64, kshape, activation='relu', padding='same')(input_padded)
conv1 = Conv2D(64, kshape, activation='relu', padding='same')(conv1)
conv1 = Conv2D(64, kshape, activation='relu', padding='same')(conv1)
pool1 = MaxPooling2D(pool_size=(2, 2))(conv1)
conv2 = Conv2D(128, kshape, activation='relu', padding='same')(pool1)
conv2 = Conv2D(128, kshape, activation='relu', padding='same')(conv2)
pool2 = MaxPooling2D(pool_size=(2, 2))(conv2)
conv3 = Conv2D(256, kshape, activation='relu', padding='same')(pool2)
conv3 = Conv2D(256, kshape, activation='relu', padding='same')(conv3)
pool3 = MaxPooling2D(pool_size=(2, 2))(conv3)
conv4 = Conv2D(512, kshape, activation='relu', padding='same')(pool3)
conv4 = Conv2D(512, kshape, activation='relu', padding='same')(conv4)
pool4 = MaxPooling2D(pool_size=(2, 2))(conv4)
conv5 = Conv2D(1024, kshape, activation='relu', padding='same')(pool4)
conv5 = Conv2D(1024, kshape, activation='relu', padding='same')(conv5)
up1 = concatenate([UpSampling2D(size=(2, 2))(conv5), conv4], axis=-1)
conv6 = Conv2D(512, kshape, activation='relu', padding='same')(up1)
conv6 = Conv2D(512, kshape, activation='relu', padding='same')(conv6)
up2 = concatenate([UpSampling2D(size=(2, 2))(conv6), conv3], axis=-1)
conv7 = Conv2D(256, kshape, activation='relu', padding='same')(up2)
conv7 = Conv2D(256, kshape, activation='relu', padding='same')(conv7)
up3 = concatenate([UpSampling2D(size=(2, 2))(conv7), conv2], axis=-1)
conv8 = Conv2D(128, kshape, activation='relu', padding='same')(up3)
conv8 = Conv2D(128, kshape, activation='relu', padding='same')(conv8)
up4 = concatenate([UpSampling2D(size=(2, 2))(conv8), conv1], axis=-1)
conv9 = Conv2D(128, kshape, activation='relu', padding='same')(up4)
conv9 = Conv2D(128, kshape, activation='relu', padding='same')(conv9)
conv10 = Conv2D(channels, (1, 1), activation='linear')(conv9)
res = Add()([conv10, input_padded]) # Residual
out = Cropping2D(cropping=(Hpad, Wpad))(
res) # Crop to go back to desired image dimensions
model = Model(inputs=inputs, outputs=out)
return model
def CAE(input_shape=(28, 28, 1), filters=[32, 64, 128, 10]):
model = Sequential()
if input_shape[0] % 8 == 0:
pad3 = 'same'
else:
pad3 = 'valid'
model.add(InputLayer(input_shape))
model.add(Conv2D(filters[0], 5, strides=2, padding='same', activation='relu', name='conv1'))
model.add(Conv2D(filters[1], 5, strides=2, padding='same', activation='relu', name='conv2'))
model.add(Conv2D(filters[2], 3, strides=2, padding=pad3, activation='relu', name='conv3'))
model.add(Flatten())
model.add(Dense(units=filters[3], name='embedding'))
model.add(Dense(units=filters[2]*int(input_shape[0]/8)*int(input_shape[0]/8), activation='relu'))
model.add(Reshape((int(input_shape[0]/8), int(input_shape[0]/8), filters[2])))
model.add(Conv2DTranspose(filters[1], 3, strides=2, padding=pad3, activation='relu', name='deconv3'))
model.add(Conv2DTranspose(filters[0], 5, strides=2, padding='same', activation='relu', name='deconv2'))
model.add(Conv2DTranspose(input_shape[2], 5, strides=2, padding='same', name='deconv1'))
if model.layers[0].input_shape != model.layers[-1].output_shape:
crop = abs(model.layers[0].input_shape[1] - model.layers[-1].output_shape[1])//2
model.add(Cropping2D(crop))
encoder = Model(inputs=model.input, outputs=model.get_layer('embedding').output)
return model, encoder
def crop_noise(noise_tensor, size, block):
"""
Crops the noise_tensor to the target size.
"""
cut = (noise_tensor.shape[1] - size) // 2
crop = Cropping2D(cut, name=f"G_Noise_Crop_block_{block}")(noise_tensor)
return crop
def __init__(self, depth):
super().__init__()
self.model_layers = []
self.model_layers.append(
ZeroPadding2D(padding=15, data_format='channels_last'))
self.model_layers.append(
Conv2D(filters=64,
kernel_size=(3, 3),
strides=(1, 1),
kernel_initializer='Orthogonal',
padding='same',
activation='relu'))
for i in range(depth - 2):
self.model_layers.append(
Conv2D(filters=64,
kernel_size=(3, 3),
strides=(1, 1),
kernel_initializer='Orthogonal',
padding='same',
use_bias=False))
self.model_layers.append(BatchNormalization())
self.model_layers.append(Activation('relu'))
self.model_layers.append(
Conv2D(filters=3,
kernel_size=(3, 3),
strides=(1, 1),
kernel_initializer='Orthogonal',
padding='same',
use_bias=False))
self.model_layers.append(
Cropping2D(cropping=15, data_format='channels_last'))
self.subtract = (Subtract())
def _upsampling_block(x,
skip,
nb_filters,
crop_px,
conv_padding,
conv_layer_type=Conv2D):
skip = Cropping2D(crop_px)(skip)
x = UpSampling2D(size=(2, 2))(x)
x = conv_layer_type(nb_filters,
2,
activation='relu',
padding='same',
kernel_initializer='he_normal')(x)
x = concatenate([skip, x], axis=3)
x = conv_layer_type(nb_filters,
3,
activation='relu',
padding=conv_padding,
kernel_initializer='he_normal')(x)
x = conv_layer_type(nb_filters,
3,
activation='relu',
padding=conv_padding,
kernel_initializer='he_normal')(x)
return x
def fcn_residual_32x_18l(num_classes, name_suffix=''):
"""
February 15
Adding residual branches
:param num_classes:
:param name_suffix:
:return:
"""
coef = 3
width = 64
input_layer = Input(shape=(None, None, 3))
x = BatchNormalization()(input_layer)
x = Conv2D(
width,
3,
**conv_args,
)(x) # Makes width wide enough for addition inside skip module
for i in range(7):
y = Cropping2D(cropping=((2, 2), (2, 2)), )(x)
x = BatchNormalization()(x)
x = Conv2D(
width,
3,
**conv_args,
)(x)
x = BatchNormalization()(x)
x = Conv2D(
width,
3,
**conv_args,
)(x)
# if i % 2 == 0:
# x = AvgPool2D(pool_size=(2, 2), strides=(1, 1), padding='same')(x)
# y = Conv2D(width, 1, **conv_args)(y) # 1x1
x = add([x, y])
x = BatchNormalization()(x)
x = Conv2D(16 * 1 << coef, 2, **conv_args)(x) # fit-once
x = BatchNormalization()(x)
x = Conv2D(16 * 1 << coef, 1, **conv_args)(x)
x = BatchNormalization()(x)
x = Conv2D(num_classes, 1, kernel_initializer=he_norm)(x) # no activation
x = Softmax()(x)
model = tf.keras.Model(inputs=input_layer,
outputs=x,
name='residual_32x_64w_18l_' + name_suffix)
return model, 32
def unet_rrn(
name,
input_shapes,
output_shapes,
kernel=3,
stride=1,
activation='elu',
output_channels=2,
kinit='RandomUniform',
batch_norm=False,
padding='same',
axis=3,
crop=0,
mpadd=0,
):
nr_classes = output_channels
timeseries, input_1_height, input_1_width, input_1_channels = input_shapes[
"input_1"]
timeseries_mask_shape = input_shapes["input_2"]
inputs = Input(
(timeseries, input_1_height, input_1_width, input_1_channels))
print("ZZZZZZZZZZZZZZZZZZZZZ", timeseries_mask_shape)
mask = Input(timeseries_mask_shape)
#masks = Input(timeseries_mask_shape)
# Encoding
conv1_output, conv1_output_last, state1_h, state1_c, pool1 = encode_block(
32, inputs, kernel, stride, activation, kinit, padding, mask=mask)
conv2_output, conv2_output_last, state2_h, state2_c, pool2 = encode_block(
64, pool1, kernel, stride, activation, kinit, padding, mask=mask)
conv3_output, conv3_output_last, state3_h, state3_c, pool3 = encode_block(
128, pool2, kernel, stride, activation, kinit, padding, mask=mask)
conv4_output, conv4_output_last, state4_h, state4_c, pool4 = encode_block(
256, pool3, kernel, stride, activation, kinit, padding, mask=mask)
# Middle
conv5_output, conv5_output_last, state5_h, state5_c = encode_block(
512, pool4, kernel, stride, activation, kinit, padding, max_pool=False)
# Decoding
conv6 = conv_t_block(256, conv5_output_last, conv4_output_last, kernel,
stride, activation, kinit, padding, axis)
conv7 = conv_t_block(128, conv6, conv3_output_last, kernel, stride,
activation, kinit, padding, axis)
conv8 = conv_t_block(64, conv7, conv2_output_last, kernel, stride,
activation, kinit, padding, axis)
conv9 = conv_t_block(32, conv8, conv1_output_last, kernel, stride,
activation, kinit, padding, axis)
# Output
conv9 = BatchNormalization()(conv9) if batch_norm else conv9
conv9 = Cropping2D((mpadd, mpadd))(conv9)
conv10 = Convolution2D(nr_classes, (1, 1),
activation='softmax',
name="output_1")(conv9)
model = Model(inputs=[inputs, mask], outputs=[conv10])
return model
def encoder_decoder_model(ds):
inp = Input(shape=(None, None, 1))
pad_x = np.ceil(
ds[0][0]['width'].numpy()[0] / 8) * 8 - ds[0][0]['width'].numpy()[0]
pad_y = np.ceil(
ds[0][0]['height'].numpy()[0] / 8) * 8 - ds[0][0]['height'].numpy()[0]
x = ZeroPadding2D(((0, int(pad_x)), (0, int(pad_y))))(inp)
x = Conv2D(16, kernel_size=(3, 3), padding='same', activation='sigmoid')(x)
x = MaxPooling2D((2, 2), padding='same')(x)
x = Conv2D(8, kernel_size=(3, 3), padding='same', activation='sigmoid')(x)
x = MaxPooling2D((2, 2), padding='same')(x)
x = Conv2D(8, kernel_size=(3, 3), padding='same', activation='sigmoid')(x)
encoded = MaxPooling2D((2, 2), padding='same')(x)
x = Conv2D(8, kernel_size=(3, 3), activation='sigmoid',
padding='same')(encoded)
x = UpSampling2D((2, 2))(x)
x = Conv2D(8, kernel_size=(3, 3), activation='sigmoid', padding='same')(x)
x = UpSampling2D((2, 2))(x)
x = Conv2D(16, kernel_size=(3, 3), activation='sigmoid', padding='same')(x)
x = UpSampling2D((2, 2))(x)
x = Cropping2D(((0, int(pad_x)), (0, int(pad_y))))(x)
decoded = Conv2D(1,
kernel_size=(3, 3),
activation='sigmoid',
padding='same')(x)
return inp, encoded, decoded
def get_lenet_model():
"""
LeNet architecture
:return: keras model
"""
inputs = Input(shape=(160, 320, 3)) # input as received from simulation
x = Lambda(lambda img: img / 255.0 - 0.5)(inputs) # normalization
x = Cropping2D(cropping=((70, 25), (0, 0)))(
x) # cropping out top and bottom of the image
x = Conv2D(filters=6, kernel_size=(5, 5), activation="relu")(x)
x = MaxPooling2D(pool_size=(2, 2))(x)
x = Conv2D(filters=16, kernel_size=(5, 5), activation="relu")(x)
x = MaxPooling2D(pool_size=(2, 2))(x)
x = Flatten()(x)
x = Dense(120, activation='relu')(x)
x = Dropout(0.5)(x)
x = Dense(84, activation='relu')(x)
x = Dropout(0.5)(x)
out = Dense(1, activation='linear')(x)
model = Model(inputs=inputs, outputs=out)
model.compile(loss='mse',
optimizer=tf.keras.optimizers.Adam(learning_rate=1e-4))
model.summary()
# needs graphviz installed
tf.keras.utils.plot_model(model, to_file='data/lenet.png')
return model
def vanilla_unet(in_shape):
input = Input((in_shape, in_shape, 3))
x = input
# Downsampling path
down_layers = []
filters = 64
for _ in range(4):
x = conv2d_block(x, filters)
down_layers.append(x)
x = MaxPooling2D((2, 2), strides=2)(x)
filters *= 2 # Number of filters doubled with each layer
x = conv2d_block(x, filters)
for conv in reversed(down_layers):
filters //= 2
x = Conv2DTranspose(filters, (2, 2), strides=(2, 2), padding='same')(x)
ch, cw = crop_shape(get_shape(conv), get_shape(x))
conv = Cropping2D((ch, cw))(conv)
x = concatenate([x, conv])
x = conv2d_block(x, filters)
output = Conv2D(1, (1, 1), activation='sigmoid')(x)
return Model(input, output)
def _adjust_block(p, ip, filters, weight_decay=5e-5, id=None):
'''
Adjusts the input `p` to match the shape of the `input`
or situations where the output number of filters needs to
be changed
# Arguments:
p: input tensor which needs to be modified
ip: input tensor whose shape needs to be matched
filters: number of output filters to be matched
weight_decay: l2 regularization weight
id: string id
# Returns:
an adjusted Keras tensor
'''
channel_dim = 1 if K.image_data_format() == 'channels_first' else -1
img_dim = 2 if K.image_data_format() == 'channels_first' else -2
with K.name_scope('adjust_block'):
if p is None:
p = ip
elif p._keras_shape[img_dim] != ip._keras_shape[img_dim]:
with K.name_scope('adjust_reduction_block_%s' % id):
p = Activation('relu', name='adjust_relu_1_%s' % id)(p)
p1 = AveragePooling2D((1, 1), strides=(2, 2), padding='valid',
name='adjust_avg_pool_1_%s' % id)(p)
p1 = Conv2D(filters // 2, (1, 1), padding='same', use_bias=False,
kernel_regularizer=l2(weight_decay),
name='adjust_conv_1_%s' % id,
kernel_initializer='he_normal')(p1)
p2 = ZeroPadding2D(padding=((0, 1), (0, 1)))(p)
p2 = Cropping2D(cropping=((1, 0), (1, 0)))(p2)
p2 = AveragePooling2D((1, 1), strides=(2, 2), padding='valid',
name='adjust_avg_pool_2_%s' % id)(p2)
p2 = Conv2D(filters // 2, (1, 1), padding='same', use_bias=False,
kernel_regularizer=l2(weight_decay),
name='adjust_conv_2_%s' % id,
kernel_initializer='he_normal')(p2)
p = concatenate([p1, p2], axis=channel_dim)
p = BatchNormalization(axis=channel_dim, momentum=_BN_DECAY,
epsilon=_BN_EPSILON,
name='adjust_bn_%s' % id)(p)
elif p._keras_shape[channel_dim] != filters:
with K.name_scope('adjust_projection_block_%s' % id):
p = Activation('relu')(p)
p = Conv2D(filters, (1, 1), strides=(1, 1), padding='same',
name='adjust_conv_projection_%s' % id, use_bias=False,
kernel_regularizer=l2(weight_decay),
kernel_initializer='he_normal')(p)
p = BatchNormalization(axis=channel_dim, momentum=_BN_DECAY,
epsilon=_BN_EPSILON,
name='adjust_bn_%s' % id)(p)
return p
def _vertical_blindspot_network(x):
""" Blind-spot network; adapted from noise2noise GitHub
Each row of output only sees input pixels above that row
"""
skips = [x]
n = x
n = _vshifted_conv(n, 48, 'enc_conv0')
n = _vshifted_conv(n, 48, 'enc_conv1')
n = _vshifted_pool(n)
skips.append(n)
n = _vshifted_conv(n, 48, 'enc_conv2')
n = _vshifted_pool(n)
skips.append(n)
n = _vshifted_conv(n, 48, 'enc_conv3')
n = _vshifted_pool(n)
skips.append(n)
n = _vshifted_conv(n, 48, 'enc_conv4')
n = _vshifted_pool(n)
skips.append(n)
n = _vshifted_conv(n, 48, 'enc_conv5')
n = _vshifted_pool(n)
n = _vshifted_conv(n, 48, 'enc_conv6')
#-----------------------------------------------
n = UpSampling2D(2)(n)
n = Concatenate(axis=3)([n, skips.pop()])
n = _vshifted_conv(n, 96, 'dec_conv5')
n = _vshifted_conv(n, 96, 'dec_conv5b')
n = UpSampling2D(2)(n)
n = Concatenate(axis=3)([n, skips.pop()])
n = _vshifted_conv(n, 96, 'dec_conv4')
n = _vshifted_conv(n, 96, 'dec_conv4b')
n = UpSampling2D(2)(n)
n = Concatenate(axis=3)([n, skips.pop()])
n = _vshifted_conv(n, 96, 'dec_conv3')
n = _vshifted_conv(n, 96, 'dec_conv3b')
n = UpSampling2D(2)(n)
n = Concatenate(axis=3)([n, skips.pop()])
n = _vshifted_conv(n, 96, 'dec_conv2')
n = _vshifted_conv(n, 96, 'dec_conv2b')
n = UpSampling2D(2)(n)
n = Concatenate(axis=3)([n, skips.pop()])
n = _vshifted_conv(n, 96, 'dec_conv1a')
n = _vshifted_conv(n, 96, 'dec_conv1b')
# final pad and crop for blind spot
n = ZeroPadding2D([[1, 0], [0, 0]])(n)
n = Cropping2D([[0, 1], [0, 0]])(n)
return n
def __call__(self, x):
h_kernel_size = self.kernel_size // 2+1
if self.mask_type == 'A':
h_kernel_size -= 1
output = ZeroPadding2D(((0, 0), (h_kernel_size, 0)))(x)
output = Convolution2D(self.filters, (1, h_kernel_size))(output)
output = Cropping2D(((0, 0), (0, 1)))(output)
return output