def getModel(cloud_dim, precipitation_dim, wind_dim, humidity_dim, pressure_dim, temperature_dim,
ocean_dim=None, other_dim=None, target_shape=1, name="resnetLSTM"):
# input to the model
input_array = getInput(cloud_dim, precipitation_dim, wind_dim, humidity_dim, pressure_dim, temperature_dim,
ocean_dim)
if other_dim:
other_in = Input(shape=other_dim, name="OtherParam")
input_array.append(other_in)
x = concatenate_together([expandForCNN(i) for i in input_array])
for _ in range(8):
x = BatchNormalization()(x)
x = ResidualBlock(x, filters=8, kernel_size=3, strides=(1, 1), padding='same', shortcut=True)
x = MaxPooling2D(pool_size=(2, 2))(x)
x = BatchNormalization()(x)
x = Conv2D(16, kernel_size=(3, 3), padding='same', activation=relu)(x)
x = MaxPooling2D(pool_size=(2, 2))(x)
x = BatchNormalization()(x)
x = Flatten()(x) if other_dim is None else flattenAll([x, other_in])
x = BatchNormalization()(x)
# unify output layer
x = Dense(256, activation=relu)(x)
x = BatchNormalization()(x)
# last stage processing
x = LSTM(128)(expand_dims(x, -1))
pred = getOutput(x, target_shape)
m = Model(inputs=input_array, outputs=pred, name=name)
return m
def getModel(cloud_dim,
precipitation_dim,
wind_dim,
humidity_dim,
pressure_dim,
temperature_dim,
ocean_dim=None,
other_dim=None,
target_shape=1,
name="vcnnLSTM"):
# input to the model
input_array = getInput(cloud_dim, precipitation_dim, wind_dim,
humidity_dim, pressure_dim, temperature_dim,
ocean_dim)
if other_dim:
other_in = Input(shape=other_dim, name="OtherParam")
input_array.append(other_in)
# CNN
x = [
AutoSetDenseOrCNN(i,
horizontal=True,
dropout=False,
activation=relu,
padding="valid") for i in input_array
]
x1 = concatenate_together(list(filter(lambda i: len(i.shape) == 4, x)),
axis=1)
x2 = concatenate_together(list(filter(lambda i: len(i.shape) != 4, x)),
axis=1)
x1 = MaxPooling2D(pool_size=(2, 1))(Conv2D(
16, kernel_size=(2, 1), padding='same', activation=relu)(
BatchNormalization()(x1))) if x1 is not None else None
x2 = Dense(x2.shape[1] * 2, activation=relu)(
BatchNormalization()(x2)) if x2 is not None else None
if x1 is not None and x2 is not None:
x = flattenAll([x1, x2])
else:
x = Flatten()(x1) if x1 is not None else Flatten()(x2)
# unify output layer
x = BatchNormalization()(x)
x = Dense(256, activation=relu)(x)
# last stage processing
x = BatchNormalization()(x)
x = LSTM(128)(expand_dims(x, -1))
pred = getOutput(x, target_shape)
m = Model(inputs=input_array, outputs=pred, name=name)
return m
def getModel(cloud_dim, precipitation_dim, wind_dim, humidity_dim, pressure_dim, temperature_dim,
ocean_dim=None, other_dim=None, target_shape=1, name="ann"):
# input to the model
input_array = getInput(cloud_dim, precipitation_dim, wind_dim, humidity_dim, pressure_dim, temperature_dim,
ocean_dim)
if other_dim:
other_in = Input(shape=other_dim, name="OtherParam")
input_array.append(other_in)
# flatten all data and concatenate together
x = flattenAll(input_array)
x = BatchNormalization()(x)
x = Dense(512, activation=relu)(x)
x = BatchNormalization()(x)
# unify output layer
x = Dense(256, activation=relu)(x)
x = BatchNormalization()(x)
# last stage processing
x = LeakyReLU()(Dense(128)(x))
x = BatchNormalization()(x)
pred = getOutput(x, target_shape)
m = Model(inputs=input_array, outputs=pred, name=name)
return m
def getModel(cloud_dim,
precipitation_dim,
wind_dim,
humidity_dim,
pressure_dim,
temperature_dim,
ocean_dim=None,
other_dim=None,
target_shape=1,
name="resnet"):
# input to the model
input_array = getInput(cloud_dim, precipitation_dim, wind_dim,
humidity_dim, pressure_dim, temperature_dim,
ocean_dim)
if other_dim:
other_in = Input(shape=other_dim, name="OtherParam")
input_array.append(other_in)
x = concatenate_together([expandForCNN(i) for i in input_array])
for _ in range(8):
x = invertedResidualBlock(x,
in_channel=1,
out_channel=8,
kernel_size=3,
expansion_rate=8,
strides=(1, 1))
x = dsConv2D(x,
in_channel=8,
out_channel=8,
kernel_size=3,
strides=1,
padding="valid")
x = dsConv2D(x,
in_channel=8,
out_channel=16,
kernel_size=3,
strides=1,
padding="valid")
x = dsConv2D(x,
in_channel=16,
out_channel=16,
kernel_size=3,
strides=1,
padding="valid")
x = dsConv2D(x,
in_channel=16,
out_channel=32,
kernel_size=3,
strides=1,
padding="valid")
x = dsConv2D(x,
in_channel=32,
out_channel=32,
kernel_size=3,
strides=1,
padding="valid")
# x = dsConv2D(x, in_channel=32, out_channel=64, kernel_size=3, strides=1, padding="valid")
# x = dsConv2D(x, in_channel=64, out_channel=64, kernel_size=3, strides=1, padding="valid")
x = Flatten()(x)
# unify output layer
x = Dense(256, activation=relu)(x)
# last stage processing
x = LeakyReLU()(Dense(128)(x))
pred = getOutput(x, target_shape)
m = Model(inputs=input_array, outputs=pred, name=name)
return m