def conv1d_model(shape_X, shape_Y, drp=0.3, krnl=(3, 3), dilate=0, mpool=0):
n_timesteps = shape_X[1]
n_features = shape_X[2]
n_outputs = shape_Y[1]
model = Sequential()
model.add(Conv1D(filters=16, kernel_size=krnl[0], activation='relu', input_shape=(n_timesteps, n_features)))
if not dilate:
model.Name = "1D TCN"
model.add(Conv1D(filters=16, kernel_size=krnl[1], activation='relu'))
else:
model.Name = "1D {} dilated TCN".format(dilate)
model.add(Conv1D(filters=16, kernel_size=krnl[1], activation='relu', dilation_rate=dilate))
if mpool:
model.add(MaxPooling1D(pool_size=mpool))
model.add(Dropout(drp))
model.add(Flatten())
model.add(Dense(100, activation='relu'))
model.add(Dense(50, activation='relu'))
model.add(Dense(n_outputs, activation='linear'))
model.compile(loss='mean_squared_error', optimizer='adam', metrics=['mape'])
return model
def depthwise_model(shape_X, shape_Y, drp=0.3, krnl=(3, 3), dilate=0, mpool=0):
n_timesteps = shape_X[2]
n_features = shape_X[3]
n_outputs = shape_Y[1]
model = Sequential()
model.add(DepthwiseConv2D(input_shape=(1, n_timesteps, n_features),
kernel_size=(1, krnl[0]),
depth_multiplier=4,
activation='relu',
padding='valid'))
if mpool:
model.add(MaxPooling2D(pool_size=(1,mpool)))
if not dilate:
model.Name = "1D TCN"
model.add(DepthwiseConv2D(input_shape=(1, n_timesteps, n_features),
kernel_size=(1, krnl[1]),
depth_multiplier=16,
activation='relu',
padding='valid'))
else:
model.Name = "1D {} dilated TCN".format(dilate)
model.add(DepthwiseConv2D(input_shape=(1, n_timesteps, n_features),
kernel_size=(1, krnl[1]),
dilation_rate=dilate,
depth_multiplier=3,
activation='relu',
padding='valid'))
if mpool:
model.add(MaxPooling2D(pool_size=(1,mpool)))
model.add(Dropout(drp))
model.add(Flatten())
model.add(Dense(100, activation='relu'))
model.add(Dense(50, activation='relu'))
model.add(Dense(n_outputs, activation='linear'))
model.compile(loss='mean_squared_error', optimizer='adam', metrics=['mape'])
return model
