def transfer(source_model, target_path, frozen_layer):
# load data of branch
df = pd.read_pickle(target_path)
# create dataframe with netto sales, month, weekday, year
df = pd.DataFrame(data=df.values, index=df.index, columns=['netto'])
df = df.assign(month=df.index.month)
df = df.assign(weekday=df.index.weekday)
df = df.assign(year=df.index.year)
# split into train and test
train, test = split_dataset(df.values, 365)
# prepare input data for branch
n_input = 365
train_x, train_y = to_supervised(train, n_input, 365)
# load pre-trained model of source branch as base model
base_model = load_model(source_model)
# freeze specific layers of base model
for layer in base_model.layers[:frozen_layer]:
layer.trainable = False
print("frozen layers: " + str(frozen_layer))
# compile the model
base_model.compile(loss='mse', optimizer='adam')
# fit base_model with new data from branch
n_timesteps, n_features, n_outputs = train_x.shape[1], train_x.shape[2], train_y.shape[1]
input_data = [train_x[:, :, i].reshape((train_x.shape[0], n_timesteps, 1)) for i in range(n_features)]
base_model.fit(input_data, train_y, epochs=20, batch_size=16, verbose=0)
# evaluate fitted model
mape = evaluate_model(train, test, base_model)
return mape
def build_model(train):
# model parameters
n_input = 365
epochs = 20
batch_size = 16
filters = 64
kernel_size = 8
pool_size = 4
dense_1 = 500
dense_2 = 200
# prepare data
train_x, train_y = to_supervised(train, n_input, 365)
# define parameters
verbose = 2
n_timesteps, n_features, n_outputs = train_x.shape[1], train_x.shape[
2], train_y.shape[1]
# create a channel for each variable
in_layers, out_layers = list(), list()
for i in range(n_features):
inputs = Input(shape=(n_timesteps, 1))
conv1 = Conv1D(filters=filters,
kernel_size=kernel_size,
activation='relu')(inputs)
conv2 = Conv1D(filters=filters,
kernel_size=kernel_size,
activation='relu')(conv1)
conv3 = Conv1D(filters=filters,
kernel_size=kernel_size,
activation='relu')(conv2)
conv4 = Conv1D(filters=filters,
kernel_size=kernel_size,
activation='relu')(conv3)
pool1 = MaxPooling1D(pool_size=pool_size)(conv4)
flat = Flatten()(pool1)
# store layers
in_layers.append(inputs)
out_layers.append(flat)
# merge heads
merged = concatenate(out_layers)
# interpretation
dense1 = Dense(dense_1, activation='relu')(merged)
dense2 = Dense(dense_2, activation='relu')(dense1)
outputs = Dense(n_outputs)(dense2)
model = Model(inputs=in_layers, outputs=outputs)
# compile model
model.compile(loss='mse', optimizer='adam')
# fit network
input_data = [
train_x[:, :, i].reshape((train_x.shape[0], n_timesteps, 1))
for i in range(n_features)
]
model.fit(input_data,
train_y,
epochs=epochs,
batch_size=batch_size,
verbose=verbose)
# model.summary() # added
return model
def build_model(train, n_input):
# prepare data
train_x, train_y = to_supervised(train, n_input, 365)
# define parameters
epochs, batch_size, verbose = 20, 8, 0
n_timesteps, n_features, n_outputs = train_x.shape[1], train_x.shape[
2], train_y.shape[1]
# define model
model = Sequential()
model.add(
Conv1D(filters=32,
kernel_size=3,
activation='relu',
input_shape=(n_timesteps, n_features)))
model.add(Conv1D(filters=32, kernel_size=3, activation='relu'))
model.add(MaxPooling1D(pool_size=2))
model.add(Conv1D(filters=16, kernel_size=3, activation='relu'))
model.add(Conv1D(filters=16, kernel_size=3, activation='relu'))
model.add(MaxPooling1D(pool_size=2))
model.add(Flatten())
model.add(Dense(100, activation='relu'))
model.add(Dense(n_outputs))
# compile model
model.compile(loss='mse', optimizer='adam')
# fit network
model.fit(train_x,
train_y,
epochs=epochs,
batch_size=batch_size,
verbose=verbose)
# plot the model
# plot_model(model, show_shapes=True, to_file='plots/model_plot_multichannel.png')
return model
def build_model(train, n_input):
# prepare data
train_x, train_y = to_supervised(train, n_input, 365)
# define parameters
epochs, batch_size, verbose = 20, 16, 0
filters, kernel_size, pool_size = 64, 8, 4
dense_1, dense_2 = 500, 200
n_timesteps, n_features, n_outputs = train_x.shape[1], train_x.shape[
2], train_y.shape[1]
# create a channel for each variable
in_layers, out_layers = list(), list()
for i in range(n_features):
inputs = Input(shape=(n_timesteps, 1))
conv1 = Conv1D(filters=filters,
kernel_size=kernel_size,
activation='relu')(inputs)
conv2 = Conv1D(filters=filters,
kernel_size=kernel_size,
activation='relu')(conv1)
conv3 = Conv1D(filters=filters,
kernel_size=kernel_size,
activation='relu')(conv2)
conv4 = Conv1D(filters=filters,
kernel_size=kernel_size,
activation='relu')(conv3)
pool1 = MaxPooling1D(pool_size=pool_size)(conv4)
flat = Flatten()(pool1)
# store layers
in_layers.append(inputs)
out_layers.append(flat)
# merge heads
merged = concatenate(out_layers)
# interpretation by fully-connected layers
dense1 = Dense(dense_1, activation='relu')(merged)
dense2 = Dense(dense_2, activation='relu')(dense1)
outputs = Dense(n_outputs)(dense2)
model = Model(inputs=in_layers, outputs=outputs)
# compile model
model.compile(loss='mse', optimizer='adam')
# fit network
input_data = [
train_x[:, :, i].reshape((train_x.shape[0], n_timesteps, 1))
for i in range(n_features)
]
model.fit(input_data,
train_y,
epochs=epochs,
batch_size=batch_size,
verbose=verbose)
# model.summary()
# model.save('temp/CNN_multihead_model_1.h5') # approx. 55 MB per net
# plot the model
# plot_model(model, show_shapes=True, to_file='plots/model_plot_multihead.png')
return model