def test_stacks(stack_cls, block_types, expected_rmse):
fdw = 28
fw = 7
x_train, y_train, x_test, y_test = simple_seq_data(nrows=1000,
freq="1H",
fdw=fdw,
fw=fw,
test_size=0.2)
inputs = keras.Input(shape=(fdw, 1))
outputs = keras.layers.Reshape((fdw, ))(inputs)
stack = stack_cls(
fdw=fdw,
fw=fw,
block_types=block_types,
block_units=8,
block_theta_units=8,
block_kernel_initializer=get_initializer("glorot_uniform",
RANDOM_SEED),
)
if stack_cls == ResidualInputStack:
_, forecast = stack([outputs, outputs])
else:
_, forecast = stack(outputs)
model = keras.Model(inputs=inputs, outputs=forecast)
model.compile(optimizer=keras.optimizers.Adam(0.01),
loss=keras.losses.MeanSquaredError())
model.fit(x_train, y_train, epochs=5, batch_size=32, shuffle=False)
y_pred = model.predict(x_test)
assert np.all(np.isfinite(y_pred))
error = rmse(y_test, y_pred)
assert error < expected_rmse
def test_positional_encoding():
fdw = 28
fw = 7
x_train, y_train, x_test, y_test = simple_seq_data(nrows=1000,
freq="1H",
fdw=fdw,
fw=fw,
test_size=0.2)
inputs = keras.Input(shape=(fdw, 1))
outputs = PositionalEncoding(8)(inputs)
outputs = keras.layers.Concatenate()([inputs, outputs])
outputs = keras.layers.Flatten()(outputs)
outputs = keras.layers.Dense(
fw,
kernel_initializer=get_initializer("glorot_uniform", RANDOM_SEED),
activation="linear",
)(outputs)
model = keras.Model(inputs=inputs, outputs=outputs)
model.compile(optimizer=keras.optimizers.Adam(0.01),
loss=keras.losses.MeanSquaredError())
model.fit(x_train, y_train, epochs=5, batch_size=32, shuffle=False)
y_pred = model.predict(x_test)
assert np.all(np.isfinite(y_pred))
error = rmse(y_test, y_pred)
assert error < 0.5
def nbeast_residual_test_scenarious():
args = {
"block_units": 8,
"block_theta_units": 8,
"block_layers": 4,
"block_kernel_initializer": get_initializer("glorot_uniform",
RANDOM_SEED),
"block_bias_initializer": "zeros",
}
tests = [
[
StackDef(StackTypes.RESIDUAL_INPUT,
block_types=[BlockTypes.GENERIC, BlockTypes.TREND],
**args),
StackDef(StackTypes.RESIDUAL_INPUT,
block_types=[BlockTypes.SEASONAL, BlockTypes.TREND],
**args),
],
[
StackDef(StackTypes.RESIDUAL_INPUT,
block_types=[BlockTypes.GENERIC, BlockTypes.GENERIC],
**args),
],
]
return tests
def test_nbeats_residual_parameters_validation():
args = {
"block_units": 8,
"block_theta_units": 8,
"block_layers": 4,
"block_kernel_initializer": get_initializer("glorot_uniform",
RANDOM_SEED),
"block_bias_initializer": "zeros",
}
with pytest.raises(ValueError) as excinfo:
NBEATSResidual(
fdw=28,
fw=12,
stacks=[
StackDef(StackTypes.RESIDUAL_INPUT,
block_types=[BlockTypes.GENERIC, BlockTypes.TREND],
**args),
StackDef(StackTypes.LAST_FORWARD,
block_types=[BlockTypes.GENERIC, BlockTypes.TREND],
**args),
],
)
assert str(excinfo.value) == (
"RESIDUAL-INPUT model supports RESIDUAL-INPUT stacks only. "
"Found: {<StackTypes.LAST_FORWARD: 3>}")
def test_base_block():
fdw = 28
fw = 7
x_train, y_train, x_test, y_test = simple_seq_data(nrows=1000,
freq="1H",
fdw=fdw,
fw=fw,
test_size=0.2)
inputs = keras.Input(shape=(fdw, 1))
outputs = keras.layers.Reshape((fdw, ))(inputs)
_, outputs = Block(
units=8,
theta_units=8,
kernel_initializer=get_initializer("glorot_uniform", RANDOM_SEED),
)(inputs)
outputs = keras.layers.Flatten()(outputs)
outputs = keras.layers.Dense(fw)(outputs)
model = keras.Model(inputs=inputs, outputs=outputs)
model.compile(optimizer=keras.optimizers.Adam(0.01),
loss=keras.losses.MeanSquaredError())
model.fit(x_train, y_train, epochs=5, batch_size=32, shuffle=False)
y_pred = model.predict(x_test)
assert np.all(np.isfinite(y_pred))
error = rmse(y_test, y_pred)
assert error < 0.5
def test_multi_head_attention_lookahead_mask():
fdw = 28
fw = 7
attention_dim = 32
num_heads = 4
x_train, y_train, x_test, y_test = simple_seq_data(nrows=1000,
freq="1H",
fdw=fdw,
fw=fw,
test_size=0.2)
triu = np.triu(np.ones((fdw, fdw)))
mask = np.stack([triu for _ in range(x_train.shape[0])])
mask = np.expand_dims(mask, axis=1)
inputs = keras.Input(shape=(fdw, 1))
lookahead_mask = keras.Input(shape=(1, fdw, fdw))
outputs, attention_weights = MultiHeadAttention(
attention_dim=attention_dim,
num_heads=num_heads,
kernel_initializer=get_initializer("glorot_uniform", RANDOM_SEED),
)([inputs, inputs, inputs], mask=lookahead_mask)
outputs = keras.layers.Reshape(
(fdw * attention_dim * num_heads, ))(outputs)
outputs = keras.layers.Dense(
fw,
kernel_initializer=get_initializer("glorot_uniform", RANDOM_SEED),
activation="linear",
)(outputs)
model = keras.Model(inputs=[inputs, lookahead_mask], outputs=outputs)
model.compile(
optimizer=keras.optimizers.Adam(0.01),
loss=keras.losses.MeanSquaredError(),
)
model.fit([x_train, mask], y_train, epochs=5, batch_size=32, shuffle=False)
mask = np.stack([triu for _ in range(x_test.shape[0])])
mask = np.expand_dims(mask, axis=1)
y_pred = model.predict([x_test, mask])
assert np.all(np.isfinite(y_pred))
error = rmse(y_test, y_pred)
assert error < 0.5
def test_multi_head_attention_padding_mask():
fdw = 28
fw = 7
attention_dim = 32
num_heads = 4
x_train, y_train, x_test, y_test = simple_seq_data(nrows=1000,
freq="1H",
fdw=fdw,
fw=fw,
test_size=0.2)
random_state = np.random.RandomState(RANDOM_SEED)
mask = (random_state.random(
(x_train.shape[0], 1, 1, x_train.shape[1])) > 0.3).astype(np.int)
inputs = keras.Input(shape=(fdw, 1))
padding_mask = keras.Input(shape=(1, 1, fdw))
outputs, attention_weights = MultiHeadAttention(
attention_dim=attention_dim,
num_heads=num_heads,
kernel_initializer=get_initializer("glorot_uniform", RANDOM_SEED),
)([inputs, inputs, inputs], mask=padding_mask)
outputs = keras.layers.Reshape(
(fdw * attention_dim * num_heads, ))(outputs)
outputs = keras.layers.Dense(
fw,
kernel_initializer=get_initializer("glorot_uniform", RANDOM_SEED),
activation="linear",
)(outputs)
model = keras.Model(inputs=[inputs, padding_mask], outputs=outputs)
model.compile(
optimizer=keras.optimizers.Adam(0.01),
loss=keras.losses.MeanSquaredError(),
)
model.fit([x_train, mask], y_train, epochs=5, batch_size=32, shuffle=False)
mask = (random_state.random(
(x_test.shape[0], 1, 1, x_test.shape[1])) > 0.3).astype(np.int)
y_pred = model.predict([x_test, mask])
assert np.all(np.isfinite(y_pred))
error = rmse(y_test, y_pred)
assert error < 0.5
def test_transformer(num_layers, layer_norm_epsilon, dff):
fdw = 28
fw = 7
x_train, y_train, x_test, y_test = simple_seq_data(nrows=1000,
freq="1H",
fdw=fdw,
fw=fw,
test_size=0.2)
transformer = Transformer(
num_layers=num_layers,
attention_dim=32,
num_heads=4,
hidden_activation="linear",
dff=dff,
hidden_kernel_initializer=get_initializer("glorot_uniform",
RANDOM_SEED),
attention_kernel_initializer=get_initializer("glorot_uniform",
RANDOM_SEED),
pwffn_kernel_initializer=get_initializer("glorot_uniform",
RANDOM_SEED),
output_kernel_initializer=get_initializer("glorot_uniform",
RANDOM_SEED),
layer_norm_epsilon=layer_norm_epsilon,
)
transformer.compile(loss=keras.losses.MeanSquaredError(),
optimizer=keras.optimizers.Adam(0.001))
decoder_inputs = create_decoder_inputs(y_train, go_token=0)
transformer.fit([x_train, decoder_inputs],
y_train,
epochs=5,
batch_size=32,
verbose=1)
dec_inp = create_empty_decoder_inputs(x_test.shape[0], go_token=0)
y_pred, weights = transformer.predict([x_test, dec_inp])
assert rmse(y_test, y_pred) < 1.1
assert len(weights["encoder_attention"]) == num_layers
assert len(weights["decoder_attention"]) == num_layers
assert len(weights["encoder_decoder_attention"]) == num_layers
def test_multi_head_attention():
fdw = 28
fw = 7
attention_dim = 32
num_heads = 4
x_train, y_train, x_test, y_test = simple_seq_data(nrows=1000,
freq="1H",
fdw=fdw,
fw=fw,
test_size=0.2)
inputs = keras.Input(shape=(fdw, 1))
outputs, attention_weights = MultiHeadAttention(
attention_dim=attention_dim,
num_heads=num_heads,
kernel_initializer=get_initializer("glorot_uniform", RANDOM_SEED),
)([inputs, inputs, inputs])
outputs = keras.layers.Reshape(
(fdw * attention_dim * num_heads, ))(outputs)
outputs = keras.layers.Dense(
fw,
kernel_initializer=get_initializer("glorot_uniform", RANDOM_SEED),
activation="linear",
)(outputs)
model = keras.Model(inputs=inputs, outputs=outputs)
model.compile(optimizer=keras.optimizers.Adam(0.01),
loss=keras.losses.MeanSquaredError())
model.fit(x_train, y_train, epochs=5, batch_size=32, shuffle=False)
y_pred = model.predict(x_test)
assert np.all(np.isfinite(y_pred))
error = rmse(y_test, y_pred)
assert error < 0.5
def nbeats_test_scenarious():
args = {
"block_units": 8,
"block_theta_units": 8,
"block_layers": 4,
"block_kernel_initializer": get_initializer("glorot_uniform",
RANDOM_SEED),
"block_bias_initializer": "zeros",
}
tests = [
[
StackDef(StackTypes.NBEATS_DRESS,
block_types=[BlockTypes.GENERIC, BlockTypes.TREND],
**args),
],
[
StackDef(StackTypes.PARALLEL,
block_types=[BlockTypes.TREND, BlockTypes.GENERIC],
**args),
StackDef(StackTypes.PARALLEL,
block_types=[BlockTypes.TREND, BlockTypes.SEASONAL],
**args),
StackDef(StackTypes.PARALLEL,
block_types=[BlockTypes.SEASONAL, BlockTypes.SEASONAL],
**args),
],
[
StackDef(StackTypes.NO_RESIDUAL,
block_types=[BlockTypes.TREND, BlockTypes.SEASONAL],
**args),
StackDef(StackTypes.NO_RESIDUAL,
block_types=[BlockTypes.SEASONAL, BlockTypes.TREND],
**args),
],
]
return tests
