def test_lr(self):
""" Logistic regression.
"""
model = Sequential()
model.add(Dense(1, use_bias=False, input_dim=4))
model.compile(loss='mse')
def fn(vec):
a, b, c, d = vec
return 4 * a + 2 * b + 3 * c + d
inputs = np.random.rand(10, 4)
outputs = np.zeros(inputs.shape[0])
for i in range(inputs.shape[0]):
outputs[i] = fn(inputs[i, :])
model.train_function = scipy_optimizer.make_train_function(model,
verbose=0,
maxiter=30)
hist = model.fit(inputs, outputs, epochs=1, verbose=False)
self.assertTrue('loss' in hist.history)
layers = [layer for layer in model.layers if layer.weights]
w = layers[0].get_weights()[0].reshape(-1)
numpy.testing.assert_almost_equal(w, [4.0, 2.0, 3.0, 1.0], decimal=4)
def test_mult_inputs(self):
""" Multiple input model
"""
def test_fn(x, y):
return 2.0 * x + 4.0 * y + 1.0
def make_model():
x = Input(shape=(1, ))
y = Input(shape=(1, ))
join = Concatenate()([x, y])
z = Dense(1)(join)
return Model([x, y], z)
model = make_model()
model.compile(loss='mse')
X = np.random.rand(10) # pylint: disable=invalid-name
Y = np.random.rand(10) # pylint: disable=invalid-name
Z = np.vectorize(test_fn)(X, Y) # pylint: disable=invalid-name
model.train_function = scipy_optimizer.make_train_function(model,
verbose=0,
maxiter=100)
hist = model.fit([X, Y], Z, verbose=False)
self.assertLess(hist.history['loss'][-1], 0.1)
def test_1dim(self):
""" Input data with rank 1.
"""
def test_fn(x):
if x > 0.5:
return 1
return 0
def make_model():
inp = Input(shape=(1, ))
kinit = keras.initializers.RandomUniform(0.0, 1.0)
h_layer = Dense(1, kernel_initializer=kinit,
activation='relu')(inp)
outp = Dense(1, activation='sigmoid')(h_layer)
return Model(inp, outp)
model = make_model()
model.compile(loss='mse')
model.train_function = scipy_optimizer.make_train_function(model,
verbose=0,
maxiter=15)
X = np.random.rand(100) # pylint: disable=invalid-name
y = np.vectorize(test_fn)(X)
X_train, X_test, y_train, y_test = train_test_split( # pylint: disable=invalid-name
X,
y,
test_size=0.20,
random_state=42)
model.fit(X_train, y_train, verbose=False)
self.assertLessEqual(model.evaluate(X_test, y_test), 1.0e-5)
def test_2layer(self):
""" logistic regression using an hidden layer
"""
model = Sequential()
model.add(Dense(3, use_bias=False, input_dim=4))
model.add(Dense(1, use_bias=False))
model.compile(loss='mse')
def fn(vec):
a, b, c, d = vec
return a * b + 2 * b + 3 * c + d
inputs = np.random.rand(10, 4)
outputs = np.zeros(inputs.shape[0])
for i in range(inputs.shape[0]):
outputs[i] = fn(inputs[i, :])
model.train_function = scipy_optimizer.make_train_function(model,
verbose=0,
maxiter=15)
model.fit(inputs, outputs, verbose=False)
pred = model.predict(inputs)
delta = outputs - pred.reshape(-1)
self.assertEqual(delta.shape, outputs.shape)
self.assertLess(delta.sum(), 0.01)
def test_graph_mode(self):
""" Ensure that the model is executed in graph mode.
"""
def custom_layer(x):
assert not tf.executing_eagerly()
return tf.reduce_sum(x, axis=-1)
model = Sequential()
model.add(InputLayer(input_shape=(4, )))
model.add(Dense(2))
model.add(Lambda(custom_layer))
model.compile(loss="mse")
def objective_fn(x):
a = np.dot(x, np.array([1, 2, 3, 4])[:, np.newaxis])
b = np.dot(x, np.array([5, 6, 7, 8])[:, np.newaxis])
return a + b
x_data = np.random.uniform(size=40).reshape(10, 4)
y = np.apply_along_axis(objective_fn, -1, x_data)
model.train_function = scipy_optimizer.make_train_function(model,
verbose=0,
maxiter=25)
hist = model.fit(x_data, y, epochs=1, verbose=False)
self.assertLess(hist.history['loss'][-1], 1.0e-3)
def test_bfgs(self):
""" Embedding generation using method bfgs.
"""
matrix = make_test_matrix((10, 10), 50)
generator = MatrixDataGenerator(matrix.tocoo())
model = make_embedding_model(matrix.shape, 3)
model.train_function = scipy_optimizer.make_train_function(
model, method='bfgs', verbose=0, maxiter=200)
hist = model.fit(generator, verbose=False)
self.assertLess(hist.history['loss'][-1], 1.0e-3)
def test_val_data(self):
""" Validation metrics
"""
def test_fn(x):
if x > 0.8:
return 2
return 0
def make_model():
inp = Input(shape=(1, ))
h_layer = Dense(
1,
kernel_initializer=keras.initializers.RandomUniform(0.0, 1.0),
activation='relu')(inp)
outp = Dense(1, activation='sigmoid')(h_layer)
return Model(inp, outp)
model = make_model()
model.compile(loss='mse', metrics=['mae'])
X = np.random.rand(200) # pylint: disable=invalid-name
y = np.vectorize(test_fn)(X)
X_train, X_test, y_train, y_test = train_test_split( # pylint: disable=invalid-name
X,
y,
test_size=0.20,
random_state=42)
model.train_function = scipy_optimizer.make_train_function(model,
verbose=0,
maxiter=50)
hist = model.fit(X_train,
y_train,
validation_data=(X_test, y_test),
verbose=False)
self.assertLessEqual(hist.history['loss'][-1], 0.2)
self.assertTrue('val_loss' in hist.history)
self.assertTrue('val_mean_absolute_error' in hist.history
or 'val_mae' in hist.history)
def test_non_trainable(self):
"""BatchNormalization uses non-trainable weights.
"""
model = Sequential()
model.add(Dense(3, use_bias=False, input_dim=4))
model.add(BatchNormalization())
model.add(Dense(1, use_bias=False))
model.compile(loss='mse')
def fn(vec):
a, b, c, d = vec
return a * b + 2 * b + 3 * c + d
inputs = np.random.rand(10, 4)
outputs = np.zeros(inputs.shape[0])
for i in range(inputs.shape[0]):
outputs[i] = fn(inputs[i, :])
model.train_function = scipy_optimizer.make_train_function(model,
verbose=0,
maxiter=50)
hist = model.fit(inputs, outputs, verbose=False)
self.assertLessEqual(hist.history['loss'][-1], 1.0e3)