def test_defaults_sequence2():
seq = DefaultsSequence(input_dim=(3, 4, 4),
lists=[
Convolutional(num_filters=10,
stride=(2, 2),
filter_size=(3, 3)),
BatchNormalization(),
Rectifier(),
Flattener(),
Linear(output_dim=10),
BatchNormalization(),
Rectifier(),
Linear(output_dim=12),
BatchNormalization(),
Rectifier()
])
seq.weights_init = Constant(1.0)
seq.biases_init = Constant(0.0)
seq.push_allocation_config()
seq.push_initialization_config()
seq.initialize()
x = T.tensor4('input')
y = seq.apply(x)
func_ = theano.function([x], [y])
x_val = np.ones((1, 3, 4, 4), dtype=theano.config.floatX)
res = func_(x_val)[0]
assert_allclose(res.shape, (1, 12))
def test_batchnorm_training():
layer = BatchNormalization(input_dim=5)
layer.initialize()
x = T.matrix()
x_val = np.ones((6, 5), dtype=theano.config.floatX)
x_val[0, 0] = 10.0
y = layer.apply(x)
_func = theano.function([x], y)
ret = _func(x_val)
assert_allclose(ret[0, 0], 2.23606801)
assert_allclose(ret[1:5, 0], -0.44721359)
assert_allclose(ret[0:5, 1:5], 0)
def test_batchnorm_training():
layer = BatchNormalization(
input_dim = 5)
layer.initialize()
x = T.matrix()
x_val = np.ones((6, 5), dtype=theano.config.floatX)
x_val[0,0] = 10.0
y = layer.apply(x)
_func = theano.function([x], y)
ret = _func(x_val)
assert_allclose(ret[0,0], 2.23606801)
assert_allclose(ret[1:5, 0], -0.44721359)
assert_allclose(ret[0:5,1:5], 0)
def test_batchnorm_infer():
layer = BatchNormalization(input_dim = (4, 5, 5))
layer.accumulate = True
layer.initialize()
x = T.tensor4("features")
x_val = [np.ones((6, 4, 5, 5), dtype=theano.config.floatX) for _ in range(2)]
x_val[0][0,0,0,0] = 10.0
x_val[1][0,0,0,0] = -200.0
y = layer.apply(x)
dataset = IterableDataset(dict(features=x_val))
data_stream = DataStream(dataset)
cg = ComputationGraph([y])
infer_population(data_stream, cg, 2)
assert layer.use_population == True
assert_allclose(layer.u.get_value(), np.array([0.72, 2, 2, 2]))
assert_allclose(layer.n.get_value(), np.array([2]))
def test_batchnormconv_training():
layer = BatchNormalization(input_dim = (4, 5, 5))
layer.initialize()
x = T.tensor4()
x_val = np.ones((6, 4, 5, 5), dtype=theano.config.floatX)
x_val[0,0,0,0] = 10.0
y = layer.apply(x)
_func = theano.function([x], y)
ret = _func(x_val)
assert_equal(ret.shape, (6, 4, 5, 5))
assert_allclose(ret[0, 0, 0, 0], 12.20655537)
assert_allclose(ret[0, 0, 1, 0], -0.08192328)
assert_allclose(ret[0, 0, 3, 3], -0.08192328)
assert_allclose(ret[1, 0, 0, 0], -0.08192328)
assert_allclose(ret[0:6, 1:4, 0:5, 0:5], 0)
def test_batchnormconv_training():
layer = BatchNormalization(input_dim=(4, 5, 5))
layer.initialize()
x = T.tensor4()
x_val = np.ones((6, 4, 5, 5), dtype=theano.config.floatX)
x_val[0, 0, 0, 0] = 10.0
y = layer.apply(x)
_func = theano.function([x], y)
ret = _func(x_val)
assert_equal(ret.shape, (6, 4, 5, 5))
assert_allclose(ret[0, 0, 0, 0], 12.20655537)
assert_allclose(ret[0, 0, 1, 0], -0.08192328)
assert_allclose(ret[0, 0, 3, 3], -0.08192328)
assert_allclose(ret[1, 0, 0, 0], -0.08192328)
assert_allclose(ret[0:6, 1:4, 0:5, 0:5], 0)
def test_batchnorm_get_set():
layer = BatchNormalization(input_dim = (4, 5, 5))
layer.accumulate = True
layer.initialize()
x = T.tensor4("features")
x_val = [np.ones((6, 4, 5, 5), dtype=theano.config.floatX) for _ in range(2)]
x_val[0][0,0,0,0] = 10.0
x_val[1][0,0,0,0] = -200.0
y = layer.apply(x)
dataset = IterableDataset(dict(features=x_val))
data_stream = DataStream(dataset)
cg = Model([y])
infer_population(data_stream, cg, 2)
values_dict = get_batchnorm_parameter_values(cg)
assert len(values_dict.keys()) == 3
assert_allclose(values_dict['/batchnormalization.n'], np.array([2]))
values_dict['/batchnormalization.n'] = np.array([5.], dtype="float32")
set_batchnorm_parameter_values(cg, values_dict)
assert_allclose(layer.n.get_value(), np.array([5.]))
def test_batchnorm_rolling():
layer = BatchNormalization(input_dim=5, rolling_accumulate=True)
layer.initialize()
x = T.matrix()
x_val = np.ones((6, 5), dtype=theano.config.floatX)
x_val[0, 0] = 10.0
y = layer.apply(x)
cg = ComputationGraph([y])
_func = cg.get_theano_function()
for i in range(100):
ret = _func(x_val)
u = layer.u.get_value()
assert_allclose(u[0], 1.58491838)
assert_allclose(u[1], 0.6339674)
s = layer.s.get_value()
assert_allclose(s[0], 7.13214684)
assert_allclose(s[1], 0.)
def test_batchnorm_rolling():
layer = BatchNormalization(
input_dim = 5, rolling_accumulate=True)
layer.initialize()
x = T.matrix()
x_val = np.ones((6, 5), dtype=theano.config.floatX)
x_val[0,0] = 10.0
y = layer.apply(x)
cg = ComputationGraph([y])
_func = cg.get_theano_function()
for i in range(100):
ret = _func(x_val)
u = layer.u.get_value()
assert_allclose(u[0], 1.58491838)
assert_allclose(u[1], 0.6339674)
s = layer.s.get_value()
assert_allclose(s[0], 7.13214684)
assert_allclose(s[1], 0.)
def test_defaults_sequence1():
seq = DefaultsSequence(input_dim=9,
lists=[
Linear(output_dim=10),
BatchNormalization(),
Rectifier(),
Linear(output_dim=12),
BatchNormalization(),
Rectifier()
])
seq.weights_init = Constant(1.0)
seq.biases_init = Constant(0.0)
seq.push_allocation_config()
seq.push_initialization_config()
seq.initialize()
x = T.matrix('input')
y = seq.apply(x)
func_ = theano.function([x], [y])
x_val = np.ones((1, 9), dtype=theano.config.floatX)
res = func_(x_val)[0]
assert_allclose(res.shape, (1, 12))
def test_batchnorm_infer():
layer = BatchNormalization(input_dim=(4, 5, 5))
layer.accumulate = True
layer.initialize()
x = T.tensor4("features")
x_val = [
np.ones((6, 4, 5, 5), dtype=theano.config.floatX) for _ in range(2)
]
x_val[0][0, 0, 0, 0] = 10.0
x_val[1][0, 0, 0, 0] = -200.0
y = layer.apply(x)
dataset = IterableDataset(dict(features=x_val))
data_stream = DataStream(dataset)
cg = ComputationGraph([y])
infer_population(data_stream, cg, 2)
assert layer.use_population == True
assert_allclose(layer.u.get_value(), np.array([0.72, 2, 2, 2]))
assert_allclose(layer.n.get_value(), np.array([2]))
def test_batchnorm_get_set():
layer = BatchNormalization(input_dim=(4, 5, 5))
layer.accumulate = True
layer.initialize()
x = T.tensor4("features")
x_val = [
np.ones((6, 4, 5, 5), dtype=theano.config.floatX) for _ in range(2)
]
x_val[0][0, 0, 0, 0] = 10.0
x_val[1][0, 0, 0, 0] = -200.0
y = layer.apply(x)
dataset = IterableDataset(dict(features=x_val))
data_stream = DataStream(dataset)
cg = Model([y])
infer_population(data_stream, cg, 2)
values_dict = get_batchnorm_parameter_values(cg)
assert len(values_dict.keys()) == 3
assert_allclose(values_dict['/batchnormalization.n'], np.array([2]))
values_dict['/batchnormalization.n'] = np.array([5.], dtype="float32")
set_batchnorm_parameter_values(cg, values_dict)
assert_allclose(layer.n.get_value(), np.array([5.]))
