def test_DataGenerator_workers():
"""Tests probflow.data.DataGenerator w/ multiple worker processes"""
# Data
x = np.random.randn(100, 3).astype('float32')
w = np.random.randn(3, 1).astype('float32')
y = x @ w
# Fit a model with 1 worker
model = pf.LinearRegression(3)
model.fit(x, y, batch_size=10, epochs=10, num_workers=1)
# Fit a model with 4 workers
model = pf.LinearRegression(3)
model.fit(x, y, batch_size=10, epochs=10, num_workers=4)
def test_example_linear_regression_multiple():
"""Tests example_linear_regression multiple linear regression"""
# TODO: generate data
# Create and fit the model
model = pf.LinearRegression()
model.fit(x, y)
def test_dumps_and_loads_before_fitting():
model1 = pf.LinearRegression(7)
model2 = pf.loads(model1.dumps())
assert isinstance(model2, pf.LinearRegression)
assert model1 is not model2
post1 = model1.posterior_mean()
post2 = model2.posterior_mean()
for k in post1:
assert isclose(post1[k], post2[k])
def test_dumps_and_loads_after_fitting():
model1 = pf.LinearRegression(7)
x, y = get_test_data(1024, 7)
model1.fit(x, y, epochs=2)
model2 = pf.loads(model1.dumps())
assert isinstance(model2, pf.LinearRegression)
assert model1 is not model2
post1 = model1.posterior_mean()
post2 = model2.posterior_mean()
for k in post1:
assert isclose(post1[k], post2[k])
def test_dump_and_load_before_fitting(tmpdir):
model1 = pf.LinearRegression(7)
fname = str(tmpdir.join("test_model.pkl"))
model1.save(fname)
model2 = pf.load(fname)
assert isinstance(model2, pf.LinearRegression)
assert model1 is not model2
post1 = model1.posterior_mean()
post2 = model2.posterior_mean()
for k in post1:
assert isclose(post1[k], post2[k])
def test_dump_and_load_after_fitting(tmpdir):
model1 = pf.LinearRegression(7)
x, y = get_test_data(1024, 7)
model1.fit(x, y, epochs=2)
fname = str(tmpdir.join("test_model.pkl"))
model1.save(fname)
model2 = pf.load(fname)
assert isinstance(model2, pf.LinearRegression)
assert model1 is not model2
post1 = model1.posterior_mean()
post2 = model2.posterior_mean()
for k in post1:
assert isclose(post1[k], post2[k])
def test_linear_regression_times():
times = []
for N, D, backend, eager in product(ns, ds, backends, eagers):
pf.set_backend(backend)
model = pf.LinearRegression(D)
x, y = get_data(N, D)
model.fit(x, y, epochs=2,
eager=eager) # don't include compilation time
t0 = time.time()
model.fit(x, y, epochs=EPOCHS, eager=eager)
t1 = time.time()
times.append({
"N": N,
"D": D,
"backend": backend,
"eager": eager,
"seconds": t1 - t0,
})
df = pd.DataFrame.from_records(times)
print(df)
def test_linear_regression():
"""Test that a linear regression recovers the true parameters"""
# Set random seed
np.random.seed(1234)
tf.random.set_seed(1234)
# Generate data
N = 1000
D = 5
x = np.random.randn(N, D).astype('float32')
w = np.random.randn(D, 1)
b = np.random.randn()
std = np.exp(np.random.randn())
noise = std * np.random.randn(N, 1)
y = x @ w + b + noise
y = y.astype('float32')
# Create and fit model
model = pf.LinearRegression(D)
model.fit(x, y, batch_size=100, epochs=1000, learning_rate=1e-2)
# Compute and check confidence intervals on the weights
lb, ub = model.posterior_ci('weights')
assert np.all(lb < w)
assert np.all(ub > w)
# Compute and check confidence intervals on the bias
lb, ub = model.posterior_ci('bias')
assert lb < b
assert ub > b
# Compute and check confidence intervals on the std
lb, ub = model.posterior_ci('std')
assert lb < std
assert ub > std