def test_log_likelihood():
"""
Computes the log-likelihood "by hand" for a simple example and ensures
that the one returned by xlogit is the same
"""
P = 1 # Without panel data
betas = np.array([.1, .1, .1, .1])
X_, y_ = X.reshape(N, P, J, K), y.reshape(N, P, J, 1)
# Compute log likelihood using xlogit
model = MixedLogit()
model._rvidx, model._rvdist = np.array([True, True]), np.array(['n', 'n'])
draws = model._get_halton_draws(N, R, K) # (N,Kr,R)
panel_info = np.ones((N, P))
obtained_loglik, _ = model._loglik_gradient(betas, X_, y_, panel_info,
draws, None, None)
# Compute expected log likelihood "by hand"
Br = betas[None, [0, 1], None] + draws * betas[None, [2, 3], None]
eXB = np.exp(np.einsum('npjk,nkr -> npjr', X_, Br))
p = eXB / np.sum(eXB, axis=2, keepdims=True)
expected_loglik = -np.sum(
np.log((y_ * p).sum(axis=2).prod(axis=1).mean(axis=1)))
assert pytest.approx(expected_loglik, obtained_loglik)
def test_log_likelihood():
"""
Computes the log-likelihood "by hand" for a simple example and ensures
that the one returned by xlogit is the same
"""
betas = np.array([.1, .1, .1, .1])
X_, y_ = X.reshape(N, J, K), y.reshape(N, J, 1)
# Compute log likelihood using xlogit
model = MixedLogit()
model._rvidx, model._rvdist = np.array([True, True]), np.array(['n', 'n'])
draws = model._generate_halton_draws(N, R, K) # (N,Kr,R)
obtained_loglik = model._loglik_gradient(betas,
X_,
y_,
None,
draws,
None,
None, {
'samples': N,
'draws': R
},
return_gradient=False)
# Compute expected log likelihood "by hand"
Br = betas[None, [0, 1], None] + draws * betas[None, [2, 3], None]
eXB = np.exp(np.einsum('njk,nkr -> njr', X_, Br))
p = eXB / np.sum(eXB, axis=1, keepdims=True)
expected_loglik = -np.sum(np.log((y_ * p).sum(axis=1).mean(axis=1)))
assert expected_loglik == pytest.approx(obtained_loglik)
def test__balance_panels():
"""
Ensures that unbalanced panels are properly balanced when required
"""
X_, y_ = X.reshape(N, J, K), y.reshape(N, J, 1)
model = MixedLogit()
X_, y_, panel_info = model._balance_panels(X_, y_, panels)
assert np.array_equal(panel_info, np.array([[1, 1], [1, 0]]))
assert X_.shape == (4, 2, 2)
def test__transform_betas():
"""
Check that betas are properly transformed to random draws
"""
betas = np.array([.1, .1, .1, .1])
# Compute log likelihood using xlogit
model = MixedLogit()
model._rvidx, model._rvdist = np.array([True, True]), np.array(['n', 'n'])
draws = model._get_halton_draws(N, R, K) # (N,Kr,R)
expected_betas = betas[None, [0, 1], None] + \
draws*betas[None, [2, 3], None]
_, obtained_betas = model._transform_betas(betas, draws)
assert np.allclose(expected_betas, obtained_betas)
def test_gpu_not_available():
"""
Ensures that xlogit detects that GPU is not available based on CuPy's
installation status
"""
assert not MixedLogit.check_if_gpu_available()
def test_lrtest():
"""
Ensures a correct result of the lrtest. The comparison values were
obtained from comparison with lrtest in R's lmtest package
"""
general = MixedLogit()
general.loglikelihood = 1312
restricted = MixedLogit()
restricted.loglikelihood = -1305
general.loglikelihood = -1312
general.coeff_ = np.zeros(4)
restricted.coeff_ = np.zeros(2)
obtained = lrtest(general, restricted)
expected = {'pval': 0.0009118819655545164, 'chisq': 14, 'degfree': 2}
def test_validate_inputs():
"""
Covers potential mistakes in parameters of the fit method that xlogit
should be able to identify
"""
model = MixedLogit()
with pytest.raises(ValueError): # wrong distribution
model.fit(X,
y,
varnames=varnames,
alts=alts,
ids=ids,
n_draws=10,
maxiter=0,
verbose=0,
halton=True,
randvars={'a': 'fake'})
with pytest.raises(ValueError): # wrong var name
model.fit(X,
y,
varnames=varnames,
alts=alts,
ids=ids,
n_draws=10,
maxiter=0,
verbose=0,
halton=True,
randvars={'fake': 'n'})
def test_fit():
"""
Ensures the log-likelihood works for a single iterations with the default
initial coefficients. The value of -1.794 was computed by hand for
comparison purposes
"""
# There is no need to initialize a random seed as the halton draws produce
# reproducible results
model = MixedLogit()
model.fit(X,
y,
varnames=varnames,
alts=alts,
n_draws=10,
panels=panels,
ids=ids,
randvars=randvars,
maxiter=0,
verbose=0,
halton=True)
assert pytest.approx(model.loglikelihood, -1.79451632)
def test_fit():
"""
Ensures the log-likelihood works for multiple iterations with the default
initial coefficients. The value of -1.473423 was computed by hand for
comparison purposes
"""
# There is no need to initialize a random seed as the halton draws produce
# reproducible results
model = MixedLogit()
model.fit(X,
y,
varnames,
alts,
ids,
randvars,
n_draws=10,
panels=panels,
maxiter=0,
verbose=0,
halton=True,
init_coeff=np.repeat(.1, 4))
assert model.loglikelihood == pytest.approx(-1.473423)
def test_predict():
"""
Ensures that returned choice probabilities are consistent.
"""
# There is no need to initialize a random seed as the halton draws produce
# reproducible results
P = 1 # Without panel data
betas = np.array([.1, .1, .1, .1])
X_ = X.reshape(N, P, J, K)
model = MixedLogit()
model._rvidx, model._rvdist = np.array([True, True]), np.array(['n', 'n'])
model.alternatives = np.array([1, 2])
model.coeff_ = betas
model.randvars = randvars
model._isvars, model._asvars, model._varnames = [], varnames, varnames
model._fit_intercept = False
model.coeff_names = np.array(["a", "b", "sd.a", "sd.b"])
#model.fit(X, y, varnames, alts, ids, randvars, verbose=0, halton=True)
y_pred, proba, freq = model.predict(X,
varnames,
alts,
ids,
n_draws=R,
return_proba=True,
return_freq=True)
# Compute choice probabilities by hand
draws = model._get_halton_draws(N, R, K) # (N,Kr,R)
Br = betas[None, [0, 1], None] + draws * betas[None, [2, 3], None]
V = np.einsum('npjk,nkr -> npjr', X_, Br)
V[V > 700] = 700
eV = np.exp(V)
e_proba = eV / np.sum(eV, axis=2, keepdims=True)
expec_proba = e_proba.prod(axis=1).mean(axis=-1)
expec_ypred = model.alternatives[np.argmax(expec_proba, axis=1)]
alt_list, counts = np.unique(expec_ypred, return_counts=True)
expec_freq = dict(
zip(list(alt_list), list(np.round(counts / np.sum(counts), 3))))
assert np.array_equal(expec_ypred, y_pred)
assert expec_freq == freq
X = df[varnames].values
y = df['choice'].values
randvars = {'meals': 'n', 'petfr': 'n', 'emipp': 'n'}
alts = df['alt']
ids = df['id']
panels = None
batch_size = 5000
if not use_gpu:
device.disable_gpu_acceleration()
if profile:
ini_ram = curr_ram()
profiler = Profiler().start(measure_gpu_mem=use_gpu)
np.random.seed(0)
model = MixedLogit()
model.fit(X,
y,
varnames,
alts=alts,
ids=ids,
n_draws=n_draws,
panels=panels,
verbose=0,
randvars=randvars,
batch_size=batch_size)
if profile:
ellapsed, max_ram, max_gpu = profiler.stop()
log("{:6} {:7.2f} {:11.2f} {:7.3f} {:7.3f} {}".format(
n_draws, ellapsed, model.loglikelihood, max_ram - ini_ram, max_gpu,
"""
This file executes the benchmark. Check the README.md file in this folder
to make sure all the requirments are satisfied.
"""
import os
from tools import init_profiler_output_files, log
import sys
# sys.path.append("../../") # Path of xlogit library root folder.
from xlogit import MixedLogit
MixedLogit.check_if_gpu_available()
mini = len(sys.argv) == 2 and sys.argv[1] == 'mini'
init_profiler_output_files()
def profile_range_draws(command, r_draws, dataset, usegpu=False):
log("\n\n=== " + dataset + " dataset. " + command.split()[1] +
('(using GPU)' if usegpu else '') + " ===")
log("Ndraws Time(s) Log-Likeli. RAM(GB) GPU(GB) Converg.")
for r in range(1, r_draws + 1):
os.system("{} {} {} {} prof".format(command, r * 100, dataset,
usegpu * 1))
def profile_range_draws_and_cores(command, r_draws, r_cores):
log("\n\n=== artificial dataset. " + command.split()[1] + " ===")
for n_draws in r_draws:
for n_cores in r_cores:
os.system("{} {} {}".format(command, n_draws, n_cores))
alt_list=['TRAIN', 'SM', 'CAR'],
empty_val=0,
varying=['TT', 'CO', 'AV'],
alt_is_prefix=True)
# ===== STEP 3. CREATE MODEL SPECIFICATION =====
df['ASC_TRAIN'] = np.ones(len(df)) * (df['alt'] == 'TRAIN')
df['ASC_CAR'] = np.ones(len(df)) * (df['alt'] == 'CAR')
df['TT'], df['CO'] = df['TT'] / 100, df['CO'] / 100 # Scale variables
annual_pass = (df['GA'] == 1) & (df['alt'].isin(['TRAIN', 'SM']))
df.loc[annual_pass, 'CO'] = 0 # Cost zero for pass holders
# ===== STEP 4. ESTIMATE MODEL PARAMETERS =====
from xlogit import MixedLogit
varnames = ['ASC_CAR', 'ASC_TRAIN', 'CO', 'TT']
model = MixedLogit()
model.fit(X=df[varnames],
y=df['CHOICE'],
varnames=varnames,
alts=df['alt'],
ids=df['custom_id'],
panels=df["ID"],
avail=df['AV'],
randvars={'TT': 'n'},
n_draws=1500)
model.summary()
"""
OUTPUT:
Estimation time= 1.3 seconds
---------------------------------------------------------------------------
Coefficient Estimate Std.Err. z-val P>|z|
'nonsig1', 'nonsig2', 'nonsig3'
]
X = df[varnames].values
y = df['choice'].values
randvars = {'meals': 'n', 'petfr': 'n', 'emipp': 'n'}
alts = [1, 2, 3]
panels = None
if not use_gpu:
device.disable_gpu_acceleration()
if profile:
ini_ram = curr_ram()
profiler = Profiler().start(measure_gpu_mem=use_gpu)
np.random.seed(0)
model = MixedLogit()
model.fit(X,
y,
varnames,
alts=alts,
n_draws=n_draws,
panels=panels,
verbose=0,
randvars=randvars)
if profile:
ellapsed, max_ram, max_gpu = profiler.stop()
log("{:6} {:7.2f} {:11.2f} {:7.3f} {:7.3f} {}".format(
n_draws, ellapsed, model.loglikelihood, max_ram - ini_ram, max_gpu,
model.convergence))
profiler.export('xlogit' + ('_gpu' if use_gpu else ''), dataset, n_draws,