def test_lalonde_defaults():
w1, t1, y1 = load_lalonde()
w2, t2, y2 = load_lalonde(rct_version='dw',
obs_version='psid1',
data_format=NUMPY)
np.testing.assert_array_equal(w1, w2)
np.testing.assert_array_equal(t1, t2)
np.testing.assert_array_equal(y1, y2)
def test_lalonde_original_cps1():
(w, t, y) = load_lalonde(rct_version='lalonde', obs_version='cps1')
n = 15992 + 297
assert w.shape == (n, 7)
assert t.shape == (n, )
assert y.shape == (n, )
ate = y[t == 1].mean() - y[t == 0].mean()
assert round(ate) == -8870
def test_lalonde_original_psid1():
(w, t, y) = load_lalonde(rct_version='lalonde', obs_version='psid1')
n = 2490 + 297
assert w.shape == (n, 7)
assert t.shape == (n, )
assert y.shape == (n, )
ate = y[t == 1].mean() - y[t == 0].mean()
assert round(ate) == -15578
def test_lalonde_cps1():
(w, t, y) = load_lalonde(rct_version='dw', obs_version='cps1')
n = 15992 + 185
assert w.shape == (n, 8)
assert t.shape == (n, )
assert y.shape == (n, )
ate = y[t == 1].mean() - y[t == 0].mean()
assert round(ate) == -8498
def test_lalonde_psid1():
(w, t, y) = load_lalonde(rct_version='dw', obs_version='psid1')
n = 2490 + 185
assert w.shape == (n, 8)
assert t.shape == (n, )
assert y.shape == (n, )
ate = y[t == 1].mean() - y[t == 0].mean()
assert round(ate) == -15205
def test_lalonde_original_rct():
(w, t, y) = load_lalonde(rct_version='lalonde', rct=True)
n = 425 + 297
assert w.shape == (n, 7)
assert t.shape == (n, )
assert y.shape == (n, )
ate = y[t == 1].mean() - y[t == 0].mean()
assert round(ate) == 886
def test_lalonde_dw_rct():
(w, t, y) = load_lalonde(rct_version='dw', rct=True)
n = 260 + 185
assert w.shape == (n, 8)
assert t.shape == (n, )
assert y.shape == (n, )
ate = y[t == 1].mean() - y[t == 0].mean()
assert round(ate) == 1794
def get_data(args):
data_name = args.data.lower()
ate = None
ites = None
if data_name == "lalonde" or data_name == "lalonde_psid" or data_name == "lalonde_psid1":
w, t, y = load_lalonde(obs_version="psid", dataroot=args.dataroot)
elif data_name == "lalonde_rct":
w, t, y = load_lalonde(rct=True, dataroot=args.dataroot)
elif data_name == "lalonde_cps" or data_name == "lalonde_cps1":
w, t, y = load_lalonde(obs_version="cps", dataroot=args.dataroot)
elif data_name.startswith("lbidd"):
# Valid string formats: lbidd_<link>_<n> and lbidd_<link>_<n>_counterfactual
# Valid <link> options: linear, quadratic, cubic, exp, and log
# Valid <n> options: 1k, 2.5k, 5k, 10k, 25k, and 50k
options = data_name.split("_")
link = options[1]
n = options[2]
observe_counterfactuals = (len(options) == 4) and (options[3]
== "counterfactual")
d = load_lbidd(n=n,
observe_counterfactuals=observe_counterfactuals,
link=link,
dataroot=args.dataroot,
return_ate=True,
return_ites=True)
ate = d["ate"]
ites = d['ites']
if observe_counterfactuals:
w, t, y = d["obs_counterfactual_w"], d["obs_counterfactual_t"], d[
"obs_counterfactual_y"]
else:
w, t, y = d["w"], d["t"], d["y"]
elif data_name == "ihdp":
d = load_ihdp(return_ate=True, return_ites=True)
w, t, y, ate, ites = d["w"], d["t"], d["y"], d['ate'], d['ites']
elif data_name == "ihdp_counterfactual":
d = load_ihdp(observe_counterfactuals=True)
w, t, y = d["w"], d["t"], d["y"]
elif data_name == "twins":
d = load_twins(dataroot=args.dataroot)
w, t, y = d["w"], d["t"], d["y"]
else:
raise (Exception("dataset {} not implemented".format(args.data)))
return ites, ate, w, t, y
loss_y = self.outcome_distribution.loss(y, y_)
loss = loss_t + loss_y
return loss, loss_t, loss_y
if __name__ == "__main__":
from data.lalonde import load_lalonde
import matplotlib.pyplot as plt
import pprint
pp = pprint.PrettyPrinter(indent=4)
dataset = 1
network_params = _DEFAULT_TARNET.copy()
if dataset == 1:
w, t, y = load_lalonde()
dist = distributions.MixedDistribution([0.0], distributions.LogLogistic())
training_params = TrainingParams(lr=0.0005, batch_size=128, num_epochs=100, verbose=False)
early_stop = True
ignore_w = False
elif dataset == 2:
w, t, y = load_lalonde(rct=True)
# dist = distributions.MixedDistribution([0.0], distributions.LogLogistic())
dist = distributions.FactorialGaussian()
training_params = TrainingParams(lr=0.001, batch_size=64, num_epochs=200)
early_stop = True
ignore_w = False
elif dataset == 3:
w, t, y = load_lalonde(obs_version="cps1")
dist = distributions.MixedDistribution(
[0.0, 25564.669921875], distributions.SigmoidFlow(ndim=10)
def test_lalonde_torch():
(w, t, y) = load_lalonde(data_format=TORCH)
assert all(isinstance(x, torch.Tensor) for x in (w, t, y))
def test_lalonde_pandas_single():
df = load_lalonde(data_format=PANDAS_SINGLE)
assert isinstance(df, pd.DataFrame)
def test_lalonde_pandas():
w, t, y = load_lalonde(data_format=PANDAS)
assert isinstance(w, pd.DataFrame) and isinstance(
t, pd.Series) and isinstance(y, pd.Series)
