-
Notifications
You must be signed in to change notification settings - Fork 0
/
csqte_sims.py
171 lines (160 loc) · 7.64 KB
/
csqte_sims.py
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
# Reproduces Figure 2 from "Robust and Agnostic Learning of Conditional Distributional Treatment Effects"
# Generic imports
import numpy as np
import os
from tqdm import tqdm
from econml.sklearn_extensions.linear_model import StatsModelsLinearRegression
from sklearn.linear_model import LogisticRegression
from sklearn.ensemble import RandomForestRegressor
from sklearn_quantile import RandomForestQuantileRegressor
from joblib import Parallel, delayed
import matplotlib
import matplotlib.pyplot as plt
matplotlib.use('Agg')
# Custom imports
from utils import write_results_to_file, PlottingSuite, ggplot_style_log
from nuisance import (
RFKernel, RBFKernel, KernelSuperquantileRegressor
)
from dgp import lognormal_dgp, lognormal_true_effect, run_simulation
from cdte import CSQTE
# Get true OLS coefs
def get_true_OLS_coefs(n, p, cond_mean_func, cond_std_func, tau, tail):
X = np.random.uniform(0, 1, size=(n, p))
true_effects = lognormal_true_effect(X, cond_mean_func, cond_std_func, tau)
ols = StatsModelsLinearRegression(cov_type="nonrobust")
ols.fit(X, true_effects[f"superquantile_{tail}"])
return ols.coef_
# DGP Definition
cond_mean_func = lambda X, A: X[:, 0] + X[:, 1]*A
cond_std_func = lambda X, A: 0.2
propensity_func = lambda X: 6*(X[:, 0]-0.5)
tail="right"
dgp="lognormal"
np.random.seed(1)
p = 10
tau = 0.75
n_test = 500
X_test = np.random.uniform(0, 1, size=(n_test, p))
true_effects = lognormal_true_effect(X_test, cond_mean_func, cond_std_func, tau)
true_coefs = get_true_OLS_coefs(500000, p, cond_mean_func, cond_std_func, tau, tail)
# Define nuisances and last stages to loop over
nuiss = ["SQRF", "OLS", "SQRF_RBF"]
final_stages = ["OLS", "RF"]
# Define experiment parameters
n_iter = 100
ns = [100, 200, 400, 800, 1600, 3200, 6400, 12800]
for nuis in nuiss:
for final_stage in final_stages:
print(f"Running CSQTE experiments with nuisance {nuis} and final stage {final_stage}...")
for n in tqdm(ns):
#############
# Nuisances #
#############
# Propensity and quantile model
propensity_model = LogisticRegression(solver="saga", penalty="elasticnet", l1_ratio=0.5)
quantile_model = RandomForestQuantileRegressor(
n_estimators=50,
min_samples_leaf=0.05,
q=tau)
# Superquantile model
if nuis == "SQRF":
superquantile_model = KernelSuperquantileRegressor(
kernel=RFKernel(
RandomForestRegressor(n_estimators=50,
min_samples_leaf=0.05)),
tau=tau,
tail=tail)
elif nuis == "OLS":
superquantile_model = StatsModelsLinearRegression()
elif nuis == "SQRF_RBF":
# RBF scale via Silverman's rule of thumb
RBF_scale = 0.9*n**(-1/(4+p))
superquantile_model = KernelSuperquantileRegressor(
kernel=RBFKernel(scale=RBF_scale),
tau=tau,
tail=tail)
# Final stage model
if final_stage == "OLS":
csqte_model = StatsModelsLinearRegression(cov_type="HC1")
elif final_stage == "RF":
csqte_model = RandomForestRegressor(n_estimators=50, min_samples_leaf=0.05)
###################
# CSQTE Estimator #
###################
CSQTE_est = CSQTE(
propensity_model=propensity_model,
quantile_model=quantile_model,
superquantile_model=superquantile_model,
csqte_model=csqte_model,
nested_quantiles=True if nuis=="OLS" else False,
tau=tau,
tail=tail)
trained_models = Parallel(n_jobs=-2, backend='loky', verbose=1)(
delayed(run_simulation)(lognormal_dgp, CSQTE_est,
random_state,
n=n, p=p,
cond_mean_func=cond_mean_func,
cond_std_func=cond_std_func,
propensity_func=propensity_func,
) for random_state in np.arange(n_iter))
# Save results for future usage
write_results_to_file(
trained_models,
X_test, true_effects, true_coefs,
n, p, tau, tail=tail, nuis=nuis, final_stage=final_stage, dgp=dgp)
# Make plots, one per nuisance class
for nuis in nuiss:
if not os.path.exists("plots"):
os.makedirs("plots")
#############################
# Process results from file #
#############################
final_stage = "OLS"
ps = PlottingSuite(n_iter, ns, p, tau, tail, nuis, final_stage, dgp)
mse_ols = ps.plot_mse(show_plugin_final=True, log_yscale=True, show=False)
coverage_ols = ps.plot_coverage(coef_idx=1, show_plugin_final=True, show=False)
final_stage = "RF"
ps = PlottingSuite(n_iter, ns, p, tau, tail, nuis, final_stage, dgp)
mse_rf = ps.plot_mse(show_plugin_final=True, log_yscale=True, show=False)
#############
# MSE Plots #
#############
plugin_all = True
ax = ggplot_style_log(figsize=(7, 4), log_y=True)
# CSQTE + OLS
plt.plot(ns, mse_ols[0][0], label="CSQTE+OLS", color="C0", zorder=5)
plt.fill_between(ns, mse_ols[0][0] - mse_ols[0][1], mse_ols[0][0] + mse_ols[0][1], alpha=0.3, color="C0", zorder=4)
# CSQTE + RF
plt.plot(ns, mse_rf[0][0], label="CSQTE+RF", color="C3", zorder=5)
plt.fill_between(ns, mse_rf[0][0] - mse_rf[0][1], mse_rf[0][0] + mse_rf[0][1], alpha=0.3, color="C3", zorder=4)
# Plugin
plt.plot(ns, mse_ols[1][0], label="Plugin", color="black", ls="--", zorder=5)
plt.fill_between(ns, mse_ols[1][0] - mse_ols[1][1], mse_ols[1][0] + mse_ols[1][1], alpha=0.3, color="gray", zorder=4)
if plugin_all:
# Plugin + OLS
plt.plot(ns, mse_ols[2][0], label="Plugin+OLS", color="C0", ls='--', zorder=5)
plt.fill_between(ns, mse_ols[2][0] - mse_ols[2][1], mse_ols[2][0] + mse_ols[2][1], alpha=0.3, color="C0")
# Plugin + RF
plt.plot(ns, mse_rf[2][0], label="Plugin+RF", color="C3", ls='--', zorder=5)
plt.fill_between(ns, mse_rf[2][0] - mse_rf[2][1], mse_rf[2][0] + mse_rf[2][1], alpha=0.3, color="C3", zorder=4)
ax.set_xlabel("n")
ax.set_ylabel("Mean Squared Error")
#ax.set_yticks([0.3, 1, 3, 10, 30], [0.3, 1, 3, 10, 30])
ax.set_yticks([0.3, 1, 3], [0.3, 1, 3])
plt.legend(loc=(1.02, 0.44))
plt.savefig(f"plots/MSE_CSQTE_tau_{tau}_tail_{tail}_nuis_{nuis}_plugin_all_{plugin_all}.pdf", bbox_inches="tight", dpi=100)
##################
# Coverage Plots #
##################
ax = ggplot_style_log(figsize=(7, 4), log_y=False)
plt.plot(ns, coverage_ols[0][0], label="CSQTE+OLS", zorder=5)
plt.fill_between(ns, coverage_ols[0][0] - coverage_ols[0][1], coverage_ols[0][0] + coverage_ols[0][1], alpha=0.3, zorder=4)
plt.plot(ns, coverage_ols[1][0], label="Plugin+OLS", zorder=5, color="black", ls="--")
plt.fill_between(ns, coverage_ols[1][0] - coverage_ols[1][1], coverage_ols[1][0] + coverage_ols[1][1],
alpha=0.3, zorder=4, color="gray")
ax.set_xlabel("n")
ax.set_ylabel("Coverage")
plt.yticks([0, 0.2, 0.4, 0.6, 0.8, 1])
plt.legend(loc=(1.02, 0.77))
plt.savefig(f"plots/Coverage_CSQTE_tau_{tau}_tail_{tail}_nuis_{nuis}.pdf", bbox_inches="tight", dpi=100)