def ___rmse_over_ratio(rmse, weight, outfolder):
stats = rmse[["ratio", "rmse_best"]]
filename = "%s/praise_rmse_over_ratio_%.1f" % (outfolder, weight)
np.savetxt(filename, stats, fmt="%.2f\t&\t%.8f", newline="\t\\\\\n")
data = stats.set_index("ratio")
min_ratio = data.idxmin()
Plotter.plot_data_over_ratios(data,
"rmse_%.1f" % weight,
outfolder,
ylog=True)
stats_extra = rmse[["ratio", "rmse_extra_best"]]
filename_extra = "%s/praise_extra_rmse_over_ratio_%.1f" % (outfolder,
weight)
np.savetxt(filename_extra,
stats_extra,
fmt="%.2f\t&\t%.8f",
newline="\t\\\\\n")
data_extra = stats_extra.set_index("ratio")
min_ratio_extra = data_extra.idxmin()
Plotter.plot_data_over_ratios(data_extra,
"rmse_extra_%.1f" % weight,
outfolder,
ylog=True)
return min_ratio.values[0], min_ratio_extra.values[0]
def save_feature_importances(self, outfolder):
# compute feature importances for each weight
importances = np.empty(shape=(0, 0))
for weight in self.dataset.weights:
lstats = self.dataset.lstats[weight]
clsset = ClassificationSet.sanitize_and_init(
self.features.features, lstats.winners, lstats.costs)
clf = ExtraTreesClassifier()
clf = clf.fit(clsset.X, clsset.y)
if importances.shape[0] == 0:
importances = clf.feature_importances_
else:
importances = np.vstack(
[importances, clf.feature_importances_])
# sort feature names by average importance
sorted_feature_names = [
name for _, name in sorted(
zip(importances.mean(axis=0), self.features.features.columns))
][::-1]
importances = pd.DataFrame(data=importances,
columns=self.features.features.columns)
importances = importances[sorted_feature_names]
feats = pd.DataFrame(columns=["order", "value", "name", "error"]) #, \
#dtype={"order": np.int64, "value": np.float_, "name":np.object_, "error": np.float_})
feats["order"] = np.arange(len(self.features.features.columns))[::-1]
feats["value"] = importances.mean(axis=0).values
feats["error"] = importances.std(axis=0).values
feats["name"] = sorted_feature_names
feats.to_csv("%s/feats" % outfolder,
sep="&",
index=False,
line_terminator="\\\\\n")
Plotter.plot_feature_importances(importances, outfolder, 30)
def fit_time(raw_sample_stats, outfolder="/tmp"):
sample_times = pd.DataFrame()
for ratio in raw_sample_stats.ratios:
sample_times = sample_times.append(
raw_sample_stats.get_times(ratio).agg("mean"),
ignore_index=True)
sample_times = sample_times.set_index(raw_sample_stats.ratios)
# save xdata, ydata to file
sample_times.to_csv("%s/times_over_ratio" % outfolder,
sep='&',
line_terminator='\\\\\n')
# plot welfare over ratios
Plotter.plot_data_over_ratios(sample_times, "time", outfolder)
xdata = 10000 * sample_times.index
for algo in sample_times.columns:
print('==== %s ====' % algo)
ydata = sample_times[algo]
popt, pcov = curve_fit(func_nlogn,
xdata,
ydata,
bounds=(0, np.inf))
print("time nlogn:", popt,
np.sqrt(np.diag(pcov)) * 100. / popt, "%")
popt, pcov = curve_fit(func_poly2,
xdata,
ydata,
bounds=(0, np.inf))
print("time n2:", popt, np.sqrt(np.diag(pcov)) * 100. / popt, "%")
popt, pcov = curve_fit(func_poly3,
xdata,
ydata,
bounds=(0, np.inf))
print("time n3:", popt, np.sqrt(np.diag(pcov)) * 100. / popt, "%")
popt, pcov = curve_fit(func_poly321,
xdata,
ydata,
bounds=(0, np.inf))
print("time n321:", popt,
np.sqrt(np.diag(pcov)) * 100. / popt, "%")
popt, pcov = curve_fit(func_n3logn,
xdata,
ydata,
bounds=(0, np.inf))
print("time n3logn:", popt,
np.sqrt(np.diag(pcov)) * 100. / popt, "%")
popt, pcov = curve_fit(func_n2logn,
xdata,
ydata,
bounds=(0, np.inf))
print("time n2logn:", popt,
np.sqrt(np.diag(pcov)) * 100. / popt, "%")
popt, pcov = curve_fit(func_nlogn_n,
xdata,
ydata,
bounds=(0, np.inf))
print("time nlogn+n:", popt,
np.sqrt(np.diag(pcov)) * 100. / popt, "%")
def plot(self, outfolder="/tmp"):
outfile = "%s/random_%s" % (outfolder, self.name)
welfares = self.__get_normalized_welfares()
data = []
for algo in self.algos:
data.append(welfares[welfares.algorithm == algo].welfare.values)
print("[%s] min = %.2f, max = %.2f" %
(algo, welfares[welfares.algorithm == algo].welfare.min(),
welfares[welfares.algorithm == algo].welfare.max()))
Plotter.boxplot_random(data, self.algos, outfile)
def plot(self, outfolder="/tmp"):
outfile_welfare = "%s/welfare_%s" % (outfolder, self.name)
outfile_time = "%s/time_%s" % (outfolder, self.name)
welfares = self.get_welfares_feasible()
times = self.get_times_feasible()
# normalize welfare and time by values of optimal algorithm (cplex)
welfares = welfares.div(welfares.CPLEX, axis=0).multiply(100., axis=0)
times = times.div(times.CPLEX, axis=0).multiply(100., axis=0)
Plotter.boxplot_average_case(welfares.values,
self.algos,
outfile_welfare,
ylabel="% of optimal welfare (CPLEX)")
Plotter.boxplot_average_case(
times.values,
self.algos,
outfile_time,
top=100000,
bottom=0.01,
ylog=True,
ylabel="% of time of optimal algorithm (CPLEX)")
def plot(self, outfolder="/tmp"):
Plotter.plot_feature_heatmap(self, outfolder)
def plot(self, outfolder="/tmp"):
Plotter.plot_breakdown(self, outfolder)