În noua versiune (v0.18.0) din ahead pachet, am extins forecast::thetaf Funcție pentru a sprijini modelele liniare generalizate (GLMS) și a adăugat un mecanism de atenție.
Atenția este utilizată pe scară largă în rețelele neuronale curente (pentru că tind să uite; învinovăți -o pe gradienți 🙂) pentru a se concentra pe anumite părți ale datelor de intrare atunci când se fac predicții.
În acest caz, ajută modelul să învețe care părți ale seriei de timp sunt mai importante pentru prognoză, folosind medii ponderate ale observațiilor trecute.
Mai multe despre asta mai târziu într -o hârtie. La sfârșitul acestui post este prezentat un link către un caiet care conține Python și R.
Începeți cu:
options(repos = c(
techtonique = "https://r-packages.techtonique.net",
CRAN = "https://cloud.r-project.org"
))
install.packages("ahead")
library(forecast) library(ahead) # glm.nb par(mfrow=c(2,1)) obj1 <- suppressWarnings(ahead::glmthetaf(USAccDeaths, h=25L, fit_func=MASS::glm.nb, attention = TRUE, type_pi = "conformal-split", method = "adj")) plot(obj1, main="With attention") obj2 <- suppressWarnings(ahead::glmthetaf(USAccDeaths, h=25L, fit_func=MASS::glm.nb, attention = FALSE, type_pi = "conformal-split", method = "adj")) plot(obj2, main="Without attention") # glm par(mfrow=c(2,1)) obj1 <- suppressWarnings(ahead::glmthetaf(USAccDeaths, h=25L, fit_func=stats::glm, attention = TRUE, type_pi = "conformal-split", method = "adj")) plot(obj1, main="With attention") obj2 <- suppressWarnings(ahead::glmthetaf(USAccDeaths, h=25L, fit_func=stats::glm, attention = FALSE, type_pi = "conformal-split", method = "adj")) plot(obj2, main="Without attention") # rlm par(mfrow=c(2,1)) obj1 <- suppressWarnings(ahead::glmthetaf(USAccDeaths, h=25L, fit_func=MASS::rlm, attention = TRUE, type_pi = "conformal-split", method = "adj")) plot(obj1, main="With attention") obj2 <- suppressWarnings(ahead::glmthetaf(USAccDeaths, h=25L, fit_func=MASS::rlm, attention = FALSE, type_pi = "conformal-split", method = "adj")) plot(obj2, main="Without attention") # lqs par(mfrow=c(2,1)) obj1 <- suppressWarnings(ahead::glmthetaf(USAccDeaths, h=25L, fit_func=MASS::lqs, attention = TRUE, type_pi = "conformal-split", method = "adj")) plot(obj1, main="With attention") obj2 <- suppressWarnings(ahead::glmthetaf(USAccDeaths, h=25L, fit_func=MASS::lqs, attention = FALSE, type_pi = "conformal-split", method = "adj")) plot(obj2, main="Without attention") # lm par(mfrow=c(2,1)) obj1 <- suppressWarnings(ahead::glmthetaf(USAccDeaths, h=25L, fit_func=stats::lm, attention = TRUE, type_pi = "conformal-split", method = "adj")) plot(obj1, main="With attention") obj2 <- suppressWarnings(ahead::glmthetaf(USAccDeaths, h=25L, fit_func=stats::lm, attention = FALSE, type_pi = "conformal-split", method = "adj")) plot(obj2, main="Without attention") # gam par(mfrow=c(2,1)) obj1 <- suppressWarnings(ahead::glmthetaf(USAccDeaths, h=25L, fit_func=gam::gam, attention = TRUE, type_pi = "conformal-split", method = "adj")) plot(obj1, main="With attention") obj2 <- suppressWarnings(ahead::glmthetaf(USAccDeaths, h=25L, fit_func=gam::gam, attention = FALSE, type_pi = "conformal-split", method = "adj")) plot(obj2, main="Without attention") # rq par(mfrow=c(2,1)) obj1 <- suppressWarnings(ahead::glmthetaf(USAccDeaths, h=25L, fit_func=quantreg::rq, attention = TRUE, type_pi = "conformal-split", method = "adj")) plot(obj1, main="With attention") obj2 <- suppressWarnings(ahead::glmthetaf(USAccDeaths, h=25L, fit_func=quantreg::rq, attention = FALSE, type_pi = "conformal-split", method = "adj")) plot(obj2, main="Without attention")
# glm.nb par(mfrow=c(2,1)) obj1 <- suppressWarnings(ahead::glmthetaf(AirPassengers, h=25L, fit_func=MASS::glm.nb, attention = TRUE, type_pi = "conformal-split", method = "adj")) plot(obj1, main="With attention") obj2 <- suppressWarnings(ahead::glmthetaf(AirPassengers, h=25L, fit_func=MASS::glm.nb, attention = FALSE, type_pi = "conformal-split", method = "adj")) plot(obj2, main="Without attention") # glm par(mfrow=c(2,1)) obj1 <- suppressWarnings(ahead::glmthetaf(AirPassengers, h=25L, fit_func=stats::glm, attention = TRUE, type_pi = "conformal-split", method = "adj")) plot(obj1, main="With attention") obj2 <- suppressWarnings(ahead::glmthetaf(AirPassengers, h=25L, fit_func=stats::glm, attention = FALSE, type_pi = "conformal-split", method = "adj")) plot(obj2, main="Without attention") # rlm par(mfrow=c(2,1)) obj1 <- suppressWarnings(ahead::glmthetaf(AirPassengers, h=25L, fit_func=MASS::rlm, attention = TRUE, type_pi = "conformal-split", method = "adj")) plot(obj1, main="With attention") obj2 <- suppressWarnings(ahead::glmthetaf(AirPassengers, h=25L, fit_func=MASS::rlm, attention = FALSE, type_pi = "conformal-split", method = "adj")) plot(obj2, main="Without attention") # lm par(mfrow=c(2,1)) obj1 <- suppressWarnings(ahead::glmthetaf(AirPassengers, h=25L, fit_func=stats::lm, attention = TRUE, type_pi = "conformal-split", method = "adj")) plot(obj1, main="With attention") obj2 <- suppressWarnings(ahead::glmthetaf(AirPassengers, h=25L, fit_func=stats::lm, attention = FALSE, type_pi = "conformal-split", method = "adj")) plot(obj2, main="Without attention") # lqs par(mfrow=c(2,1)) obj1 <- suppressWarnings(ahead::glmthetaf(AirPassengers, h=25L, fit_func=MASS::lqs, attention = TRUE, type_pi = "conformal-split", method = "adj")) plot(obj1, main="With attention") obj2 <- suppressWarnings(ahead::glmthetaf(AirPassengers, h=25L, fit_func=MASS::lqs, attention = FALSE, type_pi = "conformal-split", method = "adj")) plot(obj2, main="Without attention") # gam par(mfrow=c(2,1)) obj1 <- suppressWarnings(ahead::glmthetaf(AirPassengers, h=25L, fit_func=gam::gam, attention = TRUE, type_pi = "conformal-split", method = "adj")) plot(obj1, main="With attention") obj2 <- suppressWarnings(ahead::glmthetaf(AirPassengers, h=25L, fit_func=gam::gam, attention = FALSE, type_pi = "conformal-split", method = "adj")) plot(obj2, main="Without attention") # rq par(mfrow=c(2,1)) obj1 <- suppressWarnings(ahead::glmthetaf(AirPassengers, h=25L, fit_func=quantreg::rq, attention = TRUE, type_pi = "conformal-split", method = "adj")) plot(obj1, main="With attention") obj2 <- suppressWarnings(ahead::glmthetaf(AirPassengers, h=25L, fit_func=quantreg::rq, attention = FALSE, type_pi = "conformal-split", method = "adj")) plot(obj2, main="Without attention")
from rpy2.robjects.packages import importr
from rpy2.robjects import r
import rpy2.robjects as ro
import numpy as np
import matplotlib.pyplot as plt
from rpy2.robjects import pandas2ri
# Import required R packages
ahead = importr('ahead')
mass = importr('MASS')
base = importr('base')
stats = importr('stats')
# Get the data and fit the model
with localconverter(ro.default_converter):
# Get AirPassengers data
data = r('USAccDeaths')
data = np.array(data)
# Fit the model
fit = r('''
suppressWarnings(
ahead::glmthetaf(
USAccDeaths,
h=25L,
fit_func=MASS::glm.nb,
attention=TRUE,
type_pi="conformal-split",
method="adj"
)
)
''')
# Extract predictions and intervals
forecasts = np.array(fit.rx2('mean'))
lower = np.array(fit.rx2('lower'))
upper = np.array(fit.rx2('upper'))
# Create time indices
time_train = np.arange(len(data))
time_test = np.arange(len(data), len(data) + len(forecasts))
# Create the plot
plt.figure(figsize=(12, 6))
# Plot training data
plt.plot(time_train, data, 'b-', label="Observed", alpha=0.7)
# Plot forecasts and prediction intervals
plt.plot(time_test, forecasts, 'r--', label="Forecast")
plt.fill_between(time_test, lower, upper,
color="r", alpha=0.2,
label="95% Prediction Interval")
# Customize the plot
plt.title('USAccDeaths Forecast with Attention')
plt.xlabel('Time')
plt.ylabel('-')
plt.legend()
plt.grid(True, alpha=0.3)
# Show the plot
plt.tight_layout()
plt.show()
