RAMitchell

import numpy as np
import xgboost as xgb
import time

print("Xgboost version: {}".format(xgb.__version__))

n = 10000000
query_groups = 100000
group_size = n / query_groups
m = 10

RAMitchell

import cupy as cp
import GPUtil
import xgboost as xgb
import time

print("Xgboost version: {}".format(xgb.__version__))

n_train = 10000
n_test = 1000
iterations = 20

RAMitchell

import numpy as np
import GPUtil
import xgboost as xgb

print("Xgboost version: {}".format(xgb.__version__))
n = 10000
m = 100
X = np.random.randn(n, m)
y = np.random.randn(n)
exp_models = []

RAMitchell

import unittest
import numpy as np
from sklearn.metrics import log_loss
import xgboost as xgb
print("Xgboost version: {}".format(xgb.__version__))

np.random.seed(1994)
kRows = 1000
kCols = 64
kClasses = 4

RAMitchell

import cupy
import torch
import GPUtil
import xgboost as xgb

from torch.utils.dlpack import to_dlpack
from torch.utils.dlpack import from_dlpack

mem_before = GPUtil.getGPUs()[0].memoryUsed
# Create a PyTorch tensor.

RAMitchell

import xgboost as xgb
import cupy as cp
import GPUtil
import numpy as np
import matplotlib.pyplot as plt
import seaborn as sns

plt.style.use("seaborn")
dmatrix_type = [xgb.DMatrix, xgb.DeviceQuantileDMatrix]
test_sizes = [(300000, 1000), (30000, 10000)]

RAMitchell

library(Rcpp)
# set the compiler flags in ~/Documents/.R/Makevars or ~/Documents/.R/Makevars
# by adding the following line to it
# CXXFLAGS=-O3 -Wall -mtune=native -funroll-loops -mavx -mfma
sourceCpp("avx_test.cc")

curve(approximate_sigmoid(x) - 1/(1 + exp(-x)), -12, 12, n = 1000); grid()

# the exp4096 mostly underestimates the exp
curve(log(exp4096(x)) - x, -9, 9, n = 1000); grid()