Ei Habr!
O le tautoulu talu ai, na talimalo ai Kaggle i se tauvaga e faavasega ai ata tusilima, Quick Draw Doodle Recognition, lea, faatasi ai ma isi, o se vaega o R-saienitisi na auai: , и . Matou te le faʻamatalaina auiliili le tauvaga; ua uma ona faia i totonu .
O le taimi lenei e leʻi manuia i le faʻatoʻaga pine, ae o le tele o mea taua na maua, o lea ou te fia taʻu atu ai i le nuʻu e uiga i le tele o mea sili ona manaia ma aoga i luga o Kagle ma galuega i aso uma. Faatasi ai ma autu na talanoaina: olaga faigata e aunoa ma OpenCV, JSON parsing (o faʻataʻitaʻiga nei o loʻo suʻesuʻeina le tuʻufaʻatasia o le C++ code i tusitusiga poʻo afifi i le R faʻaaoga Rcpp), fa'avasegaina o fa'amaumauga ma fa'amautu o le fofo mulimuli. O code uma mai le fe'au i se fomu e talafeagai mo le fa'atinoina o lo'o avanoa i totonu .
Faʻamatalaga:
1. Tu'u lelei fa'amaumauga mai le CSV i totonu o le MonetDB database
O faʻamaumauga i lenei tauvaga e le tuʻuina atu i foliga o ata ua saunia, ae i le tulaga o 340 CSV faila (tasi faila mo vasega taʻitasi) o loʻo i ai JSONs faʻatasi ai ma faʻamaufaʻailoga. O le fa'afeso'ota'i o nei manatu i laina, tatou te maua ai se ata mulimuli e 256x256 pika. E faʻapea foʻi mo faʻamaumauga taʻitasi o loʻo i ai se faʻailoga e faʻaalia ai pe saʻo le ata na iloa e le faʻavasegaina na faʻaaogaina i le taimi na aoina ai le faʻamaumauga, o se numera lua mataitusi o le atunuu o loʻo nofo ai le tusitala o le ata, o se faʻamatalaga tulaga ese, se faʻailoga taimi. ma se igoa vasega e fetaui ma le igoa faila. O se fa'amatalaga faigofie o fa'amaumauga muamua e mamafa le 7.4 GB i le fa'amaumauga ma pe a ma le 20 GB pe a uma ona tatala, o fa'amaumauga atoa pe a uma ona tatala e ave le 240 GB. Na fa'amautinoa e le au fa'atonu o fa'aliliuga uma e lua na toe faia ata tutusa, o lona uiga o le fa'asologa atoa e le toe fa'aaogaina. I soo se tulaga, o le teuina o le 50 miliona ata i faila kalafi poʻo i le tulaga o faʻasologa na vave lava ona manatu e le aoga, ma na matou filifili e tuʻufaʻatasia uma faila CSV mai le faʻamaumauga. train_simplified.zip i totonu o faʻamaumauga faʻatasi ma faʻasologa mulimuli ane o ata o le tele manaʻomia "i luga o le lele" mo vaega taʻitasi.
O se faiga fa'amaonia lelei na filifilia e pei o le DBMS MonetDB, e taʻua o se faʻatinoga mo R o se afifi . O le afifi e aofia ai se faʻapipiʻi faʻapipiʻi o le database server ma faʻatagaina oe e piki saʻo le server mai se R session ma galulue faʻatasi ai iina. Fausia se faʻamaumauga ma fesoʻotaʻi i ai e faia i le tasi poloaiga:
con <- DBI::dbConnect(drv = MonetDBLite::MonetDBLite(), Sys.getenv("DBDIR"))Matou te manaʻomia le faia o ni laulau se lua: tasi mo faʻamaumauga uma, o le isi mo faʻamatalaga o auaunaga e uiga i faila na sii mai (e aoga pe a tupu se mea ma e tatau ona toe amata le faagasologa pe a uma ona sii mai ni faila):
Fausia laulau
if (!DBI::dbExistsTable(con, "doodles")) {
DBI::dbCreateTable(
con = con,
name = "doodles",
fields = c(
"countrycode" = "char(2)",
"drawing" = "text",
"key_id" = "bigint",
"recognized" = "bool",
"timestamp" = "timestamp",
"word" = "text"
)
)
}
if (!DBI::dbExistsTable(con, "upload_log")) {
DBI::dbCreateTable(
con = con,
name = "upload_log",
fields = c(
"id" = "serial",
"file_name" = "text UNIQUE",
"uploaded" = "bool DEFAULT false"
)
)
}O le auala sili ona vave e utaina ai faʻamatalaga i totonu o faʻamaumauga o le kopi saʻo o faila CSV e faʻaaoga ai le SQL - command COPY OFFSET 2 INTO tablename FROM path USING DELIMITERS ',','n','"' NULL AS '' BEST EFFORTfea tablename - igoa laulau ma path - le ala i le faila. Aʻo galue ma le faʻamaumauga, na maua ai le faʻatinoina o le fausiaina unzip i le R e le galue saʻo ma le tele o faila mai le faʻamaumauga, o lea na matou faʻaogaina ai le faiga unzip (faʻaaogaina le parakalafa getOption("unzip")).
Galuega mo le tusitusi i le database
#' @title Извлечение и загрузка файлов
#'
#' @description
#' Извлечение CSV-файлов из ZIP-архива и загрузка их в базу данных
#'
#' @param con Объект подключения к базе данных (класс `MonetDBEmbeddedConnection`).
#' @param tablename Название таблицы в базе данных.
#' @oaram zipfile Путь к ZIP-архиву.
#' @oaram filename Имя файла внури ZIP-архива.
#' @param preprocess Функция предобработки, которая будет применена извлечённому файлу.
#' Должна принимать один аргумент `data` (объект `data.table`).
#'
#' @return `TRUE`.
#'
upload_file <- function(con, tablename, zipfile, filename, preprocess = NULL) {
# Проверка аргументов
checkmate::assert_class(con, "MonetDBEmbeddedConnection")
checkmate::assert_string(tablename)
checkmate::assert_string(filename)
checkmate::assert_true(DBI::dbExistsTable(con, tablename))
checkmate::assert_file_exists(zipfile, access = "r", extension = "zip")
checkmate::assert_function(preprocess, args = c("data"), null.ok = TRUE)
# Извлечение файла
path <- file.path(tempdir(), filename)
unzip(zipfile, files = filename, exdir = tempdir(),
junkpaths = TRUE, unzip = getOption("unzip"))
on.exit(unlink(file.path(path)))
# Применяем функция предобработки
if (!is.null(preprocess)) {
.data <- data.table::fread(file = path)
.data <- preprocess(data = .data)
data.table::fwrite(x = .data, file = path, append = FALSE)
rm(.data)
}
# Запрос к БД на импорт CSV
sql <- sprintf(
"COPY OFFSET 2 INTO %s FROM '%s' USING DELIMITERS ',','n','"' NULL AS '' BEST EFFORT",
tablename, path
)
# Выполнение запроса к БД
DBI::dbExecute(con, sql)
# Добавление записи об успешной загрузке в служебную таблицу
DBI::dbExecute(con, sprintf("INSERT INTO upload_log(file_name, uploaded) VALUES('%s', true)",
filename))
return(invisible(TRUE))
}Afai e te manaʻomia le suia o le laulau aʻo leʻi tusia i le database, ua lava e pasi i le finauga preprocess galuega e fa'aliliuina ai fa'amaumauga.
Fa'ailoga mo le fa'asolo fa'asologa o fa'amaumauga i totonu o fa'amaumauga:
Tusia fa'amaumauga ile fa'amaumauga
# Список файлов для записи
files <- unzip(zipfile, list = TRUE)$Name
# Список исключений, если часть файлов уже была загружена
to_skip <- DBI::dbGetQuery(con, "SELECT file_name FROM upload_log")[[1L]]
files <- setdiff(files, to_skip)
if (length(files) > 0L) {
# Запускаем таймер
tictoc::tic()
# Прогресс бар
pb <- txtProgressBar(min = 0L, max = length(files), style = 3)
for (i in seq_along(files)) {
upload_file(con = con, tablename = "doodles",
zipfile = zipfile, filename = files[i])
setTxtProgressBar(pb, i)
}
close(pb)
# Останавливаем таймер
tictoc::toc()
}
# 526.141 sec elapsed - копирование SSD->SSD
# 558.879 sec elapsed - копирование USB->SSDOle taimi ole utaina o fa'amatalaga e mafai ona fesuisuia'i e fa'atatau ile saoasaoa uiga ole ta'avale fa'aaoga. I lo matou tulaga, o le faitau ma le tusitusi i totonu o le SSD e tasi poʻo le mai le flash drive (puna faila) i le SSD (DB) e itiiti ifo i le 10 minute.
E mana'omia ni nai sekone e fai ai se koluma ma se fa'ailoga o le vasega numera ma se koluma fa'asino (ORDERED INDEX) fa'atasi ai ma numera o laina e fa'ata'ita'iina ai fa'ata'ita'iga pe a faia ni vaega:
Fausia o Koluma Faaopoopo ma Faasino Upu
message("Generate lables")
invisible(DBI::dbExecute(con, "ALTER TABLE doodles ADD label_int int"))
invisible(DBI::dbExecute(con, "UPDATE doodles SET label_int = dense_rank() OVER (ORDER BY word) - 1"))
message("Generate row numbers")
invisible(DBI::dbExecute(con, "ALTER TABLE doodles ADD id serial"))
invisible(DBI::dbExecute(con, "CREATE ORDERED INDEX doodles_id_ord_idx ON doodles(id)"))Ina ia foia le faafitauli o le fatuina o se vaega i luga o le lele, matou te manaʻomia le ausia o le saoasaoa maualuga o le aveeseina o laina faʻafuaseʻi mai le laulau. doodles. Mo lenei mea na matou faʻaaogaina 3 togafiti. O le mea muamua o le faʻaititia o le dimensionality o le ituaiga o loʻo teuina le ID mataʻituina. I le seti faʻamaumauga muamua, o le ituaiga e manaʻomia e teu ai le ID o bigint, ae o le numera o faʻamatalaga e mafai ai ona faʻafetaui a latou faʻamatalaga, tutusa ma le numera faʻasologa, i le ituaiga int. O le sailiga e sili atu le vave i lenei tulaga. O le togafiti lona lua o le faʻaaogaina ORDERED INDEX - na matou oʻo mai i lenei faʻaiuga faʻapitoa, ina ua uma ona faʻaogaina mea uma na maua . O le lona tolu o le faʻaaogaina o faʻataʻitaʻiga fesili. O le ute o le metotia o le faʻatinoina o le faʻatonuga tasi PREPARE faʻatasi ai ma le faʻaogaina mulimuli ane o se faʻamatalaga saunia pe a fatuina se tele o fesili o le ituaiga tutusa, ae o le mea moni o loʻo i ai se avanoa pe a faʻatusatusa i se faigofie. SELECT na aliali mai o lo'o i totonu ole va'aiga o mea sese fa'afuainumera.
O le faʻagasologa o le tuʻuina atu o faʻamatalaga e le sili atu i le 450 MB o le RAM. O lona uiga, o le auala faʻamatalaina e mafai ai ona e faʻanofoina faʻamaumauga e mamafa le sefulu o gigabytes i luga o le toetoe lava o soʻo se meafaigaluega faʻatupeina, e aofia ai nisi masini laupapa e tasi, lea e manaia tele.
Pau lava le mea o loʻo totoe o le fuaina lea o le saoasaoa o le toe maua mai (faʻafuaseʻi) faʻamaumauga ma iloilo le faʻavasegaina pe a faʻataʻitaʻiina vaega o lapopoa eseese:
Fa'ailoga fa'amaumauga
library(ggplot2)
set.seed(0)
# Подключение к базе данных
con <- DBI::dbConnect(MonetDBLite::MonetDBLite(), Sys.getenv("DBDIR"))
# Функция для подготовки запроса на стороне сервера
prep_sql <- function(batch_size) {
sql <- sprintf("PREPARE SELECT id FROM doodles WHERE id IN (%s)",
paste(rep("?", batch_size), collapse = ","))
res <- DBI::dbSendQuery(con, sql)
return(res)
}
# Функция для извлечения данных
fetch_data <- function(rs, batch_size) {
ids <- sample(seq_len(n), batch_size)
res <- DBI::dbFetch(DBI::dbBind(rs, as.list(ids)))
return(res)
}
# Проведение замера
res_bench <- bench::press(
batch_size = 2^(4:10),
{
rs <- prep_sql(batch_size)
bench::mark(
fetch_data(rs, batch_size),
min_iterations = 50L
)
}
)
# Параметры бенчмарка
cols <- c("batch_size", "min", "median", "max", "itr/sec", "total_time", "n_itr")
res_bench[, cols]
# batch_size min median max `itr/sec` total_time n_itr
# <dbl> <bch:tm> <bch:tm> <bch:tm> <dbl> <bch:tm> <int>
# 1 16 23.6ms 54.02ms 93.43ms 18.8 2.6s 49
# 2 32 38ms 84.83ms 151.55ms 11.4 4.29s 49
# 3 64 63.3ms 175.54ms 248.94ms 5.85 8.54s 50
# 4 128 83.2ms 341.52ms 496.24ms 3.00 16.69s 50
# 5 256 232.8ms 653.21ms 847.44ms 1.58 31.66s 50
# 6 512 784.6ms 1.41s 1.98s 0.740 1.1m 49
# 7 1024 681.7ms 2.72s 4.06s 0.377 2.16m 49
ggplot(res_bench, aes(x = factor(batch_size), y = median, group = 1)) +
geom_point() +
geom_line() +
ylab("median time, s") +
theme_minimal()
DBI::dbDisconnect(con, shutdown = TRUE)
2. Saunia vaega
O le faagasologa atoa o sauniuniga o vaega e aofia ai laasaga nei:
- Fa'asalalau nisi o JSON o lo'o i ai vectors o manoa ma fa'amaopoopo o togi.
- Tusi laina lanu e faʻatatau i faʻamaopoopo o togi i luga o se ata o le tele manaʻomia (mo se faʻataʻitaʻiga, 256x256 poʻo le 128x128).
- Fa'aliliuina ata e maua i se tensor.
I le avea ai o se vaega o le tauvaga i fatu Python, o le faafitauli na foia muamua i le faʻaaogaina OpenCV. O se tasi o analogues sili ona faigofie ma sili ona manino i le R o le a pei o lenei:
Fa'atinoina o le JSON i le Tensor Conversion i le R
r_process_json_str <- function(json, line.width = 3,
color = TRUE, scale = 1) {
# Парсинг JSON
coords <- jsonlite::fromJSON(json, simplifyMatrix = FALSE)
tmp <- tempfile()
# Удаляем временный файл по завершению функции
on.exit(unlink(tmp))
png(filename = tmp, width = 256 * scale, height = 256 * scale, pointsize = 1)
# Пустой график
plot.new()
# Размер окна графика
plot.window(xlim = c(256 * scale, 0), ylim = c(256 * scale, 0))
# Цвета линий
cols <- if (color) rainbow(length(coords)) else "#000000"
for (i in seq_along(coords)) {
lines(x = coords[[i]][[1]] * scale, y = coords[[i]][[2]] * scale,
col = cols[i], lwd = line.width)
}
dev.off()
# Преобразование изображения в 3-х мерный массив
res <- png::readPNG(tmp)
return(res)
}
r_process_json_vector <- function(x, ...) {
res <- lapply(x, r_process_json_str, ...)
# Объединение 3-х мерных массивов картинок в 4-х мерный в тензор
res <- do.call(abind::abind, c(res, along = 0))
return(res)
}O le ata e faia e faʻaaoga ai meafaigaluega masani R ma faʻapolopolo i se PNG le tumau o loʻo teuina i le RAM (i luga o Linux, o loʻo i totonu o le lisi o lisi le tumau R. /tmp, faʻapipiʻi i le RAM). Ona faitau lea o le faila lea o se laina tolu-dimensional ma numera e amata mai i le 0 i le 1. E taua lenei mea ona o se BMP sili atu masani o le a faitau i se laina mata ma lanu lanu hex.
Se'i o tatou su'e le i'uga:
zip_file <- file.path("data", "train_simplified.zip")
csv_file <- "cat.csv"
unzip(zip_file, files = csv_file, exdir = tempdir(),
junkpaths = TRUE, unzip = getOption("unzip"))
tmp_data <- data.table::fread(file.path(tempdir(), csv_file), sep = ",",
select = "drawing", nrows = 10000)
arr <- r_process_json_str(tmp_data[4, drawing])
dim(arr)
# [1] 256 256 3
plot(magick::image_read(arr))
O le vaega lava ia o le a faia e pei ona taua i lalo:
res <- r_process_json_vector(tmp_data[1:4, drawing], scale = 0.5)
str(res)
# num [1:4, 1:128, 1:128, 1:3] 1 1 1 1 1 1 1 1 1 1 ...
# - attr(*, "dimnames")=List of 4
# ..$ : NULL
# ..$ : NULL
# ..$ : NULL
# ..$ : NULLO lenei faʻatinoga na foliga mai e le sili ona lelei ia i matou, talu ai o le fausiaina o vaega tetele e umi se taimi, ma na matou filifili e faʻaoga lelei le poto masani a matou paaga e ala i le faʻaogaina o se faletusi mamana. OpenCV. I lena taimi e leai se afifi saunia mo R (e leai se mea i le taimi nei), o lea la o se faʻatinoga itiiti o galuega manaʻomia na tusia i le C ++ faʻatasi ai ma le tuʻufaʻatasia i le R code faʻaaoga. Rcpp.
Ina ia foia le faafitauli, o afifi ma faletusi na faʻaaogaina:
-
OpenCV mo le galue i ata ma tusi laina. Fa'aaogaina faletusi fa'apipi'i muamua ma faila fa'aulu, fa'apea fo'i feso'ota'iga malosi.
-
xtensor mo le galulue faatasi ma arrays multidimensional ma tensors. Na matou faʻaogaina faila faila o loʻo aofia i le R package o le igoa tutusa. O le faletusi e mafai ai ona e galue i le tele o vaega, e le gata i le laina tele ma le koluma fa'asologa.
-
ndjson mo le fa'avasegaina o le JSON. O lenei faletusi o loʻo faʻaaogaina i xtensor otometi pe a iai i totonu o le poloketi.
-
RcppThread mo le faʻatulagaina o le tele o filo faʻasologa o se vector mai le JSON. Fa'aaogā faila fa'aulu ua saunia e lenei afifi. Mai sili ona lauiloa RcppParallel O le afifi, faatasi ai ma isi mea, o loʻo i ai se masini faʻalavelave faʻaogaina.
E taua le maitauina o xtensor na avea ma se atua: i le faaopoopo atu i le mea moni o loʻo i ai le tele o faʻatinoga ma le maualuga o le faʻatinoga, o ana atinaʻe na foliga mai e tali mai ma taliina fesili vave ma auiliili. Faatasi ai ma la latou fesoasoani, na mafai ai ona faʻatinoina suiga o OpenCV matrices i le xtensor tensors, faʻapea foʻi ma se auala e tuʻufaʻatasia ai ata 3-dimensional tensors i se 4-dimensional tensor o le saʻo (le vaega lava ia).
Mea e a'oa'oina ai le Rcpp, xtensor ma le RcppThread
Ina ia tuʻufaʻatasia faila e faʻaogaina faila faila ma fesoʻotaʻiga malosi ma faletusi faʻapipiʻi i luga o le polokalama, matou te faʻaogaina le masini faʻapipiʻi faʻatinoina i totonu o le afifi. Rcpp. Ina ia otometi ona maua auala ma fuʻa, na matou faʻaaogaina se faʻaoga lauiloa Linux pkg-config.
Fa'atinoina o le Rcpp plugin mo le fa'aogaina o le OpenCV library
Rcpp::registerPlugin("opencv", function() {
# Возможные названия пакета
pkg_config_name <- c("opencv", "opencv4")
# Бинарный файл утилиты pkg-config
pkg_config_bin <- Sys.which("pkg-config")
# Проврека наличия утилиты в системе
checkmate::assert_file_exists(pkg_config_bin, access = "x")
# Проверка наличия файла настроек OpenCV для pkg-config
check <- sapply(pkg_config_name,
function(pkg) system(paste(pkg_config_bin, pkg)))
if (all(check != 0)) {
stop("OpenCV config for the pkg-config not found", call. = FALSE)
}
pkg_config_name <- pkg_config_name[check == 0]
list(env = list(
PKG_CXXFLAGS = system(paste(pkg_config_bin, "--cflags", pkg_config_name),
intern = TRUE),
PKG_LIBS = system(paste(pkg_config_bin, "--libs", pkg_config_name),
intern = TRUE)
))
})O se taunuuga o le faʻaogaina o le plugin, o tau nei o le a suitulaga i le faagasologa o le tuʻufaʻatasia:
Rcpp:::.plugins$opencv()$env
# $PKG_CXXFLAGS
# [1] "-I/usr/include/opencv"
#
# $PKG_LIBS
# [1] "-lopencv_shape -lopencv_stitching -lopencv_superres -lopencv_videostab -lopencv_aruco -lopencv_bgsegm -lopencv_bioinspired -lopencv_ccalib -lopencv_datasets -lopencv_dpm -lopencv_face -lopencv_freetype -lopencv_fuzzy -lopencv_hdf -lopencv_line_descriptor -lopencv_optflow -lopencv_video -lopencv_plot -lopencv_reg -lopencv_saliency -lopencv_stereo -lopencv_structured_light -lopencv_phase_unwrapping -lopencv_rgbd -lopencv_viz -lopencv_surface_matching -lopencv_text -lopencv_ximgproc -lopencv_calib3d -lopencv_features2d -lopencv_flann -lopencv_xobjdetect -lopencv_objdetect -lopencv_ml -lopencv_xphoto -lopencv_highgui -lopencv_videoio -lopencv_imgcodecs -lopencv_photo -lopencv_imgproc -lopencv_core"Ole tulafono ole fa'atinoga mo le fa'avasegaina o le JSON ma le fa'atupuina o se vaega mo le tu'uina atu ile fa'ata'ita'iga o lo'o tu'uina atu i lalo ole fa'aleaga. Muamua, faʻaopoopo se lisi o galuega faʻapitonuʻu e suʻe ai faila faila (manaʻomia mo ndjson):
Sys.setenv("PKG_CXXFLAGS" = paste0("-I", normalizePath(file.path("src"))))Faʻatinoina o le JSON i le tensor liua i C++
// [[Rcpp::plugins(cpp14)]]
// [[Rcpp::plugins(opencv)]]
// [[Rcpp::depends(xtensor)]]
// [[Rcpp::depends(RcppThread)]]
#include <xtensor/xjson.hpp>
#include <xtensor/xadapt.hpp>
#include <xtensor/xview.hpp>
#include <xtensor-r/rtensor.hpp>
#include <opencv2/core/core.hpp>
#include <opencv2/highgui/highgui.hpp>
#include <opencv2/imgproc/imgproc.hpp>
#include <Rcpp.h>
#include <RcppThread.h>
// Синонимы для типов
using RcppThread::parallelFor;
using json = nlohmann::json;
using points = xt::xtensor<double,2>; // Извлечённые из JSON координаты точек
using strokes = std::vector<points>; // Извлечённые из JSON координаты точек
using xtensor3d = xt::xtensor<double, 3>; // Тензор для хранения матрицы изоображения
using xtensor4d = xt::xtensor<double, 4>; // Тензор для хранения множества изображений
using rtensor3d = xt::rtensor<double, 3>; // Обёртка для экспорта в R
using rtensor4d = xt::rtensor<double, 4>; // Обёртка для экспорта в R
// Статические константы
// Размер изображения в пикселях
const static int SIZE = 256;
// Тип линии
// См. https://en.wikipedia.org/wiki/Pixel_connectivity#2-dimensional
const static int LINE_TYPE = cv::LINE_4;
// Толщина линии в пикселях
const static int LINE_WIDTH = 3;
// Алгоритм ресайза
// https://docs.opencv.org/3.1.0/da/d54/group__imgproc__transform.html#ga5bb5a1fea74ea38e1a5445ca803ff121
const static int RESIZE_TYPE = cv::INTER_LINEAR;
// Шаблон для конвертирования OpenCV-матрицы в тензор
template <typename T, int NCH, typename XT=xt::xtensor<T,3,xt::layout_type::column_major>>
XT to_xt(const cv::Mat_<cv::Vec<T, NCH>>& src) {
// Размерность целевого тензора
std::vector<int> shape = {src.rows, src.cols, NCH};
// Общее количество элементов в массиве
size_t size = src.total() * NCH;
// Преобразование cv::Mat в xt::xtensor
XT res = xt::adapt((T*) src.data, size, xt::no_ownership(), shape);
return res;
}
// Преобразование JSON в список координат точек
strokes parse_json(const std::string& x) {
auto j = json::parse(x);
// Результат парсинга должен быть массивом
if (!j.is_array()) {
throw std::runtime_error("'x' must be JSON array.");
}
strokes res;
res.reserve(j.size());
for (const auto& a: j) {
// Каждый элемент массива должен быть 2-мерным массивом
if (!a.is_array() || a.size() != 2) {
throw std::runtime_error("'x' must include only 2d arrays.");
}
// Извлечение вектора точек
auto p = a.get<points>();
res.push_back(p);
}
return res;
}
// Отрисовка линий
// Цвета HSV
cv::Mat ocv_draw_lines(const strokes& x, bool color = true) {
// Исходный тип матрицы
auto stype = color ? CV_8UC3 : CV_8UC1;
// Итоговый тип матрицы
auto dtype = color ? CV_32FC3 : CV_32FC1;
auto bg = color ? cv::Scalar(0, 0, 255) : cv::Scalar(255);
auto col = color ? cv::Scalar(0, 255, 220) : cv::Scalar(0);
cv::Mat img = cv::Mat(SIZE, SIZE, stype, bg);
// Количество линий
size_t n = x.size();
for (const auto& s: x) {
// Количество точек в линии
size_t n_points = s.shape()[1];
for (size_t i = 0; i < n_points - 1; ++i) {
// Точка начала штриха
cv::Point from(s(0, i), s(1, i));
// Точка окончания штриха
cv::Point to(s(0, i + 1), s(1, i + 1));
// Отрисовка линии
cv::line(img, from, to, col, LINE_WIDTH, LINE_TYPE);
}
if (color) {
// Меняем цвет линии
col[0] += 180 / n;
}
}
if (color) {
// Меняем цветовое представление на RGB
cv::cvtColor(img, img, cv::COLOR_HSV2RGB);
}
// Меняем формат представления на float32 с диапазоном [0, 1]
img.convertTo(img, dtype, 1 / 255.0);
return img;
}
// Обработка JSON и получение тензора с данными изображения
xtensor3d process(const std::string& x, double scale = 1.0, bool color = true) {
auto p = parse_json(x);
auto img = ocv_draw_lines(p, color);
if (scale != 1) {
cv::Mat out;
cv::resize(img, out, cv::Size(), scale, scale, RESIZE_TYPE);
cv::swap(img, out);
out.release();
}
xtensor3d arr = color ? to_xt<double,3>(img) : to_xt<double,1>(img);
return arr;
}
// [[Rcpp::export]]
rtensor3d cpp_process_json_str(const std::string& x,
double scale = 1.0,
bool color = true) {
xtensor3d res = process(x, scale, color);
return res;
}
// [[Rcpp::export]]
rtensor4d cpp_process_json_vector(const std::vector<std::string>& x,
double scale = 1.0,
bool color = false) {
size_t n = x.size();
size_t dim = floor(SIZE * scale);
size_t channels = color ? 3 : 1;
xtensor4d res({n, dim, dim, channels});
parallelFor(0, n, [&x, &res, scale, color](int i) {
xtensor3d tmp = process(x[i], scale, color);
auto view = xt::view(res, i, xt::all(), xt::all(), xt::all());
view = tmp;
});
return res;
}O lenei code e tatau ona tuʻu i totonu o le faila src/cv_xt.cpp ma tuufaatasia ma le poloaiga Rcpp::sourceCpp(file = "src/cv_xt.cpp", env = .GlobalEnv); e manaʻomia foʻi mo galuega nlohmann/json.hpp mai . Ua vaevaeina le code i ni nai galuega:
-
to_xt- o se galuega faʻataʻitaʻi mo le suia o se ata matrix (cv::Mat) i se tensorxt::xtensor; -
parse_json- o le galuega e faʻapipiʻi se manoa JSON, faʻapipiʻi faʻamaopoopo o togi, faʻapipiʻi i totonu o se vector; -
ocv_draw_lines- mai le fua o le vector o togi, tusi laina lanu-lanu; -
process- tuʻufaʻatasia galuega o loʻo i luga ma faʻaopoopoina foi le gafatia e fua ai le ata e maua; -
cpp_process_json_str- afifi i luga o le galuegaprocess, lea e auina atu i fafo le taunuuga i se R-meafaitino (faasologa tele); -
cpp_process_json_vector- afifi i luga o le galuegacpp_process_json_str, lea e fa'atagaina ai oe e fa'agasolo se ve'a manoa i le fa'asologa o filo fa'atele.
Ina ia tusia laina lanu-lanu, na faʻaaogaina le ata lanu HSV, sosoo ai ma le liua i le RGB. Se'i o tatou su'e le i'uga:
arr <- cpp_process_json_str(tmp_data[4, drawing])
dim(arr)
# [1] 256 256 3
plot(magick::image_read(arr))
Faʻatusatusaga o le saoasaoa o faʻatinoga i R ma C ++
res_bench <- bench::mark(
r_process_json_str(tmp_data[4, drawing], scale = 0.5),
cpp_process_json_str(tmp_data[4, drawing], scale = 0.5),
check = FALSE,
min_iterations = 100
)
# Параметры бенчмарка
cols <- c("expression", "min", "median", "max", "itr/sec", "total_time", "n_itr")
res_bench[, cols]
# expression min median max `itr/sec` total_time n_itr
# <chr> <bch:tm> <bch:tm> <bch:tm> <dbl> <bch:tm> <int>
# 1 r_process_json_str 3.49ms 3.55ms 4.47ms 273. 490ms 134
# 2 cpp_process_json_str 1.94ms 2.02ms 5.32ms 489. 497ms 243
library(ggplot2)
# Проведение замера
res_bench <- bench::press(
batch_size = 2^(4:10),
{
.data <- tmp_data[sample(seq_len(.N), batch_size), drawing]
bench::mark(
r_process_json_vector(.data, scale = 0.5),
cpp_process_json_vector(.data, scale = 0.5),
min_iterations = 50,
check = FALSE
)
}
)
res_bench[, cols]
# expression batch_size min median max `itr/sec` total_time n_itr
# <chr> <dbl> <bch:tm> <bch:tm> <bch:tm> <dbl> <bch:tm> <int>
# 1 r 16 50.61ms 53.34ms 54.82ms 19.1 471.13ms 9
# 2 cpp 16 4.46ms 5.39ms 7.78ms 192. 474.09ms 91
# 3 r 32 105.7ms 109.74ms 212.26ms 7.69 6.5s 50
# 4 cpp 32 7.76ms 10.97ms 15.23ms 95.6 522.78ms 50
# 5 r 64 211.41ms 226.18ms 332.65ms 3.85 12.99s 50
# 6 cpp 64 25.09ms 27.34ms 32.04ms 36.0 1.39s 50
# 7 r 128 534.5ms 627.92ms 659.08ms 1.61 31.03s 50
# 8 cpp 128 56.37ms 58.46ms 66.03ms 16.9 2.95s 50
# 9 r 256 1.15s 1.18s 1.29s 0.851 58.78s 50
# 10 cpp 256 114.97ms 117.39ms 130.09ms 8.45 5.92s 50
# 11 r 512 2.09s 2.15s 2.32s 0.463 1.8m 50
# 12 cpp 512 230.81ms 235.6ms 261.99ms 4.18 11.97s 50
# 13 r 1024 4s 4.22s 4.4s 0.238 3.5m 50
# 14 cpp 1024 410.48ms 431.43ms 462.44ms 2.33 21.45s 50
ggplot(res_bench, aes(x = factor(batch_size), y = median,
group = expression, color = expression)) +
geom_point() +
geom_line() +
ylab("median time, s") +
theme_minimal() +
scale_color_discrete(name = "", labels = c("cpp", "r")) +
theme(legend.position = "bottom")
E pei ona mafai ona e vaʻaia, o le saoasaoa o le faʻatupulaia na matua taua tele, ma e le mafai ona maua le C ++ code e ala i le faʻatusatusaina o le R code.
3. Iterators mo le la'uina o vaega mai le database
R o loʻo i ai se igoa taʻutaʻua mo le faʻaogaina o faʻamaumauga e fetaui ma le RAM, aʻo le Python e sili atu ona faʻaalia i le faʻaogaina o faʻamatalaga faʻamatalaga, e mafai ai ona e faigofie ma faʻapitoa le faʻaogaina o faʻatusatusaga (faʻatatau e faʻaaoga ai mafaufauga i fafo). O se faʻataʻitaʻiga masani ma talafeagai mo i tatou i le tulaga o le faʻafitauli o loʻo faʻamatalaina o fesoʻotaʻiga neural loloto na aʻoaʻoina e le auala faʻasolosolo faʻasolosolo ma le faʻatusatusaina o le faʻasolosolo i laasaga taʻitasi e faʻaaoga ai se vaega itiiti o faʻamatalaga, poʻo le mini-batch.
O faʻavae aʻoaʻoga loloto o loʻo tusia i le Python ei ai vasega faʻapitoa e faʻatino ai le faʻataʻitaʻiga e faʻavae i luga o faʻamaumauga: laulau, ata i totonu o faila, faʻasologa binary, ma isi. I le R e mafai ona tatou faʻaogaina uma foliga o le faletusi Python faigata faʻatasi ai ma ona pito i tua eseese e faʻaaoga ai le afifi o le igoa lava e tasi, lea e galue i luga o le afifi toe faʻamatala. O le vaega mulimuli e tatau ona i ai se isi tusitusiga umi; e le gata ina faʻatagaina oe e taʻavale le code Python mai le R, ae faʻatagaina foi oe e faʻafeiloaʻi mea i le va o R ma Python sauniga, faʻapipiʻi otometi uma ituaiga suiga talafeagai.
Na matou faʻaumatia le manaʻoga e teu uma faʻamatalaga i le RAM e ala i le faʻaaogaina o le MonetDBite, o galuega uma "neural network" o le a faia e le uluai code i le Python, e tatau lava ona matou tusia se faʻamatalaga i luga o faʻamaumauga, talu ai e leai se mea ua saunia. mo sea tulaga i le R poʻo le Python. E na'o le lua lava mana'oga mo lea mea: e tatau ona toe fa'afo'i vaega i se matasele e le gata ma fa'asaoina lona tulaga i le va o fa'asologa (o le mea mulimuli i le R o lo'o fa'atinoina i se auala sili ona faigofie e fa'aoga tapuni). I le taimi muamua, sa manaʻomia le faʻaliliu manino o R arrays i numpy arrays i totonu o le tagata faʻataʻitaʻi, ae o le taimi nei o le afifi. faigata faia e ia lava.
O le suʻesuʻega mo aʻoaʻoga ma faʻamaumauga faʻamaonia na faʻaalia e faapea:
Iterator mo aʻoaʻoga ma faʻamaumauga faʻamaonia
train_generator <- function(db_connection = con,
samples_index,
num_classes = 340,
batch_size = 32,
scale = 1,
color = FALSE,
imagenet_preproc = FALSE) {
# Проверка аргументов
checkmate::assert_class(con, "DBIConnection")
checkmate::assert_integerish(samples_index)
checkmate::assert_count(num_classes)
checkmate::assert_count(batch_size)
checkmate::assert_number(scale, lower = 0.001, upper = 5)
checkmate::assert_flag(color)
checkmate::assert_flag(imagenet_preproc)
# Перемешиваем, чтобы брать и удалять использованные индексы батчей по порядку
dt <- data.table::data.table(id = sample(samples_index))
# Проставляем номера батчей
dt[, batch := (.I - 1L) %/% batch_size + 1L]
# Оставляем только полные батчи и индексируем
dt <- dt[, if (.N == batch_size) .SD, keyby = batch]
# Устанавливаем счётчик
i <- 1
# Количество батчей
max_i <- dt[, max(batch)]
# Подготовка выражения для выгрузки
sql <- sprintf(
"PREPARE SELECT drawing, label_int FROM doodles WHERE id IN (%s)",
paste(rep("?", batch_size), collapse = ",")
)
res <- DBI::dbSendQuery(con, sql)
# Аналог keras::to_categorical
to_categorical <- function(x, num) {
n <- length(x)
m <- numeric(n * num)
m[x * n + seq_len(n)] <- 1
dim(m) <- c(n, num)
return(m)
}
# Замыкание
function() {
# Начинаем новую эпоху
if (i > max_i) {
dt[, id := sample(id)]
data.table::setkey(dt, batch)
# Сбрасываем счётчик
i <<- 1
max_i <<- dt[, max(batch)]
}
# ID для выгрузки данных
batch_ind <- dt[batch == i, id]
# Выгрузка данных
batch <- DBI::dbFetch(DBI::dbBind(res, as.list(batch_ind)), n = -1)
# Увеличиваем счётчик
i <<- i + 1
# Парсинг JSON и подготовка массива
batch_x <- cpp_process_json_vector(batch$drawing, scale = scale, color = color)
if (imagenet_preproc) {
# Шкалирование c интервала [0, 1] на интервал [-1, 1]
batch_x <- (batch_x - 0.5) * 2
}
batch_y <- to_categorical(batch$label_int, num_classes)
result <- list(batch_x, batch_y)
return(result)
}
}O le galuega e ave e fai ma faʻaoga se fesuiaiga ma se fesoʻotaʻiga i le database, numera o laina faʻaaogaina, numera o vasega, lapoʻa lapoa, fua (scale = 1 e fetaui ma le fa'aliliuina o ata o le 256x256 pixels, scale = 0.5 — 128x128 pika), fa'ailoga lanu (color = FALSE fa'amaoti le fa'aliliuina i le fa'aefuefu pe'ā fa'aoga color = TRUE ta'i ta'itasi e tosoina i se lanu fou) ma se fa'ailoga muamua mo feso'ota'iga na mua'i a'oa'oina ile imagenet. O le mea mulimuli e manaʻomia ina ia fuaina ai tau pika mai le vaeluaga [0, 1] i le vaeluaga [-1, 1], lea na faʻaaogaina pe a aʻoaʻoina le tuʻuina atu. faigata faʻataʻitaʻiga.
O le galuega i fafo o loʻo i ai le siakiina o ituaiga finauga, se laulau data.table fa'atasi ai ma numera laina fa'afefiloi mai samples_index ma numera o faʻaputuga, faʻataʻitaʻiga ma numera maualuga o vaega, faʻapea foʻi ma se faʻamatalaga SQL mo le laʻuina o faʻamaumauga mai le faʻamaumauga. E le gata i lea, na matou faʻamatalaina se analogue vave o le galuega i totonu keras::to_categorical(). Na matou faʻaaogaina toetoe lava o faʻamatalaga uma mo aʻoaʻoga, ma tuʻu ai le afa pasene mo le faʻamaonia, o lea na faʻatapulaʻaina ai le tele o taimi e le parakalafa. steps_per_epoch pe a valaauina keras::fit_generator(), ma le tulaga if (i > max_i) na'o le galue mo le fa'amaonia le su'esu'e.
I totonu o le galuega i totonu, e toe maua mai laina fa'asino igoa mo le isi vaega, fa'amaumauga e la'u mai le fa'amaumauga ma le fa'aputuga fa'aopoopo fa'aopoopo, JSON parsing (galuega cpp_process_json_vector(), tusia i le C++) ma le fatuina o faʻasologa e fetaui ma ata. Ona faia lea o vete vevela tasi ma igoa o vasega, faʻapipiʻi faʻatasi ma tau pika ma faʻailoga e tuʻufaʻatasia i se lisi, o le tau toe foʻi mai. Ina ia faʻavavevave galuega, matou faʻaaogaina le fausiaina o faʻasino igoa i laulau data.table ma suiga e ala i le so'otaga - e aunoa ma nei afifi "chips" fa'amaumauga. laulau E faigata tele ona mafaufau i le galue lelei ma soʻo se aofaiga tele o faʻamaumauga i le R.
O faʻaiʻuga o fua faʻavave i luga o le komepiuta Core i5 e faʻapea:
Toerator benchmark
library(Rcpp)
library(keras)
library(ggplot2)
source("utils/rcpp.R")
source("utils/keras_iterator.R")
con <- DBI::dbConnect(drv = MonetDBLite::MonetDBLite(), Sys.getenv("DBDIR"))
ind <- seq_len(DBI::dbGetQuery(con, "SELECT count(*) FROM doodles")[[1L]])
num_classes <- DBI::dbGetQuery(con, "SELECT max(label_int) + 1 FROM doodles")[[1L]]
# Индексы для обучающей выборки
train_ind <- sample(ind, floor(length(ind) * 0.995))
# Индексы для проверочной выборки
val_ind <- ind[-train_ind]
rm(ind)
# Коэффициент масштаба
scale <- 0.5
# Проведение замера
res_bench <- bench::press(
batch_size = 2^(4:10),
{
it1 <- train_generator(
db_connection = con,
samples_index = train_ind,
num_classes = num_classes,
batch_size = batch_size,
scale = scale
)
bench::mark(
it1(),
min_iterations = 50L
)
}
)
# Параметры бенчмарка
cols <- c("batch_size", "min", "median", "max", "itr/sec", "total_time", "n_itr")
res_bench[, cols]
# batch_size min median max `itr/sec` total_time n_itr
# <dbl> <bch:tm> <bch:tm> <bch:tm> <dbl> <bch:tm> <int>
# 1 16 25ms 64.36ms 92.2ms 15.9 3.09s 49
# 2 32 48.4ms 118.13ms 197.24ms 8.17 5.88s 48
# 3 64 69.3ms 117.93ms 181.14ms 8.57 5.83s 50
# 4 128 157.2ms 240.74ms 503.87ms 3.85 12.71s 49
# 5 256 359.3ms 613.52ms 988.73ms 1.54 30.5s 47
# 6 512 884.7ms 1.53s 2.07s 0.674 1.11m 45
# 7 1024 2.7s 3.83s 5.47s 0.261 2.81m 44
ggplot(res_bench, aes(x = factor(batch_size), y = median, group = 1)) +
geom_point() +
geom_line() +
ylab("median time, s") +
theme_minimal()
DBI::dbDisconnect(con, shutdown = TRUE)
Afai e lava le aofaʻi o le RAM, e mafai ona e faʻavavevaveina le faʻaogaina o le database e ala i le tuʻuina atu i lenei lava RAM (32 GB e lava mo la matou galuega). I Linux, o le vaeluaga o loʻo faʻapipiʻiina e ala i le faaletonu /dev/shm, nofoia e oo atu i le afa o le gafatia o le RAM. E mafai ona e fa'ailoga atili e ala i le fa'asa'oina /etc/fstabia maua se faamaumauga e pei o tmpfs /dev/shm tmpfs defaults,size=25g 0 0. Ia mautinoa e toe faʻafou ma siaki le taunuʻuga e ala i le faʻatinoina o le poloaiga df -h.
O le suʻesuʻe mo faʻamatalaga suʻega e foliga sili atu ona faigofie, talu ai o faʻamaumauga o suʻega e fetaui lelei i le RAM:
Fa'amatalaga mo fa'amatalaga su'ega
test_generator <- function(dt,
batch_size = 32,
scale = 1,
color = FALSE,
imagenet_preproc = FALSE) {
# Проверка аргументов
checkmate::assert_data_table(dt)
checkmate::assert_count(batch_size)
checkmate::assert_number(scale, lower = 0.001, upper = 5)
checkmate::assert_flag(color)
checkmate::assert_flag(imagenet_preproc)
# Проставляем номера батчей
dt[, batch := (.I - 1L) %/% batch_size + 1L]
data.table::setkey(dt, batch)
i <- 1
max_i <- dt[, max(batch)]
# Замыкание
function() {
batch_x <- cpp_process_json_vector(dt[batch == i, drawing],
scale = scale, color = color)
if (imagenet_preproc) {
# Шкалирование c интервала [0, 1] на интервал [-1, 1]
batch_x <- (batch_x - 0.5) * 2
}
result <- list(batch_x)
i <<- i + 1
return(result)
}
}4. Filifiliga o faʻataʻitaʻiga faʻataʻitaʻiga
O le fausaga muamua na faʻaaogaina , o uiga ia o loʻo talanoaina i savali. O loʻo aofia ai e pei o tulaga masani faigata ma, e tusa ai, o loʻo maua i totonu o le afifi o le igoa lava e tasi mo R. Ae a taumafai e faʻaaogaina i ata e tasi-auala, o se mea uiga ese na aliali mai: o le tensor input e tatau lava ona i ai le fua. (batch, height, width, 3), o lona uiga, e le mafai ona suia le numera o alalaupapa. E leai se faʻatapulaʻaina i le Python, o lea na matou faanatinati ai ma tusia a matou lava faʻatinoga o lenei fausaga, mulimuli i le uluaʻi tusiga (e aunoa ma le pa'ū o loʻo i totonu o le keras version):
Mobilenet v1 fausaga
library(keras)
top_3_categorical_accuracy <- custom_metric(
name = "top_3_categorical_accuracy",
metric_fn = function(y_true, y_pred) {
metric_top_k_categorical_accuracy(y_true, y_pred, k = 3)
}
)
layer_sep_conv_bn <- function(object,
filters,
alpha = 1,
depth_multiplier = 1,
strides = c(2, 2)) {
# NB! depth_multiplier != resolution multiplier
# https://github.com/keras-team/keras/issues/10349
layer_depthwise_conv_2d(
object = object,
kernel_size = c(3, 3),
strides = strides,
padding = "same",
depth_multiplier = depth_multiplier
) %>%
layer_batch_normalization() %>%
layer_activation_relu() %>%
layer_conv_2d(
filters = filters * alpha,
kernel_size = c(1, 1),
strides = c(1, 1)
) %>%
layer_batch_normalization() %>%
layer_activation_relu()
}
get_mobilenet_v1 <- function(input_shape = c(224, 224, 1),
num_classes = 340,
alpha = 1,
depth_multiplier = 1,
optimizer = optimizer_adam(lr = 0.002),
loss = "categorical_crossentropy",
metrics = c("categorical_crossentropy",
top_3_categorical_accuracy)) {
inputs <- layer_input(shape = input_shape)
outputs <- inputs %>%
layer_conv_2d(filters = 32, kernel_size = c(3, 3), strides = c(2, 2), padding = "same") %>%
layer_batch_normalization() %>%
layer_activation_relu() %>%
layer_sep_conv_bn(filters = 64, strides = c(1, 1)) %>%
layer_sep_conv_bn(filters = 128, strides = c(2, 2)) %>%
layer_sep_conv_bn(filters = 128, strides = c(1, 1)) %>%
layer_sep_conv_bn(filters = 256, strides = c(2, 2)) %>%
layer_sep_conv_bn(filters = 256, strides = c(1, 1)) %>%
layer_sep_conv_bn(filters = 512, strides = c(2, 2)) %>%
layer_sep_conv_bn(filters = 512, strides = c(1, 1)) %>%
layer_sep_conv_bn(filters = 512, strides = c(1, 1)) %>%
layer_sep_conv_bn(filters = 512, strides = c(1, 1)) %>%
layer_sep_conv_bn(filters = 512, strides = c(1, 1)) %>%
layer_sep_conv_bn(filters = 512, strides = c(1, 1)) %>%
layer_sep_conv_bn(filters = 1024, strides = c(2, 2)) %>%
layer_sep_conv_bn(filters = 1024, strides = c(1, 1)) %>%
layer_global_average_pooling_2d() %>%
layer_dense(units = num_classes) %>%
layer_activation_softmax()
model <- keras_model(
inputs = inputs,
outputs = outputs
)
model %>% compile(
optimizer = optimizer,
loss = loss,
metrics = metrics
)
return(model)
}O le le lelei o lenei auala e iloagofie. Ou te manaʻo e faʻataʻitaʻi le tele o faʻataʻitaʻiga, ae i se isi itu, ou te le manaʻo e toe tusi tusi taʻitasi ma le lima. Na le maua foi le avanoa e faʻaaoga ai le mamafa o faʻataʻitaʻiga na aʻoaʻoina muamua ile imagenet. E pei ona masani ai, o le suʻesuʻeina o faʻamaumauga na fesoasoani. Galuega get_config() fa'atagaina oe e maua se fa'amatalaga o le fa'ata'ita'iga i se fomu e talafeagai mo le fa'asa'oina (base_model_conf$layers - se lisi masani R), ma le galuega from_config() fa'atino le suiga i tua i se mea fa'ata'ita'i:
base_model_conf <- get_config(base_model)
base_model_conf$layers[[1]]$config$batch_input_shape[[4]] <- 1L
base_model <- from_config(base_model_conf)Ole taimi nei e le faigata ona tusia se galuega lautele e maua ai soʻo se mea e tuʻuina atu faigata fa'ata'ita'iga o lo'o i ai pe leai ni mamafa ua a'oa'oina ile imagenet:
Galuega mo le utaina o fausaga ua saunia
get_model <- function(name = "mobilenet_v2",
input_shape = NULL,
weights = "imagenet",
pooling = "avg",
num_classes = NULL,
optimizer = keras::optimizer_adam(lr = 0.002),
loss = "categorical_crossentropy",
metrics = NULL,
color = TRUE,
compile = FALSE) {
# Проверка аргументов
checkmate::assert_string(name)
checkmate::assert_integerish(input_shape, lower = 1, upper = 256, len = 3)
checkmate::assert_count(num_classes)
checkmate::assert_flag(color)
checkmate::assert_flag(compile)
# Получаем объект из пакета keras
model_fun <- get0(paste0("application_", name), envir = asNamespace("keras"))
# Проверка наличия объекта в пакете
if (is.null(model_fun)) {
stop("Model ", shQuote(name), " not found.", call. = FALSE)
}
base_model <- model_fun(
input_shape = input_shape,
include_top = FALSE,
weights = weights,
pooling = pooling
)
# Если изображение не цветное, меняем размерность входа
if (!color) {
base_model_conf <- keras::get_config(base_model)
base_model_conf$layers[[1]]$config$batch_input_shape[[4]] <- 1L
base_model <- keras::from_config(base_model_conf)
}
predictions <- keras::get_layer(base_model, "global_average_pooling2d_1")$output
predictions <- keras::layer_dense(predictions, units = num_classes, activation = "softmax")
model <- keras::keras_model(
inputs = base_model$input,
outputs = predictions
)
if (compile) {
keras::compile(
object = model,
optimizer = optimizer,
loss = loss,
metrics = metrics
)
}
return(model)
}A fa'aogaina ata ala ta'itasi, e le'o fa'aaogaina ni fua fa'aa'oa'oina. E mafai ona faʻaleleia lenei mea: faʻaaoga le galuega get_weights() maua le mamafa faʻataʻitaʻiga i le tulaga o se lisi o R arrays, sui le fua o le elemene muamua o lenei lisi (e ala i le ave o le tasi lanu lanu poʻo le averesi o mea uma e tolu), ona toe utaina lea o le mamafa i totonu o le faʻataʻitaʻiga ma le galuega. set_weights(). Matou te leʻi faʻaopoopoina lenei faʻatinoga, aua i lenei laʻasaga ua uma ona manino e sili atu ona aoga le galue ma ata lanu.
Na matou faia le tele o faʻataʻitaʻiga e faʻaaoga ai le mobilenet versions 1 ma le 2, faʻapea foʻi ma le resnet34. O fale fa'aonaponei fa'aonaponei e pei o le SE-ResNeXt na fa'atino lelei i lenei tauvaga. Ae paga lea, e leʻi iai ni a matou faʻatinoga ua saunia, ma matou te leʻi tusia a matou lava mea (ae matou te mautinoa lava e tusi).
5. Parameterization o tusitusiga
Mo le faʻaogaina, o tulafono uma mo le amataina o aʻoaʻoga na mamanuina e avea o se tusitusiga e tasi, faʻaogaina le faʻaogaina e pei ona taʻua i lalo:
doc <- '
Usage:
train_nn.R --help
train_nn.R --list-models
train_nn.R [options]
Options:
-h --help Show this message.
-l --list-models List available models.
-m --model=<model> Neural network model name [default: mobilenet_v2].
-b --batch-size=<size> Batch size [default: 32].
-s --scale-factor=<ratio> Scale factor [default: 0.5].
-c --color Use color lines [default: FALSE].
-d --db-dir=<path> Path to database directory [default: Sys.getenv("db_dir")].
-r --validate-ratio=<ratio> Validate sample ratio [default: 0.995].
-n --n-gpu=<number> Number of GPUs [default: 1].
'
args <- docopt::docopt(doc)Faʻapipiʻi tusipasi e fai ma sui o le faatinoga mo R. Faatasi ai ma lana fesoasoani, o loʻo faʻalauiloaina faʻamaumauga i ni faʻatonuga faigofie e pei o Rscript bin/train_nn.R -m resnet50 -c -d /home/andrey/doodle_db poʻo ./bin/train_nn.R -m resnet50 -c -d /home/andrey/doodle_db, pe afai faila train_nn.R e mafai ona faʻatinoina (o lenei poloaiga o le a amata aʻoaʻoina le faʻataʻitaʻiga resnet50 i luga o ata e tolu-lanu e fua 128x128 pixels, e tatau ona tuʻu le faʻamaumauga i totonu o le pusa /home/andrey/doodle_db). E mafai ona e fa'aopoopoina le saoasaoa o le a'oa'oina, optimizer type, ma so'o se isi fa'ailoga fa'apitoa i le lisi. I le faagasologa o le saunia o le lomiga, na foliga mai o le tusiata fale mobilenet_v2 mai le lomiga o iai nei faigata i le R faʻaaogaina ona o suiga e le o amanaia i le R package, matou te faʻatali mo latou e toe faʻaleleia.
O lenei faiga na mafai ai ona faʻavavevave faʻataʻitaʻiga ma faʻataʻitaʻiga eseese pe a faʻatusatusa i le faʻalauiloaina masani o tusitusiga ile RStudio (matou te matauina le afifi o se isi mea e mafai. ). Ae o le aoga sili o le mafai lea ona faigofie ona faʻatautaia le faʻalauiloaina o tusitusiga i Docker pe naʻo luga ole server, e aunoa ma le faʻapipiʻiina o le RStudio mo lenei mea.
6. Fa'amauina o tusitusiga
Na matou faʻaaogaina Docker e faʻamautinoa ai le feaveaʻi o le siosiomaga mo faʻataʻitaʻiga aʻoaʻoga i le va o tagata o le 'au ma mo le faʻapipiʻiina vave i le ao. E mafai ona e amata faamasani i lenei meafaigaluega, lea e le masani ai mo se R programmer, ma faasologa o lomiga po o .
Docker e faʻatagaina oe e faia uma au lava ata mai le sasa ma faʻaoga isi ata e fai ma faʻavae mo le fatuina o oe lava. Pe a suʻesuʻeina avanoa avanoa, na matou oʻo mai i le faaiuga o le faʻapipiʻiina o le NVIDIA, CUDA + cuDNN avetaʻavale ma faletusi Python o se vaega tele o le ata, ma na matou filifili e ave le ata aloaia e fai ma faavae. tensorflow/tensorflow:1.12.0-gpu, faʻaopoopo i ai mea manaʻomia R packages.
O le faila faila mulimuli e pei o lenei:
faila faila
FROM tensorflow/tensorflow:1.12.0-gpu
MAINTAINER Artem Klevtsov <[email protected]>
SHELL ["/bin/bash", "-c"]
ARG LOCALE="en_US.UTF-8"
ARG APT_PKG="libopencv-dev r-base r-base-dev littler"
ARG R_BIN_PKG="futile.logger checkmate data.table rcpp rapidjsonr dbi keras jsonlite curl digest remotes"
ARG R_SRC_PKG="xtensor RcppThread docopt MonetDBLite"
ARG PY_PIP_PKG="keras"
ARG DIRS="/db /app /app/data /app/models /app/logs"
RUN source /etc/os-release &&
echo "deb https://cloud.r-project.org/bin/linux/ubuntu ${UBUNTU_CODENAME}-cran35/" > /etc/apt/sources.list.d/cran35.list &&
apt-key adv --keyserver keyserver.ubuntu.com --recv-keys E084DAB9 &&
add-apt-repository -y ppa:marutter/c2d4u3.5 &&
add-apt-repository -y ppa:timsc/opencv-3.4 &&
apt-get update &&
apt-get install -y locales &&
locale-gen ${LOCALE} &&
apt-get install -y --no-install-recommends ${APT_PKG} &&
ln -s /usr/lib/R/site-library/littler/examples/install.r /usr/local/bin/install.r &&
ln -s /usr/lib/R/site-library/littler/examples/install2.r /usr/local/bin/install2.r &&
ln -s /usr/lib/R/site-library/littler/examples/installGithub.r /usr/local/bin/installGithub.r &&
echo 'options(Ncpus = parallel::detectCores())' >> /etc/R/Rprofile.site &&
echo 'options(repos = c(CRAN = "https://cloud.r-project.org"))' >> /etc/R/Rprofile.site &&
apt-get install -y $(printf "r-cran-%s " ${R_BIN_PKG}) &&
install.r ${R_SRC_PKG} &&
pip install ${PY_PIP_PKG} &&
mkdir -p ${DIRS} &&
chmod 777 ${DIRS} &&
rm -rf /tmp/downloaded_packages/ /tmp/*.rds &&
rm -rf /var/lib/apt/lists/*
COPY utils /app/utils
COPY src /app/src
COPY tests /app/tests
COPY bin/*.R /app/
ENV DBDIR="/db"
ENV CUDA_HOME="/usr/local/cuda"
ENV PATH="/app:${PATH}"
WORKDIR /app
VOLUME /db
VOLUME /app
CMD bash
Mo le faigofie, o afifi na faʻaaogaina na tuʻuina i ni fesuiaiga; o le tele o tusitusiga tusitusi e kopiina i totonu o pusa i le taimi o le faʻapotopotoga. Na matou suia foi le atigi poloaiga i /bin/bash mo le faigofie o le faʻaogaina o mea /etc/os-release. O lenei mea na aloese ai mai le manaʻoga e faʻamaonia le OS version i le code.
E le gata i lea, o se tamai tusi bash na tusia e mafai ai ona e faʻalauiloa se atigipusa ma ni tulafono eseese. Mo se faʻataʻitaʻiga, e mafai ona avea nei ma tusitusiga mo le aʻoaʻoina o fesoʻotaʻiga neural na tuʻuina muamua i totonu o le koneteina, poʻo se atigi faʻatonuga mo le faʻapipiʻiina ma le mataʻituina o le gaioiga o le pusa:
Fa'amatalaga e tatala ai le koneteina
#!/bin/sh
DBDIR=${PWD}/db
LOGSDIR=${PWD}/logs
MODELDIR=${PWD}/models
DATADIR=${PWD}/data
ARGS="--runtime=nvidia --rm -v ${DBDIR}:/db -v ${LOGSDIR}:/app/logs -v ${MODELDIR}:/app/models -v ${DATADIR}:/app/data"
if [ -z "$1" ]; then
CMD="Rscript /app/train_nn.R"
elif [ "$1" = "bash" ]; then
ARGS="${ARGS} -ti"
else
CMD="Rscript /app/train_nn.R $@"
fi
docker run ${ARGS} doodles-tf ${CMD}Afai o lenei bash script e faʻatautaia e aunoa ma ni faʻamaufaʻailoga, o le a valaʻau le tusitusiga i totonu o le pusa train_nn.R fa'atasi ai ma tau fa'aletonu; pe afai o le finauga muamua o le "bash", ona amata lea o le koneteina fegalegaleai ma se atigi poloaiga. I isi tulaga uma, o le tau o finauga tulaga e suitulaga: CMD="Rscript /app/train_nn.R $@".
E taua le maitauina o faʻamaumauga ma faʻamaumauga faʻamaumauga ma faʻamaumauga, faʻapea foʻi ma le lisi mo le faʻasaoina o faʻataʻitaʻiga aʻoaʻoina, o loʻo faʻapipiʻiina i totonu o le koneteina mai le polokalama talimalo, lea e mafai ai ona e mauaina iʻuga o tusitusiga e aunoa ma ni togafiti le manaʻomia.
7. Fa'aaogaina o le tele o GPU ile Google Cloud
O se tasi o vaega o le tauvaga o le pisapisao tele o faʻamatalaga (vaai i le ulutala ata, nono mai @Leigh.plt mai le ODS slack). O vaega tetele e fesoasoani e faʻafefe ai lenei mea, ma ina ua maeʻa faʻataʻitaʻiga i luga o se PC ma le 1 GPU, na matou filifili e faʻatautaia faʻataʻitaʻiga aʻoaʻoga i luga o le tele o GPU i le ao. Fa'aaoga GoogleCloud () ona o le tele o filifiliga o fetuutuunaiga avanoa, tau talafeagai ma ponesi $300. Ona o le matapeʻapeʻa, na ou faʻatonuina se faʻataʻitaʻiga 4xV100 ma se SSD ma se tone o le RAM, ma o se mea sese tele. O sea masini e 'ai vave tupe; e mafai ona e alu fa'ata'ita'i e aunoa ma se paipa fa'amaonia. Mo faamoemoega faaleaoaoga, e sili atu le ave o le K80. Ae o le tele o le RAM na faʻaaogaina - o le ao SSD e leʻi faʻaalia i lana faʻatinoga, o lea na faʻafeiloaʻi ai le database i. dev/shm.
O le mea e sili ona fiafia i ai o le vaega o le code e nafa ma le faʻaaogaina o le tele o GPU. Muamua, o le faʻataʻitaʻiga e faia i luga o le PPU e faʻaaoga ai se pule o tala, pei o le Python:
with(tensorflow::tf$device("/cpu:0"), {
model_cpu <- get_model(
name = model_name,
input_shape = input_shape,
weights = weights,
metrics =(top_3_categorical_accuracy,
compile = FALSE
)
})Ona kopiina lea o le faʻataʻitaʻiga e leʻi tuʻufaʻatasia (e taua lenei) i se numera tuʻuina atu o GPU avanoa, ma naʻo le maeʻa ona tuʻufaʻatasia:
model <- keras::multi_gpu_model(model_cpu, gpus = n_gpu)
keras::compile(
object = model,
optimizer = keras::optimizer_adam(lr = 0.0004),
loss = "categorical_crossentropy",
metrics = c(top_3_categorical_accuracy)
)O le auala masani o le faʻamaisaina o laulau uma sei vagana ai le mea mulimuli, aʻoaʻoina le vaega mulimuli, faʻamalo ma toe aʻoaʻoina le faʻataʻitaʻiga atoa mo le tele o GPU e leʻi mafai ona faʻatinoina.
Sa mataituina aoaoga e aunoa ma le faaaogaina. laupapa tensorboard, fa'atapula'aina i matou i le fa'amauina o ogalaau ma fa'asaoina fa'ata'ita'iga ma igoa fa'amatalaga pe a uma vaitau ta'itasi:
Toe valaau
# Шаблон имени файла лога
log_file_tmpl <- file.path("logs", sprintf(
"%s_%d_%dch_%s.csv",
model_name,
dim_size,
channels,
format(Sys.time(), "%Y%m%d%H%M%OS")
))
# Шаблон имени файла модели
model_file_tmpl <- file.path("models", sprintf(
"%s_%d_%dch_{epoch:02d}_{val_loss:.2f}.h5",
model_name,
dim_size,
channels
))
callbacks_list <- list(
keras::callback_csv_logger(
filename = log_file_tmpl
),
keras::callback_early_stopping(
monitor = "val_loss",
min_delta = 1e-4,
patience = 8,
verbose = 1,
mode = "min"
),
keras::callback_reduce_lr_on_plateau(
monitor = "val_loss",
factor = 0.5, # уменьшаем lr в 2 раза
patience = 4,
verbose = 1,
min_delta = 1e-4,
mode = "min"
),
keras::callback_model_checkpoint(
filepath = model_file_tmpl,
monitor = "val_loss",
save_best_only = FALSE,
save_weights_only = FALSE,
mode = "min"
)
)8. Nai lo o se faaiuga
O le tele o faʻafitauli na matou feagai e leʻi foia:
- в faigata e leai se galuega ua saunia mo le otometi ona suʻeina le maualuga o le aʻoaʻoga (analogue
lr_finderi le faletusi vave.ai); Faatasi ai ma ni taumafaiga, e mafai ona tuʻuina atu faʻatinoga lona tolu i le R, mo se faʻataʻitaʻiga, ; - ona o se taunuuga o le vaega muamua, e le mafai ona filifili le sao sao aʻoaʻoga saoasaoa pe a faʻaaogaina le tele o GPU;
- o loʻo i ai le leai o ni fausaga faʻaonaponei neural network, aemaise lava i latou na muaʻi aʻoaʻoina ile imagenet;
- e leai se faiga fa'ata'amilosaga e tasi ma fua fa'atatau o a'oa'oga (kosine annealing sa i la matou talosaga , Faafetai ).
O a mea aoga na aʻoaʻoina mai lenei tauvaga:
- I luga o meafaigaluega maualalo-malosi, e mafai ona e galue ma le lelei (tele taimi le tele o le RAM) voluma o faʻamaumauga e aunoa ma se tiga. taga palasitika fa'amaumauga. laulau fa'asaoina le manatua ona o suiga i totonu o laulau, e 'alofia ai le kopiina, ma a fa'aoga sa'o, e toetoe lava a fa'aalia i taimi uma ona gafatia le saoasaoa maualuga i meafaigaluega uma ua tatou iloa mo gagana tusitusi. Faʻasaoina faʻamaumauga i totonu o se faʻamaumauga e mafai ai e oe, i le tele o tulaga, aua le mafaufau i le manaʻoga e faʻapipiʻi le faʻamaumauga atoa i le RAM.
- O galuega fa'agesegese ile R e mafai ona sui i mea vave ile C ++ fa'aoga le afifi Rcpp. Afai e faaopoopo i le faaaogaina RcppThread poʻo RcppParallel, matou te maua faʻasalalauga faʻapipiʻi tele-filo, o lea e leai se manaʻoga e faʻatusatusa le code i le R level.
- afifi Rcpp e mafai ona faʻaaogaina e aunoa ma le malamalama tele i le C ++, o loʻo faʻamatalaina le pito maualalo manaʻomia . Fa'aulu faila mo le tele o faletusi manaia C pei xtensor o loʻo avanoa ile CRAN, o lona uiga, o loʻo faʻatulagaina se atinaʻe mo le faʻatinoina o galuega faatino e tuʻufaʻatasia ai le C ++ code maualuga i le R. Fa'aopoopo le fa'aogaina o le fa'ailoga fa'asologa ma se su'esu'ega fa'ailoga C++ i le RStudio.
- tusipasi e fa'atagaina oe e fa'atautaia fa'amaumauga a le tagata lava ia ma fa'amaufa'ailoga. E faigofie lenei mea mo le faʻaogaina i luga o se server mamao, e aofia ai. i lalo o le fagafao. I RStudio, e le faigofie le faia o le tele o itula o faʻataʻitaʻiga ma aʻoaʻoga neural networks, ma le faʻapipiʻiina o le IDE i luga o le 'auʻaunaga lava ia e le faʻamaonia i taimi uma.
- E faʻamautinoa e Docker le faʻaogaina o le code ma le toe faʻaleleia o taunuʻuga i le va o tagata atiaʻe ma ituaiga eseese o le OS ma faletusi, faʻapea foʻi ma le faigofie o le faʻatinoina i luga o sapalai. E mafai ona e fa'alauiloa le paipa a'oa'oga atoa ile na'o le tasi le fa'atonuga.
- Google Cloud ose auala fa'atatau tupe e fa'ata'ita'i ai i masini taugata, ae e mana'omia lou filifilia ma le fa'aeteete.
- O le fuaina o le saoasaoa o vaega taʻitasi code e aoga tele, aemaise lava pe a tuʻufaʻatasia R ma C ++, ma faʻatasi ai ma le afifi. nofoa - faigofie tele foi.
I le aotelega o lenei aafiaga sa matua tauia ma o loo faaauau pea ona matou galulue e foia nisi o mataupu na laga.
puna: www.habr.com