{ "nbformat": 4, "nbformat_minor": 2, "metadata": { "colab": { "name": "lesson_11-R.ipynb", "provenance": [], "collapsed_sections": [], "toc_visible": true }, "kernelspec": { "name": "ir", "display_name": "R" }, "language_info": { "name": "R" }, "coopTranslator": { "original_hash": "6ea6a5171b1b99b7b5a55f7469c048d2", "translation_date": "2025-10-11T12:29:31+00:00", "source_file": "4-Classification/2-Classifiers-1/solution/R/lesson_11-R.ipynb", "language_code": "et" } }, "cells": [ { "cell_type": "markdown", "source": [ "# Koosta klassifikatsioonimudel: Maitsvad Aasia ja India köögid\n" ], "metadata": { "id": "zs2woWv_HoE8" } }, { "cell_type": "markdown", "source": [ "## Köögi klassifikaatorid 1\n", "\n", "Selles tunnis uurime erinevaid klassifikaatoreid, et *ennustada rahvuskööki antud koostisosade grupi põhjal.* Samal ajal õpime rohkem, kuidas algoritme saab kasutada klassifitseerimisülesannete jaoks.\n", "\n", "### [**Eeltesti küsimustik**](https://gray-sand-07a10f403.1.azurestaticapps.net/quiz/21/)\n", "\n", "### **Ettevalmistus**\n", "\n", "See tund põhineb meie [eelmisel tunnil](https://github.com/microsoft/ML-For-Beginners/blob/main/4-Classification/1-Introduction/solution/lesson_10-R.ipynb), kus:\n", "\n", "- Tegime kerge sissejuhatuse klassifikatsioonidesse, kasutades andmestikku Aasia ja India köökide kohta 😋.\n", "\n", "- Uurisime mõningaid [dplyr verbe](https://dplyr.tidyverse.org/), et ette valmistada ja puhastada oma andmeid.\n", "\n", "- Lõime kauneid visualiseeringuid, kasutades ggplot2.\n", "\n", "- Näitasime, kuidas tasakaalustamata andmetega toime tulla, neid eelprotsessides [recipes](https://recipes.tidymodels.org/articles/Simple_Example.html) abil.\n", "\n", "- Näitasime, kuidas `prep` ja `bake` meie retsepti, et kinnitada, et see töötab ootuspäraselt.\n", "\n", "#### **Eeltingimus**\n", "\n", "Selle tunni jaoks vajame järgmisi pakette, et puhastada, ette valmistada ja visualiseerida oma andmeid:\n", "\n", "- `tidyverse`: [tidyverse](https://www.tidyverse.org/) on [R-pakettide kogum](https://www.tidyverse.org/packages), mis on loodud selleks, et muuta andmeteadus kiiremaks, lihtsamaks ja lõbusamaks!\n", "\n", "- `tidymodels`: [tidymodels](https://www.tidymodels.org/) raamistik on [pakettide kogum](https://www.tidymodels.org/packages/) modelleerimiseks ja masinõppeks.\n", "\n", "- `themis`: [themis pakett](https://themis.tidymodels.org/) pakub lisaretsepti samme tasakaalustamata andmetega tegelemiseks.\n", "\n", "- `nnet`: [nnet pakett](https://cran.r-project.org/web/packages/nnet/nnet.pdf) pakub funktsioone ühe peidetud kihiga ettepoole suunatud närvivõrkude ja multinomiaalse logistilise regressiooni mudelite hindamiseks.\n", "\n", "Saate need installida järgmiselt:\n" ], "metadata": { "id": "iDFOb3ebHwQC" } }, { "cell_type": "markdown", "source": [ "`install.packages(c(\"tidyverse\", \"tidymodels\", \"DataExplorer\", \"here\"))`\n", "\n", "Teise võimalusena kontrollib allolev skript, kas teil on selle mooduli lõpetamiseks vajalikud paketid, ja installib need, kui need puuduvad.\n" ], "metadata": { "id": "4V85BGCjII7F" } }, { "cell_type": "code", "execution_count": 2, "source": [ "suppressWarnings(if (!require(\"pacman\"))install.packages(\"pacman\"))\r\n", "\r\n", "pacman::p_load(tidyverse, tidymodels, themis, here)" ], "outputs": [ { "output_type": "stream", "name": "stderr", "text": [ "Loading required package: pacman\n", "\n" ] } ], "metadata": { "colab": { "base_uri": "https://localhost:8080/" }, "id": "an5NPyyKIKNR", "outputId": "834d5e74-f4b8-49f9-8ab5-4c52ff2d7bc8" } }, { "cell_type": "markdown", "source": [ "Nüüd alustame hoogsalt!\n", "\n", "## 1. Jaga andmed treening- ja testkomplektideks.\n", "\n", "Alustame mõne sammuga meie eelmisest tunnist.\n", "\n", "### Eemalda kõige tavalisemad koostisosad, mis tekitavad segadust erinevate köökide vahel, kasutades `dplyr::select()`.\n", "\n", "Kõigile meeldib riis, küüslauk ja ingver!\n" ], "metadata": { "id": "0ax9GQLBINVv" } }, { "cell_type": "code", "execution_count": 3, "source": [ "# Load the original cuisines data\r\n", "df <- read_csv(file = \"https://raw.githubusercontent.com/microsoft/ML-For-Beginners/main/4-Classification/data/cuisines.csv\")\r\n", "\r\n", "# Drop id column, rice, garlic and ginger from our original data set\r\n", "df_select <- df %>% \r\n", " select(-c(1, rice, garlic, ginger)) %>%\r\n", " # Encode cuisine column as categorical\r\n", " mutate(cuisine = factor(cuisine))\r\n", "\r\n", "# Display new data set\r\n", "df_select %>% \r\n", " slice_head(n = 5)\r\n", "\r\n", "# Display distribution of cuisines\r\n", "df_select %>% \r\n", " count(cuisine) %>% \r\n", " arrange(desc(n))" ], "outputs": [ { "output_type": "stream", "name": "stderr", "text": [ "New names:\n", "* `` -> ...1\n", "\n", "\u001b[1m\u001b[1mRows: \u001b[1m\u001b[22m\u001b[34m\u001b[34m2448\u001b[34m\u001b[39m \u001b[1m\u001b[1mColumns: \u001b[1m\u001b[22m\u001b[34m\u001b[34m385\u001b[34m\u001b[39m\n", "\n", "\u001b[36m──\u001b[39m \u001b[1m\u001b[1mColumn specification\u001b[1m\u001b[22m \u001b[36m────────────────────────────────────────────────────────\u001b[39m\n", "\u001b[1mDelimiter:\u001b[22m \",\"\n", "\u001b[31mchr\u001b[39m (1): cuisine\n", "\u001b[32mdbl\u001b[39m (384): ...1, almond, angelica, anise, anise_seed, apple, apple_brandy, a...\n", "\n", "\n", "\u001b[36mℹ\u001b[39m Use \u001b[30m\u001b[47m\u001b[30m\u001b[47m`spec()`\u001b[47m\u001b[30m\u001b[49m\u001b[39m to retrieve the full column specification for this data.\n", "\u001b[36mℹ\u001b[39m Specify the column types or set \u001b[30m\u001b[47m\u001b[30m\u001b[47m`show_col_types = FALSE`\u001b[47m\u001b[30m\u001b[49m\u001b[39m to quiet this message.\n", "\n" ] }, { "output_type": "display_data", "data": { "text/plain": [ " cuisine almond angelica anise anise_seed apple apple_brandy apricot armagnac\n", "1 indian 0 0 0 0 0 0 0 0 \n", "2 indian 1 0 0 0 0 0 0 0 \n", "3 indian 0 0 0 0 0 0 0 0 \n", "4 indian 0 0 0 0 0 0 0 0 \n", "5 indian 0 0 0 0 0 0 0 0 \n", " artemisia ⋯ whiskey white_bread white_wine whole_grain_wheat_flour wine wood\n", "1 0 ⋯ 0 0 0 0 0 0 \n", "2 0 ⋯ 0 0 0 0 0 0 \n", "3 0 ⋯ 0 0 0 0 0 0 \n", "4 0 ⋯ 0 0 0 0 0 0 \n", "5 0 ⋯ 0 0 0 0 0 0 \n", " yam yeast yogurt zucchini\n", "1 0 0 0 0 \n", "2 0 0 0 0 \n", "3 0 0 0 0 \n", "4 0 0 0 0 \n", "5 0 0 1 0 " ], "text/markdown": [ "\n", "A tibble: 5 × 381\n", "\n", "| cuisine <fct> | almond <dbl> | angelica <dbl> | anise <dbl> | anise_seed <dbl> | apple <dbl> | apple_brandy <dbl> | apricot <dbl> | armagnac <dbl> | artemisia <dbl> | ⋯ ⋯ | whiskey <dbl> | white_bread <dbl> | white_wine <dbl> | whole_grain_wheat_flour <dbl> | wine <dbl> | wood <dbl> | yam <dbl> | yeast <dbl> | yogurt <dbl> | zucchini <dbl> |\n", "|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|\n", "| indian | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | ⋯ | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 0 |\n", "| indian | 1 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | ⋯ | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 0 |\n", "| indian | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | ⋯ | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 0 |\n", "| indian | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | ⋯ | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 0 |\n", "| indian | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | ⋯ | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 1 | 0 |\n", "\n" ], "text/latex": [ "A tibble: 5 × 381\n", "\\begin{tabular}{lllllllllllllllllllll}\n", " cuisine & almond & angelica & anise & anise\\_seed & apple & apple\\_brandy & apricot & armagnac & artemisia & ⋯ & whiskey & white\\_bread & white\\_wine & whole\\_grain\\_wheat\\_flour & wine & wood & yam & yeast & yogurt & zucchini\\\\\n", " & & & & & & & & & & ⋯ & & & & & & & & & & \\\\\n", "\\hline\n", "\t indian & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & ⋯ & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0\\\\\n", "\t indian & 1 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & ⋯ & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0\\\\\n", "\t indian & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & ⋯ & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0\\\\\n", "\t indian & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & ⋯ & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0\\\\\n", "\t indian & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & ⋯ & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 1 & 0\\\\\n", "\\end{tabular}\n" ], "text/html": [ "\n", "\n", "\n", "\t\n", "\t\n", "\n", "\n", "\t\n", "\t\n", "\t\n", "\t\n", "\t\n", "\n", "
A tibble: 5 × 381
cuisinealmondangelicaaniseanise_seedappleapple_brandyapricotarmagnacartemisia⋯whiskeywhite_breadwhite_winewhole_grain_wheat_flourwinewoodyamyeastyogurtzucchini
<fct><dbl><dbl><dbl><dbl><dbl><dbl><dbl><dbl><dbl>⋯<dbl><dbl><dbl><dbl><dbl><dbl><dbl><dbl><dbl><dbl>
indian000000000⋯0000000000
indian100000000⋯0000000000
indian000000000⋯0000000000
indian000000000⋯0000000000
indian000000000⋯0000000010
\n" ] }, "metadata": {} }, { "output_type": "display_data", "data": { "text/plain": [ " cuisine n \n", "1 korean 799\n", "2 indian 598\n", "3 chinese 442\n", "4 japanese 320\n", "5 thai 289" ], "text/markdown": [ "\n", "A tibble: 5 × 2\n", "\n", "| cuisine <fct> | n <int> |\n", "|---|---|\n", "| korean | 799 |\n", "| indian | 598 |\n", "| chinese | 442 |\n", "| japanese | 320 |\n", "| thai | 289 |\n", "\n" ], "text/latex": [ "A tibble: 5 × 2\n", "\\begin{tabular}{ll}\n", " cuisine & n\\\\\n", " & \\\\\n", "\\hline\n", "\t korean & 799\\\\\n", "\t indian & 598\\\\\n", "\t chinese & 442\\\\\n", "\t japanese & 320\\\\\n", "\t thai & 289\\\\\n", "\\end{tabular}\n" ], "text/html": [ "\n", "\n", "\n", "\t\n", "\t\n", "\n", "\n", "\t\n", "\t\n", "\t\n", "\t\n", "\t\n", "\n", "
A tibble: 5 × 2
cuisinen
<fct><int>
korean 799
indian 598
chinese 442
japanese320
thai 289
\n" ] }, "metadata": {} } ], "metadata": { "colab": { "base_uri": "https://localhost:8080/", "height": 735 }, "id": "jhCrrH22IWVR", "outputId": "d444a85c-1d8b-485f-bc4f-8be2e8f8217c" } }, { "cell_type": "markdown", "source": [ "Suurepärane! Nüüd jagame andmed nii, et 70% andmetest läheb treeninguks ja 30% testimiseks. Samuti rakendame `stratifitseerimise` tehnikat, et `säilitada iga köögi proportsioon` treening- ja valideerimisandmestikes.\n", "\n", "[rsample](https://rsample.tidymodels.org/), Tidymodelsi pakett, pakub infrastruktuuri tõhusaks andmete jagamiseks ja ümberproovimiseks:\n" ], "metadata": { "id": "AYTjVyajIdny" } }, { "cell_type": "code", "execution_count": 4, "source": [ "# Load the core Tidymodels packages into R session\r\n", "library(tidymodels)\r\n", "\r\n", "# Create split specification\r\n", "set.seed(2056)\r\n", "cuisines_split <- initial_split(data = df_select,\r\n", " strata = cuisine,\r\n", " prop = 0.7)\r\n", "\r\n", "# Extract the data in each split\r\n", "cuisines_train <- training(cuisines_split)\r\n", "cuisines_test <- testing(cuisines_split)\r\n", "\r\n", "# Print the number of cases in each split\r\n", "cat(\"Training cases: \", nrow(cuisines_train), \"\\n\",\r\n", " \"Test cases: \", nrow(cuisines_test), sep = \"\")\r\n", "\r\n", "# Display the first few rows of the training set\r\n", "cuisines_train %>% \r\n", " slice_head(n = 5)\r\n", "\r\n", "\r\n", "# Display distribution of cuisines in the training set\r\n", "cuisines_train %>% \r\n", " count(cuisine) %>% \r\n", " arrange(desc(n))" ], "outputs": [ { "output_type": "stream", "name": "stdout", "text": [ "Training cases: 1712\n", "Test cases: 736" ] }, { "output_type": "display_data", "data": { "text/plain": [ " cuisine almond angelica anise anise_seed apple apple_brandy apricot armagnac\n", "1 chinese 0 0 0 0 0 0 0 0 \n", "2 chinese 0 0 0 0 0 0 0 0 \n", "3 chinese 0 0 0 0 0 0 0 0 \n", "4 chinese 0 0 0 0 0 0 0 0 \n", "5 chinese 0 0 0 0 0 0 0 0 \n", " artemisia ⋯ whiskey white_bread white_wine whole_grain_wheat_flour wine wood\n", "1 0 ⋯ 0 0 0 0 1 0 \n", "2 0 ⋯ 0 0 0 0 1 0 \n", "3 0 ⋯ 0 0 0 0 0 0 \n", "4 0 ⋯ 0 0 0 0 0 0 \n", "5 0 ⋯ 0 0 0 0 0 0 \n", " yam yeast yogurt zucchini\n", "1 0 0 0 0 \n", "2 0 0 0 0 \n", "3 0 0 0 0 \n", "4 0 0 0 0 \n", "5 0 0 0 0 " ], "text/markdown": [ "\n", "A tibble: 5 × 381\n", "\n", "| cuisine <fct> | almond <dbl> | angelica <dbl> | anise <dbl> | anise_seed <dbl> | apple <dbl> | apple_brandy <dbl> | apricot <dbl> | armagnac <dbl> | artemisia <dbl> | ⋯ ⋯ | whiskey <dbl> | white_bread <dbl> | white_wine <dbl> | whole_grain_wheat_flour <dbl> | wine <dbl> | wood <dbl> | yam <dbl> | yeast <dbl> | yogurt <dbl> | zucchini <dbl> |\n", "|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|\n", "| chinese | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | ⋯ | 0 | 0 | 0 | 0 | 1 | 0 | 0 | 0 | 0 | 0 |\n", "| chinese | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | ⋯ | 0 | 0 | 0 | 0 | 1 | 0 | 0 | 0 | 0 | 0 |\n", "| chinese | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | ⋯ | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 0 |\n", "| chinese | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | ⋯ | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 0 |\n", "| chinese | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | ⋯ | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 0 |\n", "\n" ], "text/latex": [ "A tibble: 5 × 381\n", "\\begin{tabular}{lllllllllllllllllllll}\n", " cuisine & almond & angelica & anise & anise\\_seed & apple & apple\\_brandy & apricot & armagnac & artemisia & ⋯ & whiskey & white\\_bread & white\\_wine & whole\\_grain\\_wheat\\_flour & wine & wood & yam & yeast & yogurt & zucchini\\\\\n", " & & & & & & & & & & ⋯ & & & & & & & & & & \\\\\n", "\\hline\n", "\t chinese & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & ⋯ & 0 & 0 & 0 & 0 & 1 & 0 & 0 & 0 & 0 & 0\\\\\n", "\t chinese & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & ⋯ & 0 & 0 & 0 & 0 & 1 & 0 & 0 & 0 & 0 & 0\\\\\n", "\t chinese & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & ⋯ & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0\\\\\n", "\t chinese & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & ⋯ & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0\\\\\n", "\t chinese & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & ⋯ & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0\\\\\n", "\\end{tabular}\n" ], "text/html": [ "\n", "\n", "\n", "\t\n", "\t\n", "\n", "\n", "\t\n", "\t\n", "\t\n", "\t\n", "\t\n", "\n", "
A tibble: 5 × 381
cuisinealmondangelicaaniseanise_seedappleapple_brandyapricotarmagnacartemisia⋯whiskeywhite_breadwhite_winewhole_grain_wheat_flourwinewoodyamyeastyogurtzucchini
<fct><dbl><dbl><dbl><dbl><dbl><dbl><dbl><dbl><dbl>⋯<dbl><dbl><dbl><dbl><dbl><dbl><dbl><dbl><dbl><dbl>
chinese000000000⋯0000100000
chinese000000000⋯0000100000
chinese000000000⋯0000000000
chinese000000000⋯0000000000
chinese000000000⋯0000000000
\n" ] }, "metadata": {} }, { "output_type": "display_data", "data": { "text/plain": [ " cuisine n \n", "1 korean 559\n", "2 indian 418\n", "3 chinese 309\n", "4 japanese 224\n", "5 thai 202" ], "text/markdown": [ "\n", "A tibble: 5 × 2\n", "\n", "| cuisine <fct> | n <int> |\n", "|---|---|\n", "| korean | 559 |\n", "| indian | 418 |\n", "| chinese | 309 |\n", "| japanese | 224 |\n", "| thai | 202 |\n", "\n" ], "text/latex": [ "A tibble: 5 × 2\n", "\\begin{tabular}{ll}\n", " cuisine & n\\\\\n", " & \\\\\n", "\\hline\n", "\t korean & 559\\\\\n", "\t indian & 418\\\\\n", "\t chinese & 309\\\\\n", "\t japanese & 224\\\\\n", "\t thai & 202\\\\\n", "\\end{tabular}\n" ], "text/html": [ "\n", "\n", "\n", "\t\n", "\t\n", "\n", "\n", "\t\n", "\t\n", "\t\n", "\t\n", "\t\n", "\n", "
A tibble: 5 × 2
cuisinen
<fct><int>
korean 559
indian 418
chinese 309
japanese224
thai 202
\n" ] }, "metadata": {} } ], "metadata": { "colab": { "base_uri": "https://localhost:8080/", "height": 535 }, "id": "w5FWIkEiIjdN", "outputId": "2e195fd9-1a8f-4b91-9573-cce5582242df" } }, { "cell_type": "markdown", "source": [ "## 2. Töötamine tasakaalustamata andmetega\n", "\n", "Nagu võisite märgata nii algses andmekogumis kui ka meie treeningandmestikus, on köökide arvude jaotuses üsna suur ebavõrdsus. Korea köökide arv on *peaaegu* kolm korda suurem kui Tai köökide arv. Tasakaalustamata andmed mõjutavad sageli mudeli jõudlust negatiivselt. Paljud mudelid toimivad kõige paremini, kui vaatluste arv on võrdne, ja seetõttu kipuvad tasakaalustamata andmetega hätta jääma.\n", "\n", "Tasakaalustamata andmekogumitega tegelemiseks on peamiselt kaks viisi:\n", "\n", "- vähemuses oleva klassi vaatluste lisamine: `Üle-sampling`, näiteks SMOTE algoritmi kasutamine, mis sünteetiliselt genereerib vähemuses oleva klassi uusi näiteid, kasutades nende juhtumite lähimaid naabreid.\n", "\n", "- enamuses oleva klassi vaatluste eemaldamine: `Alam-sampling`\n", "\n", "Eelmises õppetunnis näitasime, kuidas tasakaalustamata andmekogumitega toime tulla, kasutades `recipe`-t. Retsepti võib mõelda kui plaani, mis kirjeldab, milliseid samme tuleks andmekogumile rakendada, et see oleks valmis andmeanalüüsiks. Meie puhul soovime, et meie `treeningandmestikus` oleks köökide arv ühtlaselt jaotatud. Asume asja kallale.\n" ], "metadata": { "id": "daBi9qJNIwqW" } }, { "cell_type": "code", "execution_count": 5, "source": [ "# Load themis package for dealing with imbalanced data\r\n", "library(themis)\r\n", "\r\n", "# Create a recipe for preprocessing training data\r\n", "cuisines_recipe <- recipe(cuisine ~ ., data = cuisines_train) %>% \r\n", " step_smote(cuisine)\r\n", "\r\n", "# Print recipe\r\n", "cuisines_recipe" ], "outputs": [ { "output_type": "display_data", "data": { "text/plain": [ "Data Recipe\n", "\n", "Inputs:\n", "\n", " role #variables\n", " outcome 1\n", " predictor 380\n", "\n", "Operations:\n", "\n", "SMOTE based on cuisine" ] }, "metadata": {} } ], "metadata": { "colab": { "base_uri": "https://localhost:8080/", "height": 200 }, "id": "Az6LFBGxI1X0", "outputId": "29d71d85-64b0-4e62-871e-bcd5398573b6" } }, { "cell_type": "markdown", "source": [ "Muidugi võid edasi minna ja kinnitada (kasutades prep+bake), et retsept töötab nii, nagu sa ootad - kõik köögi sildid sisaldavad `559` vaatlusi.\n", "\n", "Kuna me kasutame seda retsepti modelleerimise eeltöötlusena, teeb `workflow()` kogu ettevalmistuse ja küpsetamise meie eest, nii et me ei pea retsepti käsitsi hindama.\n", "\n", "Nüüd oleme valmis mudelit treenima 👩‍💻👨‍💻!\n", "\n", "## 3. Klassifikaatori valimine\n", "\n", "

\n", " \n", "

Kunstiteos: @allison_horst
\n" ], "metadata": { "id": "NBL3PqIWJBBB" } }, { "cell_type": "markdown", "source": [ "Nüüd peame otsustama, millist algoritmi töö jaoks kasutada 🤔.\n", "\n", "Tidymodelsis pakub [`parsnip pakett`](https://parsnip.tidymodels.org/index.html) ühtset liidest mudelitega töötamiseks erinevate mootorite (pakettide) vahel. Vaata parsnipi dokumentatsiooni, et uurida [mudelitüüpe ja mootoreid](https://www.tidymodels.org/find/parsnip/#models) ning nende vastavaid [mudeli argumente](https://www.tidymodels.org/find/parsnip/#model-args). Valik võib esmapilgul tunduda üsna segadust tekitav. Näiteks sisaldavad järgmised meetodid kõik klassifikatsioonitehnikaid:\n", "\n", "- C5.0 reeglipõhised klassifikatsioonimudelid\n", "\n", "- Paindlikud diskriminantmudelid\n", "\n", "- Lineaarsed diskriminantmudelid\n", "\n", "- Reguleeritud diskriminantmudelid\n", "\n", "- Logistilise regressiooni mudelid\n", "\n", "- Multinomiaalse regressiooni mudelid\n", "\n", "- Naivse Bayesi mudelid\n", "\n", "- Toetavate vektorite masinad\n", "\n", "- Lähimad naabrid\n", "\n", "- Otsustuspuud\n", "\n", "- Ansamblimeetodid\n", "\n", "- Neuraalvõrgud\n", "\n", "Nimekiri jätkub!\n", "\n", "### **Millist klassifikaatorit valida?**\n", "\n", "Niisiis, millist klassifikaatorit peaksid valima? Sageli on hea tulemuse leidmiseks kasulik proovida mitmeid ja võrrelda tulemusi.\n", "\n", "> AutoML lahendab selle probleemi elegantselt, tehes need võrdlused pilves ja võimaldades valida parima algoritmi sinu andmete jaoks. Proovi seda [siin](https://docs.microsoft.com/learn/modules/automate-model-selection-with-azure-automl/?WT.mc_id=academic-77952-leestott)\n", "\n", "Samuti sõltub klassifikaatori valik meie probleemist. Näiteks, kui tulemus saab olla kategoriseeritud `rohkem kui kahte klassi`, nagu meie puhul, tuleb kasutada `mitmeklassilist klassifikatsioonialgoritmi` mitte `binaarset klassifikatsiooni.`\n", "\n", "### **Parem lähenemine**\n", "\n", "Parem viis kui lihtsalt juhuslikult arvata, on järgida ideid sellelt allalaaditavalt [ML spikri lehelt](https://docs.microsoft.com/azure/machine-learning/algorithm-cheat-sheet?WT.mc_id=academic-77952-leestott). Siin avastame, et meie mitmeklassilise probleemi jaoks on mõned valikud:\n", "\n", "

\n", " \n", "

Microsofti algoritmi spikri lehe osa, mis kirjeldab mitmeklassilise klassifikatsiooni valikuid
\n" ], "metadata": { "id": "a6DLAZ3vJZ14" } }, { "cell_type": "markdown", "source": [ "### **Põhjendus**\n", "\n", "Vaatame, kas suudame erinevate lähenemiste abil leida lahenduse, arvestades meie piiranguid:\n", "\n", "- **Sügavad närvivõrgud on liiga rasked**. Arvestades meie puhast, kuid minimaalset andmekogumit ja asjaolu, et treenime mudelit lokaalselt märkmikus, on sügavad närvivõrgud selle ülesande jaoks liiga mahukad.\n", "\n", "- **Kaheklassilist klassifikaatorit ei kasutata**. Me ei kasuta kaheklassilist klassifikaatorit, seega välistame ühe-vs-kõik lähenemise.\n", "\n", "- **Otsustuspuu või logistiline regressioon võiks sobida**. Otsustuspuu võib sobida, samuti multinomiaalne regressioon/multiklassiline logistiline regressioon mitmeklassiliste andmete jaoks.\n", "\n", "- **Mitmeklassilised tõstetud otsustuspuud lahendavad teistsuguse probleemi**. Mitmeklassiline tõstetud otsustuspuu sobib kõige paremini mitteparametriliste ülesannete jaoks, näiteks järjestuste loomine, seega ei ole see meie jaoks kasulik.\n", "\n", "Samuti on tavaliselt enne keerukamate masinõppemudelite, näiteks ansamblimeetodite, kasutamist hea mõte luua kõige lihtsam võimalik mudel, et saada aimu, mis toimub. Seega alustame selles õppetükis `multinomiaalse regressiooni` mudeliga.\n", "\n", "> Logistiline regressioon on tehnika, mida kasutatakse siis, kui tulemuse muutujad on kategoorilised (või nominaalsed). Binaarse logistilise regressiooni korral on tulemuse muutujate arv kaks, samas kui multinomiaalse logistilise regressiooni korral on tulemuse muutujate arv rohkem kui kaks. Lisalugemiseks vaata [Advanced Regression Methods](https://bookdown.org/chua/ber642_advanced_regression/multinomial-logistic-regression.html).\n", "\n", "## 4. Treeni ja hinda multinomiaalse logistilise regressiooni mudelit.\n", "\n", "Tidymodels'is määratleb `parsnip::multinom_reg()` mudeli, mis kasutab lineaarseid ennustajaid mitmeklassiliste andmete ennustamiseks multinomiaalse jaotuse abil. Vaata `?multinom_reg()`, et näha erinevaid viise/mootoreid, mida saad kasutada selle mudeli sobitamiseks.\n", "\n", "Selles näites sobitame multinomiaalse regressiooni mudeli vaikimisi [nnet](https://cran.r-project.org/web/packages/nnet/nnet.pdf) mootori abil.\n", "\n", "> Valisin `penalty` väärtuse üsna juhuslikult. Selle väärtuse valimiseks on paremaid viise, näiteks `resampling` ja mudeli `tuning`, mida arutame hiljem.\n", ">\n", "> Vaata [Tidymodels: Get Started](https://www.tidymodels.org/start/tuning/), kui soovid rohkem teada saada mudeli hüperparameetrite häälestamise kohta.\n" ], "metadata": { "id": "gWMsVcbBJemu" } }, { "cell_type": "code", "execution_count": 6, "source": [ "# Create a multinomial regression model specification\r\n", "mr_spec <- multinom_reg(penalty = 1) %>% \r\n", " set_engine(\"nnet\", MaxNWts = 2086) %>% \r\n", " set_mode(\"classification\")\r\n", "\r\n", "# Print model specification\r\n", "mr_spec" ], "outputs": [ { "output_type": "display_data", "data": { "text/plain": [ "Multinomial Regression Model Specification (classification)\n", "\n", "Main Arguments:\n", " penalty = 1\n", "\n", "Engine-Specific Arguments:\n", " MaxNWts = 2086\n", "\n", "Computational engine: nnet \n" ] }, "metadata": {} } ], "metadata": { "colab": { "base_uri": "https://localhost:8080/", "height": 166 }, "id": "Wq_fcyQiJvfG", "outputId": "c30449c7-3864-4be7-f810-72a003743e2d" } }, { "cell_type": "markdown", "source": [ "Suurepärane töö 🥳! Nüüd, kui meil on retsept ja mudeli spetsifikatsioon, peame leidma viisi, kuidas need kokku panna objektiks, mis esmalt eeltöötleb andmed, seejärel sobitab mudeli eeltöödeldud andmetele ja võimaldab ka võimalikke järelprotsessimise tegevusi. Tidymodels'is nimetatakse seda mugavat objekti [`workflow`](https://workflows.tidymodels.org/) ja see hoiab mugavalt sinu modelleerimise komponente! Seda nimetaksime *Pythonis* *torustikeks*.\n", "\n", "Nii et paneme kõik kokku ühte workflow'sse!📦\n" ], "metadata": { "id": "NlSbzDfgJ0zh" } }, { "cell_type": "code", "execution_count": 7, "source": [ "# Bundle recipe and model specification\r\n", "mr_wf <- workflow() %>% \r\n", " add_recipe(cuisines_recipe) %>% \r\n", " add_model(mr_spec)\r\n", "\r\n", "# Print out workflow\r\n", "mr_wf" ], "outputs": [ { "output_type": "display_data", "data": { "text/plain": [ "══ Workflow ════════════════════════════════════════════════════════════════════\n", "\u001b[3mPreprocessor:\u001b[23m Recipe\n", "\u001b[3mModel:\u001b[23m multinom_reg()\n", "\n", "── Preprocessor ────────────────────────────────────────────────────────────────\n", "1 Recipe Step\n", "\n", "• step_smote()\n", "\n", "── Model ───────────────────────────────────────────────────────────────────────\n", "Multinomial Regression Model Specification (classification)\n", "\n", "Main Arguments:\n", " penalty = 1\n", "\n", "Engine-Specific Arguments:\n", " MaxNWts = 2086\n", "\n", "Computational engine: nnet \n" ] }, "metadata": {} } ], "metadata": { "colab": { "base_uri": "https://localhost:8080/", "height": 333 }, "id": "Sc1TfPA4Ke3_", "outputId": "82c70013-e431-4e7e-cef6-9fcf8aad4a6c" } }, { "cell_type": "markdown", "source": [ "Töövood 👌👌! **`workflow()`** saab sobitada peaaegu samamoodi nagu mudelit. Nii et aeg mudelit treenida!\n" ], "metadata": { "id": "TNQ8i85aKf9L" } }, { "cell_type": "code", "execution_count": 8, "source": [ "# Train a multinomial regression model\n", "mr_fit <- fit(object = mr_wf, data = cuisines_train)\n", "\n", "mr_fit" ], "outputs": [ { "output_type": "display_data", "data": { "text/plain": [ "══ Workflow [trained] ══════════════════════════════════════════════════════════\n", "\u001b[3mPreprocessor:\u001b[23m Recipe\n", "\u001b[3mModel:\u001b[23m multinom_reg()\n", "\n", "── Preprocessor ────────────────────────────────────────────────────────────────\n", "1 Recipe Step\n", "\n", "• step_smote()\n", "\n", "── Model ───────────────────────────────────────────────────────────────────────\n", "Call:\n", "nnet::multinom(formula = ..y ~ ., data = data, decay = ~1, MaxNWts = ~2086, \n", " trace = FALSE)\n", "\n", "Coefficients:\n", " (Intercept) almond angelica anise anise_seed apple\n", "indian 0.19723325 0.2409661 0 -5.004955e-05 -0.1657635 -0.05769734\n", "japanese 0.13961959 -0.6262400 0 -1.169155e-04 -0.4893596 -0.08585717\n", "korean 0.22377347 -0.1833485 0 -5.560395e-05 -0.2489401 -0.15657804\n", "thai -0.04336577 -0.6106258 0 4.903828e-04 -0.5782866 0.63451105\n", " apple_brandy apricot armagnac artemisia artichoke asparagus\n", "indian 0 0.37042636 0 -0.09122797 0 -0.27181970\n", "japanese 0 0.28895643 0 -0.12651100 0 0.14054037\n", "korean 0 -0.07981259 0 0.55756709 0 -0.66979948\n", "thai 0 -0.33160904 0 -0.10725182 0 -0.02602152\n", " avocado bacon baked_potato balm banana barley\n", "indian -0.46624197 0.16008055 0 0 -0.2838796 0.2230625\n", "japanese 0.90341344 0.02932727 0 0 -0.4142787 2.0953906\n", "korean -0.06925382 -0.35804134 0 0 -0.2686963 -0.7233404\n", "thai -0.21473955 -0.75594439 0 0 0.6784880 -0.4363320\n", " bartlett_pear basil bay bean beech\n", "indian 0 -0.7128756 0.1011587 -0.8777275 -0.0004380795\n", "japanese 0 0.1288697 0.9425626 -0.2380748 0.3373437611\n", "korean 0 -0.2445193 -0.4744318 -0.8957870 -0.0048784496\n", "thai 0 1.5365848 0.1333256 0.2196970 -0.0113078024\n", " beef beef_broth beef_liver beer beet\n", "indian -0.7985278 0.2430186 -0.035598065 -0.002173738 0.01005813\n", "japanese 0.2241875 -0.3653020 -0.139551027 0.128905553 0.04923911\n", "korean 0.5366515 -0.6153237 0.213455197 -0.010828645 0.27325423\n", "thai 0.1570012 -0.9364154 -0.008032213 -0.035063746 -0.28279823\n", " bell_pepper bergamot berry bitter_orange black_bean\n", "indian 0.49074330 0 0.58947607 0.191256164 -0.1945233\n", "japanese 0.09074167 0 -0.25917977 -0.118915977 -0.3442400\n", "korean -0.57876763 0 -0.07874180 -0.007729435 -0.5220672\n", "thai 0.92554006 0 -0.07210196 -0.002983296 -0.4614426\n", " black_currant black_mustard_seed_oil black_pepper black_raspberry\n", "indian 0 0.38935801 -0.4453495 0\n", "japanese 0 -0.05452887 -0.5440869 0\n", "korean 0 -0.03929970 0.8025454 0\n", "thai 0 -0.21498372 -0.9854806 0\n", " black_sesame_seed black_tea blackberry blackberry_brandy\n", "indian -0.2759246 0.3079977 0.191256164 0\n", "japanese -0.6101687 -0.1671913 -0.118915977 0\n", "korean 1.5197674 -0.3036261 -0.007729435 0\n", "thai -0.1755656 -0.1487033 -0.002983296 0\n", " blue_cheese blueberry bone_oil bourbon_whiskey brandy\n", "indian 0 0.216164294 -0.2276744 0 0.22427587\n", "japanese 0 -0.119186087 0.3913019 0 -0.15595599\n", "korean 0 -0.007821986 0.2854487 0 -0.02562342\n", "thai 0 -0.004947048 -0.0253658 0 -0.05715244\n", "\n", "...\n", "and 308 more lines." ] }, "metadata": {} } ], "metadata": { "colab": { "base_uri": "https://localhost:8080/", "height": 1000 }, "id": "GMbdfVmTKkJI", "outputId": "adf9ebdf-d69d-4a64-e9fd-e06e5322292e" } }, { "cell_type": "markdown", "source": [ "Väljund näitab koefitsiente, mida mudel õppis treeningu ajal.\n", "\n", "### Treenitud mudeli hindamine\n", "\n", "On aeg näha, kuidas mudel toimis 📏, hinnates seda testkomplekti peal! Alustame ennustuste tegemisest testkomplekti põhjal.\n" ], "metadata": { "id": "tt2BfOxrKmcJ" } }, { "cell_type": "code", "execution_count": 9, "source": [ "# Make predictions on the test set\n", "results <- cuisines_test %>% select(cuisine) %>% \n", " bind_cols(mr_fit %>% predict(new_data = cuisines_test))\n", "\n", "# Print out results\n", "results %>% \n", " slice_head(n = 5)" ], "outputs": [ { "output_type": "display_data", "data": { "text/plain": [ " cuisine .pred_class\n", "1 indian thai \n", "2 indian indian \n", "3 indian indian \n", "4 indian indian \n", "5 indian indian " ], "text/markdown": [ "\n", "A tibble: 5 × 2\n", "\n", "| cuisine <fct> | .pred_class <fct> |\n", "|---|---|\n", "| indian | thai |\n", "| indian | indian |\n", "| indian | indian |\n", "| indian | indian |\n", "| indian | indian |\n", "\n" ], "text/latex": [ "A tibble: 5 × 2\n", "\\begin{tabular}{ll}\n", " cuisine & .pred\\_class\\\\\n", " & \\\\\n", "\\hline\n", "\t indian & thai \\\\\n", "\t indian & indian\\\\\n", "\t indian & indian\\\\\n", "\t indian & indian\\\\\n", "\t indian & indian\\\\\n", "\\end{tabular}\n" ], "text/html": [ "\n", "\n", "\n", "\t\n", "\t\n", "\n", "\n", "\t\n", "\t\n", "\t\n", "\t\n", "\t\n", "\n", "
A tibble: 5 × 2
cuisine.pred_class
<fct><fct>
indianthai
indianindian
indianindian
indianindian
indianindian
\n" ] }, "metadata": {} } ], "metadata": { "colab": { "base_uri": "https://localhost:8080/", "height": 248 }, "id": "CqtckvtsKqax", "outputId": "e57fe557-6a68-4217-fe82-173328c5436d" } }, { "cell_type": "markdown", "source": [ "Suurepärane töö! Tidymodelsis saab mudeli jõudlust hinnata, kasutades [yardstick](https://yardstick.tidymodels.org/) - paketti, mida kasutatakse mudelite tõhususe mõõtmiseks jõudlusmõõdikute abil. Nagu tegime oma logistilise regressiooni tunnis, alustame segadusmaatriksi arvutamisest.\n" ], "metadata": { "id": "8w5N6XsBKss7" } }, { "cell_type": "code", "execution_count": 10, "source": [ "# Confusion matrix for categorical data\n", "conf_mat(data = results, truth = cuisine, estimate = .pred_class)\n" ], "outputs": [ { "output_type": "display_data", "data": { "text/plain": [ " Truth\n", "Prediction chinese indian japanese korean thai\n", " chinese 83 1 8 15 10\n", " indian 4 163 1 2 6\n", " japanese 21 5 73 25 1\n", " korean 15 0 11 191 0\n", " thai 10 11 3 7 70" ] }, "metadata": {} } ], "metadata": { "colab": { "base_uri": "https://localhost:8080/", "height": 133 }, "id": "YvODvsLkK0iG", "outputId": "bb69da84-1266-47ad-b174-d43b88ca2988" } }, { "cell_type": "markdown", "source": [ "Kui tegeleda mitme klassiga, on üldiselt intuitiivsem seda kujutada soojuskaardina, nagu see:\n" ], "metadata": { "id": "c0HfPL16Lr6U" } }, { "cell_type": "code", "execution_count": 11, "source": [ "update_geom_defaults(geom = \"tile\", new = list(color = \"black\", alpha = 0.7))\n", "# Visualize confusion matrix\n", "results %>% \n", " conf_mat(cuisine, .pred_class) %>% \n", " autoplot(type = \"heatmap\")" ], "outputs": [ { "output_type": "display_data", "data": { "text/plain": [ "plot without title" ], "image/png": "iVBORw0KGgoAAAANSUhEUgAAA0gAAANICAMAAADKOT/pAAADAFBMVEUAAAABAQECAgIDAwMEBAQFBQUGBgYHBwcICAgJCQkKCgoLCwsMDAwNDQ0ODg4PDw8QEBARERESEhITExMUFBQVFRUWFhYXFxcYGBgZGRkaGhobGxscHBwdHR0eHh4fHx8gICAhISEiIiIjIyMkJCQlJSUmJiYnJycoKCgpKSkqKiorKyssLCwtLS0uLi4vLy8wMDAxMTEyMjIzMzM0NDQ1NTU2NjY3Nzc4ODg5OTk6Ojo7Ozs8PDw9PT0+Pj4/Pz9AQEBBQUFCQkJDQ0NERERFRUVGRkZHR0dISEhJSUlKSkpLS0tMTExNTU1OTk5PT09QUFBRUVFSUlJTU1NUVFRVVVVWVlZXV1dYWFhZWVlaWlpbW1tcXFxdXV1eXl5fX19gYGBhYWFiYmJjY2NkZGRlZWVmZmZnZ2doaGhpaWlqampra2tsbGxtbW1ubm5vb29wcHBxcXFycnJzc3N0dHR1dXV2dnZ3d3d4eHh5eXl6enp7e3t8fHx9fX1+fn5/f3+AgICBgYGCgoKDg4OEhISFhYWGhoaHh4eIiIiJiYmKioqLi4uMjIyNjY2Ojo6Pj4+QkJCRkZGSkpKTk5OUlJSVlZWWlpaXl5eYmJiZmZmampqbm5ucnJydnZ2enp6fn5+goKChoaGioqKjo6OkpKSlpaWmpqanp6eoqKipqamqqqqrq6usrKytra2urq6vr6+wsLCxsbGysrKzs7O0tLS1tbW2tra3t7e4uLi5ubm6urq7u7u8vLy9vb2+vr6/v7/AwMDBwcHCwsLDw8PExMTFxcXGxsbHx8fIyMjJycnKysrLy8vMzMzNzc3Ozs7Pz8/Q0NDR0dHS0tLT09PU1NTV1dXW1tbX19fY2NjZ2dna2trb29vc3Nzd3d3e3t7f39/g4ODh4eHi4uLj4+Pk5OTl5eXm5ubn5+fo6Ojp6enq6urr6+vs7Ozt7e3u7u7v7+/w8PDx8fHy8vLz8/P09PT19fX29vb39/f4+Pj5+fn6+vr7+/v8/Pz9/f3+/v7////isF19AAAACXBIWXMAABJ0AAASdAHeZh94AAAgAElEQVR4nO3deWBU9b3//0+ibApWrbYuvYorXaxoaatWvVqpqG2HsCmLBAqoVXBDjCKbKMqOQUDFFVxKqyhVFLUqWKJsxg3Lz2IFGilLiEqptMX0hpzvnJkMCbx5/W5vz5k5Z+D5/OOc85nEz3w8Mw9mMjmo84gocC7qBRDtCQGJKISARBRCQCIKISARhRCQiEIISEQhBCSiEAISUQgBiSiEgEQUQkAiCiEgEYUQkIhCCEhEIQQkohACElEIAYkohIBEFEJAIgohIBGFEJCIQijHkLb+NUZVRb2Ahn26OeoVNCxWpyZWi/mbeGbnGNJl42PUmbNiVM+7nohRlz4ao/o/HqOuFc/sHEMasShGdfwsRt3+1qYYNXJ9jLq9MkZNE89sIMUkIMmAZAOSCkgyINmApAKSDEg2IKmAJAOSDUgqIMmAZAOSCkgyINmApAKSDEg2IKmAJAOSDUgqIMmAZAOSCkgyINmApAKSDEg2IKmAJAOSDUgqIMmAZAOSCkgyINmApAKSDEg2IKmAJAOSDUgqIMmAZAOSCkgyINmApAKSDEg2IKmAJAOSDUgqIMmAZAOSCkgyINmApAKSDEg2IKmAJAOSDUgqIMmAZAOSCkgyINmApAKSDEg2IKmAJAOSDUgqIMmAZAOSCkgyINmApAKSDEg2IKmAJAOSDUgqIMmAZAOSCkgyINmApAKSDEg2IKmAJAOSDUgqIMmAZAOSCkgyINmApAKSDEg2IKmAJAOSDUgqIMmAZAOSCkgyINmApAKSDEg2IKmAJAOSDUgqIMmAZAOSCkgyINmApAKSDEg2IKmAJAOSDUgqIMmAZAOSCkgyINmApAKSDEg2IKmAJAOSDUgqIMmAZAOSCkgyINmApAKSDEg2IKmAJAOSDUgqIMmAZAOSCkgyINmApAKSDEg2IKmAJAOSDUgqIMmAZAOSCkgyINmApAKSDEg2IKmAJAOSDUgqIMmAZAOSCkgyINmApAKSbI+EVLQktatJvB8FpNGtv9LoqMteTx7d/f2vND6h5M3oIS07xT0XxjyhQJpx+sGNj79pbfCJgkJ6o7V72t/f4FKdFSmkRa3dnNTBgnYHNPneY7GCVPvB1ggg3ezaTZraq+C8RYvGF7a64cbW7rLIIU1sdmR8IE1ynX4957qC9pFDGtvsiDSkywsn+D0eJaRxycWkIC1pcdzYSecUzIwTpP+0YJBOONJ/CTqncP6iI49YsGjRwqMOjhrSS03GTY0PpJNaVia3P92nImJIc5vcWZqG1LVFoInCgPRCkzF3pyF1bLa8snLdd1pGC+nTOy8uvvdLr+iVEZ2KF/hv7WoTC0f07zvf8zaP79Vl8CrPe+2qzsX3Vu8YZgHS8cf624sKF5RdO84/+plbEDGk8oWfxQjSt7/pb7vu80ngmYJBWvTa+jpIPz0ickhLFlSmIa1vVuSPR7lXI4V0w9jN6wdM94qu+fCfj3XZ5v+MVDRwi/dKl23eoPFfVD/es3pj+/e3b7xudmaYDUjD3C+fmz+6aee64Zsnfz3QdOF82BAjSFPckOV/nrFfv+AzBf6woQ7SWa3Wr18dLaRkaUiL3FB/MMdNjhLS6sTG5KbcK3ra8zYmKlKQ5nrepsQnqxKbkz8zdStblVjtedu9zDD5zyxpn+wPIUJadFsz5wp7pz5i+P1vH2i372gg7dT9+yfPz/WVwScKC1Lrlh0PdAddvyYOkJ5zd/mDN9ywKCG92b42tS9anHwvl/g4BSl9WJZINbv2ng4ls9Z7mWHye9/4cbL3QoR0T/MzRt91SWHqI4bJzh0+KdBsex6kZw/4yYzfXL7PzfGB1LKw20P3F7mL4gDpSTfVHyxzN0YJaVH77WlIS+ohpQ+XJjJv4zbNG9mhrH64uwJBeuPwE/0Xo66FTya3L44f2ragF5AatPGo7/ovRlcULo0NpLff87dd3ZwYQHrOTfIHZW54lJDWJCo876MXdgNpbWJl8usbvZotyd30wZlhFiA97VJwJrjMLL9wDwGpvrfddf7uCXdPbCCle8IFmi8kSEvcLf7gqfQLU1SQvEEjKtddd+9uIHlDS6pqXuzy+at9Pq7dPGRKZpgVSD383Wg3+PkbHkyTugVI9ZW7/v7uEXdXbCCtXOlv73fjYgBpQ4uf+4MhrixSSFvu6NJz2rbdQdo8ruslJSu82ll9Ova6+++ZYRYgvdH8mDeSuw7usRcLT/WPLnGTgVTfxq+02pjc9Xa/jwukdwsv9AfnFbwRA0iVlzZ5p7Jy7bHfDjTXHnGt3UB32qgJFxe2XbSo2H332pLzC77zRsSQ5pWW9nADSkvfiQOkTXe6Hz/wxOWFRcFnCgbp2QkTurorJ0xYvL6Paztu1OmuX6DpgkGaO2lSN9d/0qRlle8dfPTQO37QaA6QFo06qWmjlleWLVr0Zkmrps2O7fFqoNlCgNQ7fS2ZezAWkDY9+P39Gp8wZH3wiYJB6ll3Vu5dv3ZM6xZNT5kYaLaAkHrVLWZ6ZeWiC1s0Pe2ZQLPtIZDCjau/ZVz9rQKSDUgyIKmAZAOSDEgqINmAJAOSCkg2IMmApAKSDUgyIKmAZAOSDEgqINmAJAOSCkg2IMmApAKSDUgyIKmAZAOSDEgqINmAJAOSCkg2IMmApAKSDUgyIKmAZAOSDEgqINmAJAOSCkg2IMmApAKSDUgyIKmAZAOSDEgqINmAJAOSCkg2IMmApAKSDUgyIKmAZAOSDEgqINmAJAOSCkg2IMmApAKSDUgyIKmAZAOSDEgqINmAJAOSCkg2IMmApAKSDUgyIKmAZAOSDEgqINmAJAOSCkg2IMmApAKSDUgyIKmAZAOSDEgqINmAJAOSCkg2IMmApAKSDUgyIKmAZAOSDEgqINmAJAOSCkg2IMmApAKSDUgyIKmAZAOSDEgqINmAJAOSCkg2IMmApAKSDUgyIKmAZAOSDEgqINmAJAOSCkg2IMmApAKSDUgyIKmAZAOSDEgqINmAJAOSCkg2IMmApAKSDUgyIKmAZAOSDEgqINmAJAOSCkg2IMmApAKSDUgyIKmAZAOSDEgqINmAJAOSCkg2IMmApAKSDUgyIKmAZAOSDEgqINmAJAOSCkg2IMmApAKSDUgyIKmAZAOSDEiqmEDqOSJGnXJfjOpw+z0xqtOUGNU76rPRsMvEMzvHkEa+GqMuHBajznv8tRg1IOoFNGzgKzFqiHhm5xjSvZ/GqF/MjFHd3on6zWXD7ox6AQ27oypG3Sue2UCKSUCSAckGJBWQZECyAUkFJBmQbEBSAUkGJBuQVECSAckGJBWQZECyAUkFJBmQbEBSAUkGJBuQVECSAckGJBWQZECyAUkFJBmQbEBSAUkGJBuQVECSAckGJBWQZECyAUkFJBmQbEBSAUkGJBuQVECSAckGJBWQZECyAUkFJBmQbEBSAUkGJBuQVECSAckGJBWQZECyAUkFJBmQbEBSAUkGJBuQVECSAckGJBWQZECyAUkFJBmQbEBSAUkGJBuQVECSAckGJBWQZECyAUkFJBmQbEBSAUkGJBuQVECSAckGJBWQZECyAUkFJBmQbEBSAUkGJBuQVECSAckGJBWQZECyAUkFJBmQbEBSAUkGJBuQVECSAckGJBWQZECyAUkFJBmQbEBSAUkGJBuQVECSAckGJBWQZECyAUkFJBmQbEBSAUkGJBuQVECSAckGJBWQZECyAUkFJBmQbEBSAUkGJBuQVECSAckGJBWQZECyAUkFJBmQbEBSAUkGJBuQVECSAckGJBWQZECyAUkFJBmQbEBSAUkGJBuQVECSAckGJBWQZECyAUkFJBmQbEBSAUkGJBuQVECSAckGJBWQZECyhQTpraYHhzDLL/7zp/3YY9zg1MFDHQ7Z92sXz0ge3fTt5o2O6j0jWkjLTnHPhTFPGJDmt23e/OTSquAThQFpXcl/NW45bFPwiUKEVJN4f6fxpkTFrjdlGVLVmS5aSL0bH1wH6QeFF155puswc+b1BUf37HWi6xQppInNjowNpJcbtbx90jnuluAzhQEpsc9V0y9xJcEnChFS7Qdbd4W0601ZhjSp8bmRQhrWqPiyNKQS1y25/f43Z8z82qEPzJz58GEHRAnppSbjpsYG0o8O+ONnn1V9Z7+NgWcKAdJsd1ty+/Mzg78kZfGtXRLSv/29oUD6wwEll0YKadyomXWQftT0ofRND/e4zt+d7R6IEFL5ws/iA2nydH/bx/0p8EwhQOrSfF3wSVKF+9auNrFwRP++8z1v9aAuVy9Mv7WrGN6964gN3o4vZQ/SRSeujxZSsjpIh540c2aDH4tmnPDV/3TCkD5siA+kdOceGnyOECAdfW5VVWXwaapC/xmpaOAW75Uu22r7lW6rGpKGdGXptn+MKfEyX8oepIcK5n0aE0gzCs7t8/WC/S9KvQw9dNfw0/e5BkgNe9jdHnyS4JA2FfaadEzBQf0/iR+kuf5buk/+mNjoeUvSkLZ+6XmLO9RmvpT8xgVtkr0dNqQ/HdL307hAut8deuxVN15Y0Ma/qcS5Q274jyfcIyH9utlFsfjUrsId9b0Hnrqq8Gfxg7TY8zYnPi5rv93zPklDWj6kuLhboibzpeQ3lvdM9mHYkLoeviY2kB50zacndz9xtya3U6+/7LSCBJDqG7dPpw0hTBMc0l/cQWuSu8vcK7GDtCSlZX77Ws9bk4K0odPsam+pD2lJBtJuCg7pqYKHKyoquh1csS4GkGY2+6a/HeT61N3cPkUKSKmudIM+DWOeEH5GanGmv/2NuyvwTNmBtDxR6XllKUhlRTWe92j2IfVzdZ0fB0itDvO317orphQP948Gur5AqmtgQWkIs3wWCqQzjve3j7l7As+UHUjVPUq3rrs5BWllYsW/Fg5OVGUb0tsv+LU74IU34wCplytJbs8oHD+14Jv+p3ftUmMgJXvahfWJRQiQxrunk9su+7wVeKbsQPI+ur7z1e8k/uzfNKN7jylbB3bblGVI6aL9GWlonz5nu4v69Jkw86GWTdr3+6E7f+bMn7nju/c+veC4//QaoTAgzSst7eEGlJYGnyq4gcrjDipN9V7gqUKAtK71foPuLnKXB59pz7rWLmJIP657d3nVzJn3tv3KPocVJ/XM6H1046bfuGj6fzpnGJB6163rwcAzBYf0UeYt+GOBpwrjEqGP+3yt0XFj43WtXZC4+lvF1d8yrv62AUkFJBmQbEBSAUkGJBuQVECSAckGJBWQZECyAUkFJBmQbEBSAUkGJBuQVECSAckGJBWQZECyAUkFJBmQbEBSAUkGJBuQVECSAckGJBWQZECyAUkFJBmQbEBSAUkGJBuQVECSAckGJBWQZECyAUkFJBmQbEBSAUkGJBuQVECSAckGJBWQZECyAUkFJBmQbEBSAUkGJBuQVECSAckGJBWQZECyAUkFJBmQbEBSAUkGJBuQVECSAckGJBWQZECyAUkFJBmQbEBSAUkGJBuQVECSAckGJBWQZECyAUkFJBmQbEBSAUkGJBuQVECSAckGJBWQZECyAUkFJBmQbEBSAUkGJBuQVECSAckGJBWQZECyAUkFJBmQbEBSAUkGJBuQVECSAckGJBWQZECyAUkFJBmQbEBSAUkGJBuQVECSAckGJBWQZECyAUkFJBmQbEBSAUkGJBuQVECSAckGJBWQZECyAUkFJBmQbEBSAUkGJBuQVECSAckGJBWQZECyAUkFJBmQbEBSAUkGJBuQVECSAckGJBWQZECyAUkFJBmQbEBSAUkGJBuQVECSAckGJBWQZECyAUkFJBmQbEBSAUkGJBuQVECSAcl28S9jVKt2Mer4blGfjob9KOoFNOxnV8Soi8QzO8eQ7loVo7r/JUYNf/KNGDXkTzFq1PoYNU08s3MMafLaGNUz6rcJDbvtmWUxanhFjIrV+8z7xDMbSDEJSDIg2YCkApIMSDYgqYAkA5INSCogyYBkA5IKSDIg2YCkApIMSDYgqYAkA5INSCogyYBkA5IKSDIg2YCkApIMSDYgqYAkA5INSCogyYBkA5IKSDIg2YCkApIMSDYgqYAkA5INSCogyYBkA5IKSDIg2YCkApIMSDYgqYAkA5INSCogyYBkA5IKSDIg2YCkApIMSDYgqYAkA5INSCogyYBkA5IKSDIg2YCkApIMSDYgqYAkA5INSCogyYBkA5IKSDIg2YCkApIMSDYgqYAkA5INSCogyYBkA5IKSDIg2YCkApIs/yBt6n1EoUsFpFwHJFn+Qbp437a9+6UCUq4Dkiz/IH312WwBAtL/FpBk+QdpvyogRRWQZPkH6ezXgRRVQJLlH6S3f7gYSBEFJFn+QTrzv9x+R6cCUq4Dkiz/IJ3dNhOQch2QZPkHKfsBSQUkWT5C+uyFBx56+Qsg5T4gyfIP0vZBjfzLGvYfD6ScByRZ/kEa7zo+/OIL91/gHgVSrgOSLP8gfeuG9P6K7wEp1wFJln+QmsxP7+c1A1KuA5Is/yDt/3x6/2xzIOU6IMnyD9JZP672d9vanQukXAckWf5Bmldw1JWjbr/8iMJXgZTrgCTLP0jeb7/pf/z93XnZcgQkGZBkeQjJ89a/VV6ZNUZA0gFJlpeQshyQVECS5RmkVqO9VjsCUq4DkizPIJ1W6p22IyDlOiDJ8gxSTgKSCkiy/IPU5sP0/ulvASnXAUmWf5BceWr3P7c1BlKuA5Is3yC5+rhoNecBSZZvkN6/2xWl/uuQl434C5ByHZBk+QbJ8y74U7YAAel/C0iy/IPkbZyS3FTdtglIOQ9IsvyDtPIw/1OGCnfYaiDlOiDJ8g9Sh+Pf8ncfHt8JSLkOSLL8g3ToI+n9/S2AlOuAJMs/SM2eSO9/tV9MIc07t3nzk8ZW+Ie/P9k9GT2kkvSvC86OGtJrmV9cjF+2bNoPvtL4xMFLo4T0/Dn773/SmDUVFdenV3Vm9JCWneKeC2OefwvSjy6o8Xdf/ODMzC01ifdjBOnZfY8eNuYsd2PycHSzI+IA6ZeFd/n9OmpIbw5JdX7Br5ZNKmx1402nuCsihPTbfY8eOvosN6iiom/hWL+ZkUOa2OzIHEJ6ueDYASNH9Dm08OXMLbUfbI0RpNNbvLt2bcW391uz9rdNRk2KA6TuBwSfI10Yb+1eP7TDsmXfOLJs2bJFRx8cIaTTWrxdUbHmW/utqri4RaCJQoP0UpNxU3MIyXuljf9CfHJc/4bs+Lv9bbFbvrbsd2tjAelnRwafI10YkC458NVliwdO9A9/7sqigzRusr/t6d6ruPDweEAqX/hZTiF53mcf/H8N/4vF/lu7iuHdu47Y4FUnXh7cr+9SLzOuTSwc0b/vfM/bPL5Xl8GrPO+1qzoX31u9Y5gFSOnOPiS1iwWks1tVVa0NPk1VKJCeLCzJHC5tfVigqcL4sOHsQyoqzjyxomJlDCAlyzGkXfIhXVm67R9jSpKH1/3Ve7XDlszYKxq4xXulyzZv0Pgvqh/vWb2x/fvbN143OzNM/sP/XJfsy7Ah3eeGxQfSKcd0PsgdNOgvsYB0/qFvpPZvzH34gn3HRg3pHje0ouLklkUHuoOu/WhvgrTbvyHrQ9qatLC4Q21N4jnP2971lczYK5rreZsSn6xKbE7+LNWtbFVidfLrXmaY/IcXtEn2dsiQZjZrVxEfSMcU9pj5UEf30+AzBYf0ZOGg9MFU5w4vDTZXcEiPNDt/TUVFy8JL7r8n4S7YmyDt9m/I+pCWDyku7paoqUksS95w1azM2CtanHxbl/i4LJFqdu09HUpmrfcyw+T3rrg52c4XSQSGNGqfotVr4wPp/RX+trubG3im4JC6Nn49ffC7icPPL/hFtJBu36f9x8ndknJ/cLF7ai+CtNuSkDZ0ml3tLfUh+f9fzCt+nRl7RUtSkJYmquu+edO8kR3K6oe7Kyikfu7aT9bGCFK637hRgecIDGnp13/UYNTXzYgSUl93zZ/rR4+6EUB6v6yoxvMe9SE97XnVnV/LjDOQ1iZWJr9xo1ezJbmbPjgzzAqkqwvG7jiOBaTVq/3tQ25i4JkCQ3rE3eLvXip5xN/d5YZGCGlAwZj0wYoV/vYeN3ovgrR/g3b8DdkkpJWJFf9aODhRVZMYUFE9q+PfMuMMJG9oSVXNi10+f7XPx7Wbh0zJDLMB6VduZP0gDpA+KLzI37UtWBI9pKvdr/zd7wq/tyS56+qmRgfp8cwr0LLCdv7u3ILX9yJIXZO1anRG5w6nFLS5ugEkb0b3HlO2Duy2IfHiTZ37lXuZ8aYMpM3jul5SssKrndWnY6+7/54ZZgHSmmMPHDvOb8naOePGXeJ+OW7cm9FCqurnzp8w+gx3efCZAkP6uft9at/bnXz9ze0KTloSGaRVxxw4JnVBw6KK3u680SN/6PoEmS4MSPNKS3u4AaWl7+QAUrLZJ23wdyu/ObchpB2H74g5/g8FgvR+5oKyB9deWnc0LWJIG8efckDTU0uDTxQc0tmF6f3Swa2aNjuu5+uBJgsE6d3M4/RAxeo7Tm7RtPW4QI7CgNQ788zJDaSTnkrv72tdd8P2jxI7PnWLHlLYcfW3jKu/Vf8WpMavpfezm9TdsLDDqFog5SQgyfIP0hGXpna1XQ8PTgZI/7eAJMs/SLe67147atSAb7nBQMp1QJLlH6TacYf7P5EdMrwGSLkOSLL8g5Sk9Mmypau3Z4sRkHRAkuUjpG1vzfnU+x8g5T4gyfIQ0sQWzi3xhvwia5SApAKSLP8gPeDaT09CenTf8UDKdUCS5R+kk6/0tiUhebecCKRcByRZ/kFq+moa0u8aASnXAUmWf5C+9nwa0lMHACnXAUmWf5B+cs4/fUifn9QOSLkOSLL8g/T6Psdf5/r2PqDRm0DKdUCS5R8k77VT/Ssbfvj7bDkCkgxIsjyE5Hmb3ntvc9YYAUkHJFn+QToje/+JVSD9LwFJln+QvjEJSFEFJFn+QXruW7/9F5CiCUiy/IN09ndd4yOO9gNSrgOSLP8gnXle27qAlOuAJMs/SNkPSCogyfIO0rZlb24BUkQBSZZvkCa3cK5R/y/FNwIpuwFJlmeQnnEtbxh2lrtafCOQshuQZHkG6eyW/v8utm+jvwEpioAkyzNIzYf727dc1i5YBdL/X0CS5Rkkd7+/3eBeFt8JpKwGJFm+QXrQ3250LwEpioAkAxKQ/v2AJMs3SLcsSTbPlfo7IOU6IMnyDVLDgJTrgCTLM0i3NgxIuQ5IsjyDlJOApAKSDEg2IKmAJAOSDUgqIMmAZAOSCkgyINmApAKSDEg2IKmAJAOSDUgqIMmAZAOSCkgyINmApAKSDEg2IKmAJAOSDUgqIMmAZAOSCkgyINmApAKSDEg2IKmAJAOSDUgqIMmAZAOSCkgyINmApAKSDEg2IKmAJAOSDUgqIMmAZAOSCkgyINmApAKSDEg2IKmAJAOSDUgqIMmAZAOSCkgyINmApAKSDEg2IKmAJAOSDUgqIMmAZAOSCkgyINmApAKSDEg2IKmAJAOSDUgqIMmAZAOSCkgyINmApAKSDEg2IKmAJAOSDUgqIMmAZOvQM0Yd/4sYdWq7TjHqB5fGqJ/2iVEXimd2jiFNWR+jij+PUaOWboxRiWkxanTUj03DYvKKBCQVkGRAsgFJBSQZkGxAUgFJBiQbkFRAkgHJBiQVkGRAsgFJBSQZkGxAUgFJBiQbkFRAkgHJBiQVkGRAsgFJBSQZkGxAUgFJBiQbkFRAkgHJBiQVkGRAsgFJBSQZkGxAUgFJBiQbkFRAkgHJBiQVkGRAsgFJBSQZkGxAUgFJBiQbkFRAkgHJBiQVkGRAsgFJBSQZkGxAUgFJBiQbkFRAkgHJBiQVkGRAsgFJBSQZkGxAUgFJBiQbkFRAkgHJBiQVkGRAsgFJBSQZkGxAUgFJBiQbkFRAkgHJBiQVkGRAsgFJBSQZkGxAUgFJBiQbkFRAkgHJBiQVkGRAsgFJBSQZkGxAUgFJBiQbkFRAkgHJBiQVkGRAsgFJBSQZkGxAUgFJBiQbkFRAkgHJBiQVkGRAsgFJBSQZkGxAUgFJBiQbkFRAkgHJBiQVkGRAsgFJBSQZkGxAUgFJBiQbkFRAkgHJBiQVkGRAsgFJBSQZkGxAUgFJBiQbkFRAkgHJBiQVkGRAsgFJBSQZkGxAUgFJBiQbkFRAkgHJBiQVkGRAsgFJBSQZkGxAUgFJBiQbkFRAkgHJBiQVkGRAsgFJBSQZkGxAUgFJBiRbUEhvtHZP+/sbXKqzIodUduEBTdr8KoSJAkNa1No9s+tRFJBGHOWuSx1cf3zjxifcsPNtkUEK7XEKBVJN4p1oIY1tdkQa0uWFE/wejxrS2y2OmzD53IIngs8UFNK45Kl5ZpejKCB1a3xQGs2V7shLLj1s35sa3hYZpPAepz0C0twmd5amIXVtEWii0CB1bvbh559vOumY4DMFhPR8k9GT03zqj6KANKjRJT3TaA498K5p0ya0aNXwtsgghfc47RGQFr22vg7ST4+IBaSqZh393Wj3euCpAkJaPH9jHZ/6oygg3XrLtDSaMe4sf9y2YHz9bZFBCvFxCg1SzbCRNX8d36tzyYfe9sTv+k32No/v1WXwKs+rGN6964gNXm1i4Yj+fednBVKyOkhntVq/fnX0kJa54f5urpsaeKrgHzbU84kQUrI0mjvcef6gixtYf1tkkEJ8nEKDVFrypTfo1i1fPtz1b17RwFX/9AaN/6L68Z7V3pWl2/4xpsRL3rjFe6XLtuS3f74s2d+yAql1y44HuoOuXxMxpBfc3f5uiRsReKo9DdLU/Y7yB23cZTGAFOLjFBakJ/p/4a1OrPW86osXeEVPet6qxGbPq+1W5m390vMWd6j1iuZ63qbEJ8lvX9Am2dtZgdSysNtD9xe5iyKG9Iy7z9+9424KPNWeBmlawv33yNsuaOH6xgBSiI9TSJDGJv7geW+2r00O+v/GKyrzvLJEqtne8iHFxd0SNV7RYs/bnPcB240AABDoSURBVPg4+R2rpyRblxVIb7/nb7u6OdFCmucm+7vF7tbAU+1xkO4+r8C5b13qrowBpBAfp5Ag9RsxsKYO0lVPeEVLPG9pojr1tQ2dZlcnBzWpG9OQdlNYkNI94UZGC+ltN8zfzUn/gReoPQ7StGljS+5M/ow0LAaQQnycQoJUvrXPI94a/43bts7zU2bWJlYmv7LRKyuq8bxHcwVp5Up/e78bFy2kT1sk/N1wtzjwVHsgJL/v7jclBpBCfJxC+7BhRYd3vZKRX2y7r+c/Uma8oSVVNS92+XxlYsW/Fg5OVOUE0ruFF/qD8wreiBbS58VNln/++YZjvxN8pj0O0umHTp42bXDhORZX7iGF+DiF93ukx4u3VN3R89Lbkj/8pCBtHtf1kpIVnjeje48pWwd225RFSM9OmNDVXTlhwuL1fVzbcaNOd/0CTRcCpD98teXwMT9sNDf4TAEhzZ04sZu7auLEpQ2OooB0Q48ep7u2PXqMnHZFwQnFHZp/dWzD2yKDFN7jtEdca9czfYWdu3f92jGtWzQ9ZWKg2UK51m7ZRS2anvFcCBMFhFRcd2rua3AUBaSz6u69z7Rpfb7RqPlpd+58W1SQwnuc9ghIIcfV3zKu/lYByQYkGZBUQLIBSQYkFZBsQJIBSQUkG5BkQFIByQYkGZBUQLIBSQYkFZBsQJIBSQUkG5BkQFIByQYkGZBUQLIBSQYkFZBsQJIBSQUkG5BkQFIByQYkGZBUQLIBSQYkFZBsQJIBSQUkG5BkQFIByQYkGZBUQLIBSQYkFZBsQJIBSQUkG5BkQFIByQYkGZBUQLIBSQYkFZBsQJIBSQUkG5BkQFIByQYkGZBUQLIBSQYkFZBsQJIBSQUkG5BkQFIByQYkGZBUQLIBSQYkFZBsQJIBSQUkG5BkQFIByQYkGZBUQLIBSQYkFZBsQJIBSQUkG5BkQFIByQYkGZBUQLIBSQYkFZBsQJIBSQUkG5BkQFIByQYkGZBUQLIBSQYkFZBsQJIBSQUkG5BkQFIByQYkGZBUQLIBSQYkFZBsQJIBSQUkG5BkQFIByQYkGZBUQLIBSQYkFZBsQJIBSQUkG5BkQFIByQYkGZBUQLIBSQYkFZBsQJIBSQUkG5BkQFIByQYkGZBUQLIBSQYkFZBsQJIBSQUkG5BkQFIByQYkGZBUQLIBSQYkFZBsQJIBSQUkG5BkQFIByQYkGZBUQLIBSQYkFZBsQJIBSRUTSB2LY9SJfWJUm469Y1TLs2NUIurHpmEXimd2jiFNq4xRvaP+c79ht77zWYwatSlGXV8eo24Xz2wgxSQgyYBkA5IKSDIg2YCkApIMSDYgqYAkA5INSCogyYBkA5IKSDIg2YCkApIMSDYgqYAkA5INSCogyYBkA5IKSDIg2YCkApIMSDYgqYAkA5INSCogyYBkA5IKSDIg2YCkApIMSDYgqYAkA5INSCogyYBkA5IKSDIg2YCkApIMSDYgqYAkA5INSCogyYBkA5IKSDIg2YCkApIMSDYgqYAkA5INSCogyYBkA5IKSDIg2YCkApIMSDYgqYAkA5INSCogyYBkA5IKSDIg2YCkApIMSDYgqYAkA5INSCogyYBkA5IKSDIg2YCkApIMSDYgqYAkA5INSCogyYBkA5IKSDIg2YCkApIMSDYgqYAkA5INSCogyYBkA5IKSDIg2YCkApIMSDYgqYAkA5INSCogyYBkA5IKSDIg2YCkApIMSDYgqYAkA5INSCogyYBkA5IKSDIg2YCkApIMSDYgqYAkA5INSCogyYBkA5IKSDIg2YCkApIMSDYgqYAkA5INSCogyYBkA5IKSDIg2YCkApIMSDYgqYAkA5INSCogyYBkA5IKSDIg2YCkApIMSDYgqYAkA5INSCogyYBkA5IKSDIg2YCkApIMSLagkBa1dnNSBwvaHdDke49FCym5mGd2PYoW0pz/PrjJSZM+DT5RcEh/cnXNjBbSgsw6JpSXzzq7eeOT7oo1pKIlOw1rEu9nBdK4ZkekIS1pcdzYSecUzIwSkr+YZ3Y5ihbSrMKTx0443Q2OA6R1d6UqKng9WkiLh6Y6v2BW+Zz9j7p56GkFE+MKafnHBlLtB1uzAemFJmPuTkPq2Gx5ZeW677SMENLzTUZPTvOpP4oYUsuj13322cbjD40DpHSrDy8OPkkIb+0WHtqxvPyCpi+Vly894RtxhXTbiwaSLBikJQsq05DWNyvyx6Pcq9FBWjx/Yx2f+qNoIVXe8YS/6+HWxQbSZQd/FHySECB1PXB++bKm5/uHg9wT8YQ0pH2n672iV0Z0Kl7geRXDu3cdsSFrb+0q6yAtckP9wRw3OTpIyer5xAJSuk9P+0bwSUKC9Gbh2BBmCQ5pduHN5eVPuwH+8XQ3Ip6QvH7+K9I1H/7zsS7bvCtLt/1jTEkdpPXPJKvKBqTn3F3+4A03DEg7tWH5y50bzQw+T0iQOhz+lxBmCQ6p3aGLyssfcMP846fc1XGG9LTnbUxUeFu/9LzFHWrTkBa0SfZ2NiA96ab6g2XuRiDt1DPOHfWbEOYJB9KbhXeGMU1gSLMLb0xup7nb/MGz7vI4Q1rseZsTH3vLhxQXd0vUZP8VaZI/KHPDgbRTH/1qaseCgcHnCQfS5Y1XhzFNYEjdGi9Mbh90Q/3BU+6aOENakoK0odPsam9pBtJuCgnSEneLP3gq/cIEpJ0a5F4NPEcokCqP+EkY0wSG9NbXz/R3c1x/f3dP+oUp3pDKimo879HsQ9rQ4uf+YIgrA1J9fxz3O3/3azc5HpBecpPCmCYwpBnpl6Jl+5/n7wa4p2IKqf/Df89AWplY8a+FgxNV2YZUeWmTdyor1x777UBz7WmQPio8syq5+6V7Jh6Qhrvgv4z1CwrpGjcrte/Q+Pny8kX/dUKgybIIaW7nPhlI3ozuPaZsHdhtQ3YgzZ00qZvrP2nSssr3Dj566B0/aDQnQkhzJ07s5q6aOHFpg6NoIX12nfvhqImdCr5fFQ9I3d2aMKYJDCnhFqb28w48csCgk/edHldI/5eCQepVd9nU9MrKRRe2aHraM4FmCwipuG4x9zU4ihjSp5NObrb/t66uCD5TKJAuKAxjluCQ/ruw7uDpc/Zvcup9wSbbIyCFHFd/y7j6WwUkG5BkQFIByQYkGZBUQLIBSQYkFZBsQJIBSQUkG5BkQFIByQYkGZBUQLIBSQYkFZBsQJIBSQUkG5BkQFIByQYkGZBUQLIBSQYkFZBsQJIBSQUkG5BkQFIByQYkGZBUQLIBSQYkFZBsQJIBSQUkG5BkQFIByQYkGZBUQLIBSQYkFZBsQJIBSQUkG5BkQFIByQYkGZBUQLIBSQYkFZBsQJIBSQUkG5BkQFIByQYkGZBUQLIBSQYkFZBsQJIBSQUkG5BkQFIByQYkGZBUQLIBSQYkFZBsQJIBSQUkG5BkQFIByQYkGZBUQLIBSQYkFZBsQJIBSQUkG5BkQFIByQYkGZBUQLIBSQYkFZBsQJIBSQUkG5BkQFIByQYkGZBUQLIBSQYkFZBsQJIBSQUkG5BkQFIByQYkGZBUQLIBSQYkFZBsQJIBSQUkG5BkQFIByQYkGZBUQLIBSQYkFZBsQJIBSQUkG5BkQFIByQYkGZBUQLIBSQYkFZBsQJIBSQUkG5BkQFIByQYkGZBUQLIBSQYkFZBsQJIBSQUkG5BkQFIByQYkGZBUQLIBSQYkFZBsQJIBSQUkG5BkQFIByQYkGZBUQLIBSQYkVUwgzXg8Rg2MegENGzY16hU07NqoF9CwWD1O08QzO8eQiPbMgEQUQkAiCiEgEYUQkIhCCEhEIQQkohACElEIAYkohIBEFEJAIgohIBGFEJCIQghIRCGUl5CmTY56BQ2ad+emqJdQ34d3Lo16CfV9eeesqJfQoBl3ZnX6vISUaBf1Chp0R5uPo15Cfa+2eTzqJdS3tc3VUS+hQb2/n9XpgRQ0IKmAFPeApAKSDEg2IKmAJAMSUfwDElEIAYkohPIDUtGS1K4m8X7ECzFL2JSoyPmqYnAaTDWJd6Jegq7u6ZMpK+cvryDVfrA14oWYJSQh5XxVMTgNpthCWv6xgZSV85dXkGJYElLUS4hFsYV024u5efrEHNKnd15cfO+XXtErIzoVL/Bfk2sTC0f07zvf8zaP79Vl8CrPe+2qzsX3Vu8YZruGS1g9qMvVC9Nv7SqGd+86YoO340vZXkPmDqsTLw/u13epZxaQ69PjQ6oZNrLmr+N7dS750Nue+F2/yTvuNadnZ+eGtO90febpk1nH3vjW7oaxm9cPmO4VXfPhPx/rss0/A0UDt3ivdNnmDRr/RfXjPas3tn9/+8brZmeGWV9QgyXU9ivdVjUkDenK0m3/GFPi7Vhd1tdQd4c1iev+6r3aYYtZQK5Pjw+ptORLb9CtW758uOvfkutY9c8d95rTs7NL/fxXpPTTp/6k7XWQVic2JjflXtHTnrcx/ZQtmuu/n/pkVWJz8s1ut7JVidWet93LDLO+ogZL+KO/uCXpVW390vMWd6jNfCn7a6i7w5rEc8l//a6v7LqAnJ+eJKQn+n+RfMDWel71xQu8oie9+nvN6dnZpRSk9NOn/qTtdZDebF+b2hctTr5ZSXycehanD8sSqWbX3tOhZNZ6LzPM+ooaLqH9ds/7JA1p+ZDi4m6JmsyXsr+GujusSSxL3nDVrF0XkPPTU5MYm/hD5gHr/xuvKIl2x73m9OzsUgpS3f3uOGl7HaRF/nPVS/+0mIGUPlyayLxP2TRvZIey+mGWa7CE+f6TZk0K0oZOs6u9pf5TZUluIGXusCaRfI54V/x61wXk/PTUJPqNGFhTB+mqJ1LryNxrbs/OLvV7ccfTp/6k7XWQ1vifiX30wm4grU2sTH59o1ezJbmbPjgzzHoNlrA8Uen/qetDKiuq8bxHcwgpc4c1ieS7lurOr+26gJyfnppE+dY+jyQfsOQbt22d56fWkbnX3J6dXWoAqf6k7XWQvEEjKtddd+9uIHlDS6pqXuzy+at9Pq7dPGRKZpj1BTVYQnWP0q3rbk5BWplY8a+FgxNVOYOUucOaxICK6lkd/2YWkOvT43/YsKLDu17JyC+23dfzH+lPnOvuNbdnZ5f6P/z3zP3Wn7S9D9KWO7r0nLZtd5A2j+t6SckKr3ZWn4697v57Zpj1Gi7ho+s7X/1O4s/+TTO695iydWC3TbmClLnDDYkXb+rcr9wzC8j16Un9Hunx4i1Vd/S89LZ1db+6ydxrTs/OLs3t3GfHA7bjpO19kGg3NfgTNba/B93rAlLetf0j/zPtdECKS0DKuxZ2GFWbOQZSXAISUQgBiSiEgEQUQkAiCiEgEYUQkPK3X7pMp+32622Pzu169uqAlL+9PnXq1Gtd5+TWXNb9nv+4AimHASm/e92V7u7mKUDKcUDK7+ognXn28984w2vd2j8u+qp3QfLtXhuv7XFrLmze/JLsX8lLQMr36iCdd/I373mhHtKfilz5h17blq1HP3tjwS+iXeFeEpDyuzpIbd2c5HYHJK+f23Hjj74W4fL2noCU32UgNf6XZyE19a/J61UY4fL2noCU32UgHeFvd4V0tD/sx0OcizjL+V0G0tH+FkjRxVnO73aCdOpJ/vY0IEUQZzm/2wnSeYckfyja1CwJ6TL3P0DKaZzl/G4nSJPdmMp3f/ydJKQR7rangZTLOMv53U6Qqm84sknr5we08Ly/nNqoFZByGWeZKISARBRCQCIKISARhRCQiEIISEQhBCSiEAISUQgBiSiEgEQUQkAiCiEgEYXQ/wMhANIDIZLX1QAAAABJRU5ErkJggg==" }, "metadata": { "image/png": { "width": 420, "height": 420 } } } ], "metadata": { "colab": { "base_uri": "https://localhost:8080/", "height": 436 }, "id": "HsAtwukyLsvt", "outputId": "3032a224-a2c8-4270-b4f2-7bb620317400" } }, { "cell_type": "markdown", "source": [ "Tumedamad ruudud segadusmaatriksi graafikul viitavad suurele juhtumite arvule, ja loodetavasti näete tumedamate ruutude diagonaaljoont, mis näitab juhtumeid, kus ennustatud ja tegelik märgis on sama.\n", "\n", "Arvutame nüüd segadusmaatriksi kokkuvõtlikud statistilised näitajad.\n" ], "metadata": { "id": "oOJC87dkLwPr" } }, { "cell_type": "code", "execution_count": 12, "source": [ "# Summary stats for confusion matrix\n", "conf_mat(data = results, truth = cuisine, estimate = .pred_class) %>% \n", "summary()" ], "outputs": [ { "output_type": "display_data", "data": { "text/plain": [ " .metric .estimator .estimate\n", "1 accuracy multiclass 0.7880435\n", "2 kap multiclass 0.7276583\n", "3 sens macro 0.7780927\n", "4 spec macro 0.9477598\n", "5 ppv macro 0.7585583\n", "6 npv macro 0.9460080\n", "7 mcc multiclass 0.7292724\n", "8 j_index macro 0.7258524\n", "9 bal_accuracy macro 0.8629262\n", "10 detection_prevalence macro 0.2000000\n", "11 precision macro 0.7585583\n", "12 recall macro 0.7780927\n", "13 f_meas macro 0.7641862" ], "text/markdown": [ "\n", "A tibble: 13 × 3\n", "\n", "| .metric <chr> | .estimator <chr> | .estimate <dbl> |\n", "|---|---|---|\n", "| accuracy | multiclass | 0.7880435 |\n", "| kap | multiclass | 0.7276583 |\n", "| sens | macro | 0.7780927 |\n", "| spec | macro | 0.9477598 |\n", "| ppv | macro | 0.7585583 |\n", "| npv | macro | 0.9460080 |\n", "| mcc | multiclass | 0.7292724 |\n", "| j_index | macro | 0.7258524 |\n", "| bal_accuracy | macro | 0.8629262 |\n", "| detection_prevalence | macro | 0.2000000 |\n", "| precision | macro | 0.7585583 |\n", "| recall | macro | 0.7780927 |\n", "| f_meas | macro | 0.7641862 |\n", "\n" ], "text/latex": [ "A tibble: 13 × 3\n", "\\begin{tabular}{lll}\n", " .metric & .estimator & .estimate\\\\\n", " & & \\\\\n", "\\hline\n", "\t accuracy & multiclass & 0.7880435\\\\\n", "\t kap & multiclass & 0.7276583\\\\\n", "\t sens & macro & 0.7780927\\\\\n", "\t spec & macro & 0.9477598\\\\\n", "\t ppv & macro & 0.7585583\\\\\n", "\t npv & macro & 0.9460080\\\\\n", "\t mcc & multiclass & 0.7292724\\\\\n", "\t j\\_index & macro & 0.7258524\\\\\n", "\t bal\\_accuracy & macro & 0.8629262\\\\\n", "\t detection\\_prevalence & macro & 0.2000000\\\\\n", "\t precision & macro & 0.7585583\\\\\n", "\t recall & macro & 0.7780927\\\\\n", "\t f\\_meas & macro & 0.7641862\\\\\n", "\\end{tabular}\n" ], "text/html": [ "\n", "\n", "\n", "\t\n", "\t\n", "\n", "\n", "\t\n", "\t\n", "\t\n", "\t\n", "\t\n", "\t\n", "\t\n", "\t\n", "\t\n", "\t\n", "\t\n", "\t\n", "\t\n", "\n", "
A tibble: 13 × 3
.metric.estimator.estimate
<chr><chr><dbl>
accuracy multiclass0.7880435
kap multiclass0.7276583
sens macro 0.7780927
spec macro 0.9477598
ppv macro 0.7585583
npv macro 0.9460080
mcc multiclass0.7292724
j_index macro 0.7258524
bal_accuracy macro 0.8629262
detection_prevalencemacro 0.2000000
precision macro 0.7585583
recall macro 0.7780927
f_meas macro 0.7641862
\n" ] }, "metadata": {} } ], "metadata": { "colab": { "base_uri": "https://localhost:8080/", "height": 494 }, "id": "OYqetUyzL5Wz", "outputId": "6a84d65e-113d-4281-dfc1-16e8b70f37e6" } }, { "cell_type": "markdown", "source": [ "Kui keskendume mõnedele mõõdikutele, nagu täpsus, tundlikkus, ppv, siis pole algus üldse halb 🥳!\n", "\n", "## 4. Sügavamale kaevumine\n", "\n", "Küsime ühe peene küsimuse: Milliseid kriteeriume kasutatakse, et otsustada konkreetse köögi tüübi kui ennustatud tulemuse üle?\n", "\n", "Statistilised masinõppe algoritmid, nagu logistiline regressioon, põhinevad `tõenäosusel`; seega ennustab klassifikaator tegelikult tõenäosusjaotust võimalike tulemuste hulgas. Klass, millel on kõrgeim tõenäosus, valitakse kõige tõenäolisemaks tulemuseks antud vaatluste põhjal.\n", "\n", "Vaatame seda tegevuses, tehes nii kindlaid klassiprognoose kui ka tõenäosusi.\n" ], "metadata": { "id": "43t7vz8vMJtW" } }, { "cell_type": "code", "execution_count": 13, "source": [ "# Make hard class prediction and probabilities\n", "results_prob <- cuisines_test %>%\n", " select(cuisine) %>% \n", " bind_cols(mr_fit %>% predict(new_data = cuisines_test)) %>% \n", " bind_cols(mr_fit %>% predict(new_data = cuisines_test, type = \"prob\"))\n", "\n", "# Print out results\n", "results_prob %>% \n", " slice_head(n = 5)" ], "outputs": [ { "output_type": "display_data", "data": { "text/plain": [ " cuisine .pred_class .pred_chinese .pred_indian .pred_japanese .pred_korean\n", "1 indian thai 1.551259e-03 0.4587877 5.988039e-04 2.428503e-04\n", "2 indian indian 2.637133e-05 0.9999488 6.648651e-07 2.259993e-05\n", "3 indian indian 1.049433e-03 0.9909982 1.060937e-03 1.644947e-05\n", "4 indian indian 6.237482e-02 0.4763035 9.136702e-02 3.660913e-01\n", "5 indian indian 1.431745e-02 0.9418551 2.945239e-02 8.721782e-03\n", " .pred_thai \n", "1 5.388194e-01\n", "2 1.577948e-06\n", "3 6.874989e-03\n", "4 3.863391e-03\n", "5 5.653283e-03" ], "text/markdown": [ "\n", "A tibble: 5 × 7\n", "\n", "| cuisine <fct> | .pred_class <fct> | .pred_chinese <dbl> | .pred_indian <dbl> | .pred_japanese <dbl> | .pred_korean <dbl> | .pred_thai <dbl> |\n", "|---|---|---|---|---|---|---|\n", "| indian | thai | 1.551259e-03 | 0.4587877 | 5.988039e-04 | 2.428503e-04 | 5.388194e-01 |\n", "| indian | indian | 2.637133e-05 | 0.9999488 | 6.648651e-07 | 2.259993e-05 | 1.577948e-06 |\n", "| indian | indian | 1.049433e-03 | 0.9909982 | 1.060937e-03 | 1.644947e-05 | 6.874989e-03 |\n", "| indian | indian | 6.237482e-02 | 0.4763035 | 9.136702e-02 | 3.660913e-01 | 3.863391e-03 |\n", "| indian | indian | 1.431745e-02 | 0.9418551 | 2.945239e-02 | 8.721782e-03 | 5.653283e-03 |\n", "\n" ], "text/latex": [ "A tibble: 5 × 7\n", "\\begin{tabular}{lllllll}\n", " cuisine & .pred\\_class & .pred\\_chinese & .pred\\_indian & .pred\\_japanese & .pred\\_korean & .pred\\_thai\\\\\n", " & & & & & & \\\\\n", "\\hline\n", "\t indian & thai & 1.551259e-03 & 0.4587877 & 5.988039e-04 & 2.428503e-04 & 5.388194e-01\\\\\n", "\t indian & indian & 2.637133e-05 & 0.9999488 & 6.648651e-07 & 2.259993e-05 & 1.577948e-06\\\\\n", "\t indian & indian & 1.049433e-03 & 0.9909982 & 1.060937e-03 & 1.644947e-05 & 6.874989e-03\\\\\n", "\t indian & indian & 6.237482e-02 & 0.4763035 & 9.136702e-02 & 3.660913e-01 & 3.863391e-03\\\\\n", "\t indian & indian & 1.431745e-02 & 0.9418551 & 2.945239e-02 & 8.721782e-03 & 5.653283e-03\\\\\n", "\\end{tabular}\n" ], "text/html": [ "\n", "\n", "\n", "\t\n", "\t\n", "\n", "\n", "\t\n", "\t\n", "\t\n", "\t\n", "\t\n", "\n", "
A tibble: 5 × 7
cuisine.pred_class.pred_chinese.pred_indian.pred_japanese.pred_korean.pred_thai
<fct><fct><dbl><dbl><dbl><dbl><dbl>
indianthai 1.551259e-030.45878775.988039e-042.428503e-045.388194e-01
indianindian2.637133e-050.99994886.648651e-072.259993e-051.577948e-06
indianindian1.049433e-030.99099821.060937e-031.644947e-056.874989e-03
indianindian6.237482e-020.47630359.136702e-023.660913e-013.863391e-03
indianindian1.431745e-020.94185512.945239e-028.721782e-035.653283e-03
\n" ] }, "metadata": {} } ], "metadata": { "colab": { "base_uri": "https://localhost:8080/", "height": 248 }, "id": "xdKNs-ZPMTJL", "outputId": "68f6ac5a-725a-4eff-9ea6-481fef00e008" } }, { "cell_type": "markdown", "source": [ "Palju parem!\n", "\n", "✅ Kas saaksid selgitada, miks mudel on üsna kindel, et esimene vaatlus on Tai köök?\n", "\n", "## **🚀Väljakutse**\n", "\n", "Selles õppetükis kasutasid sa puhastatud andmeid, et luua masinõppemudel, mis suudab ennustada rahvuskööki koostisosade põhjal. Võta aega, et tutvuda [paljude võimalustega](https://www.tidymodels.org/find/parsnip/#models), mida Tidymodels pakub andmete klassifitseerimiseks, ja [muude viisidega](https://parsnip.tidymodels.org/articles/articles/Examples.html#multinom_reg-models), kuidas sobitada multinomiaalset regressiooni.\n", "\n", "#### TÄNUD:\n", "\n", "[`Allison Horst`](https://twitter.com/allison_horst/) suurepäraste illustratsioonide loomise eest, mis muudavad R-i kasutamise sõbralikumaks ja kaasahaaravamaks. Leia rohkem illustratsioone tema [galeriist](https://www.google.com/url?q=https://github.com/allisonhorst/stats-illustrations&sa=D&source=editors&ust=1626380772530000&usg=AOvVaw3zcfyCizFQZpkSLzxiiQEM).\n", "\n", "[Cassie Breviu](https://www.twitter.com/cassieview) ja [Jen Looper](https://www.twitter.com/jenlooper) selle mooduli algse Python-versiooni loomise eest ♥️\n", "\n", "
\n", "Oleksin lisanud mõned naljad, aga ma ei saa toidupunidest aru 😅.\n", "\n", "
\n", "\n", "Head õppimist,\n", "\n", "[Eric](https://twitter.com/ericntay), Gold Microsoft Learn Student Ambassador.\n" ], "metadata": { "id": "2tWVHMeLMYdM" } }, { "cell_type": "markdown", "metadata": {}, "source": [ "\n---\n\n**Lahtiütlus**: \nSee dokument on tõlgitud AI tõlketeenuse [Co-op Translator](https://github.com/Azure/co-op-translator) abil. Kuigi püüame tagada täpsust, palume arvestada, et automaatsed tõlked võivad sisaldada vigu või ebatäpsusi. Algne dokument selle algses keeles tuleks pidada autoriteetseks allikaks. Olulise teabe puhul soovitame kasutada professionaalset inimtõlget. Me ei vastuta selle tõlke kasutamisest tulenevate arusaamatuste või valesti tõlgenduste eest.\n" ] } ] }