{ "cells": [ { "cell_type": "code", "execution_count": null, "id": "60484ff2", "metadata": {}, "outputs": [], "source": [ "import torch\n", "import pandas as pd\n", "import seaborn as sns\n", "import numpy as np\n", "import re\n", "import unicodedata\n", "\n", "import torch.nn as nn\n", "from torch.utils.data import Dataset, WeightedRandomSampler\n", "from transformers import AutoTokenizer, AutoModel\n", "\n", "from sklearn.metrics import accuracy_score, f1_score" ] }, { "cell_type": "markdown", "id": "b8bd2483", "metadata": {}, "source": [ "### Чтение данных" ] }, { "cell_type": "code", "execution_count": 2, "id": "8c3bc1a0", "metadata": {}, "outputs": [ { "data": { "text/html": [ "
\n", "\n", "\n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", "
TitleDescriptionPriceCapacityGrapeSecondary Grape VarietiesClosureCountryUnitCharacteristicsPer bottle / case / eachTypeABVRegionStyleVintageAppellation
0The Guv'nor, SpainWe asked some of our most prized winemakers wo...£9.99 per bottle75CLTempranilloNaNNatural CorkSpain10.5Vanilla, Blackberry, Blackcurrantper bottleRedABV 14.00%NaNRich & JuicyNVNaN
1Bread & Butter 'Winemaker's Selection' Chardon...This really does what it says on the tin. It’s...£15.99 per bottle75CLChardonnayNaNNatural CorkUSA10.1Vanilla, Almond, Coconut, Green Apple, Peach, ...per bottleWhiteABV 13.50%CaliforniaRich & Toasty2021Napa Valley
2Oyster Bay Sauvignon Blanc 2022, MarlboroughOyster Bay has been an award-winning gold-stan...£12.49 per bottle75CLSauvignon BlancNaNScrewcapNew Zealand9.8Tropical Fruit, Gooseberry, Grapefruit, Grass,...per bottleWhiteABV 13.00%MarlboroughCrisp & Zesty2022NaN
3Louis Latour Mâcon-Lugny 2021/22We’ve sold this wine for thirty years – and fo...£17.99 per bottle75CLChardonnayNaNNatural CorkFrance10.1Peach, Apricot, Floral, Lemonper bottleWhiteABV 13.50%BurgundyRipe & Rounded2022Macon
4Bread & Butter 'Winemaker's Selection' Pinot N...Bread & Butter is that thing that you can coun...£15.99 per bottle75CLPinot NoirNaNNatural CorkUSA10.1Smoke, Black Cherry, Cedar, Raspberry, Red Fruitper bottleRedABV 13.50%CaliforniaSmooth & Mellow2021Napa Valley
......................................................
1285Vouvray Sec 'Expresion de Silex' 2020/21, LoireAlthough Denis Meunier is only in his late 20s...£11.99 per bottle75CLChenin BlancNaNNatural CorkFrance9.4Quince, Green Apple, Lemon, Pearper bottleWhiteABV 12.50%LoireAromatic & Floral2021Vouvray
1286Waimea Estates Pinot Noir Rosé 2021, NelsonWaimea’s vineyards are less than 1km from the ...£13.99 per bottle75CLPinot NoirNaNScrewcapNew Zealand9.8Floral, Raspberry, Red Cherry, Strawberryper bottleRoséABV 13.00%NelsonDelicate & Dry2021NaN
1287Wakefield 'Visionary' Cabernet Sauvignon 2010Made from the very best fruit nourished by 40-...£85.00 per bottle75CLCabernet SauvignonNaNScrewcapAustralia10.1Vanilla, Blackberry, Blackcurrant, Cedar, Euca...per bottleRedABV 13.50%South AustraliaSavoury & Full Bodied2010Clare Valley
1288Yalumba 'The Menzies' Cabernet Sauvignon 2015/...In 1961, Sir Robert Menzies said that Yalumba'...£37.99 per bottle75CLCabernet SauvignonNaNNatural CorkAustralia10.9Black Plum, Blackcurrant, Vanillaper bottleRedABV 14.50%South AustraliaSavoury & Full Bodied2016Coonawarra
1289Yalumba 'Virgilius' Viognier 2017, Eden ValleyYalumba's Virgilius is often regarded as Austr...£34.99 per bottle75CLViognierNaNScrewcapAustralia10.1Peach, Apricot, Floral, Honeysuckleper bottleWhiteABV 13.50%NaNAromatic & Floral2017Eden Valley
\n", "

1290 rows × 17 columns

\n", "
" ], "text/plain": [ " Title \\\n", "0 The Guv'nor, Spain \n", "1 Bread & Butter 'Winemaker's Selection' Chardon... \n", "2 Oyster Bay Sauvignon Blanc 2022, Marlborough \n", "3 Louis Latour Mâcon-Lugny 2021/22 \n", "4 Bread & Butter 'Winemaker's Selection' Pinot N... \n", "... ... \n", "1285 Vouvray Sec 'Expresion de Silex' 2020/21, Loire \n", "1286 Waimea Estates Pinot Noir Rosé 2021, Nelson \n", "1287 Wakefield 'Visionary' Cabernet Sauvignon 2010 \n", "1288 Yalumba 'The Menzies' Cabernet Sauvignon 2015/... \n", "1289 Yalumba 'Virgilius' Viognier 2017, Eden Valley \n", "\n", " Description Price \\\n", "0 We asked some of our most prized winemakers wo... £9.99 per bottle \n", "1 This really does what it says on the tin. It’s... £15.99 per bottle \n", "2 Oyster Bay has been an award-winning gold-stan... £12.49 per bottle \n", "3 We’ve sold this wine for thirty years – and fo... £17.99 per bottle \n", "4 Bread & Butter is that thing that you can coun... £15.99 per bottle \n", "... ... ... \n", "1285 Although Denis Meunier is only in his late 20s... £11.99 per bottle \n", "1286 Waimea’s vineyards are less than 1km from the ... £13.99 per bottle \n", "1287 Made from the very best fruit nourished by 40-... £85.00 per bottle \n", "1288 In 1961, Sir Robert Menzies said that Yalumba'... £37.99 per bottle \n", "1289 Yalumba's Virgilius is often regarded as Austr... £34.99 per bottle \n", "\n", " Capacity Grape Secondary Grape Varieties Closure \\\n", "0 75CL Tempranillo NaN Natural Cork \n", "1 75CL Chardonnay NaN Natural Cork \n", "2 75CL Sauvignon Blanc NaN Screwcap \n", "3 75CL Chardonnay NaN Natural Cork \n", "4 75CL Pinot Noir NaN Natural Cork \n", "... ... ... ... ... \n", "1285 75CL Chenin Blanc NaN Natural Cork \n", "1286 75CL Pinot Noir NaN Screwcap \n", "1287 75CL Cabernet Sauvignon NaN Screwcap \n", "1288 75CL Cabernet Sauvignon NaN Natural Cork \n", "1289 75CL Viognier NaN Screwcap \n", "\n", " Country Unit Characteristics \\\n", "0 Spain 10.5 Vanilla, Blackberry, Blackcurrant \n", "1 USA 10.1 Vanilla, Almond, Coconut, Green Apple, Peach, ... \n", "2 New Zealand 9.8 Tropical Fruit, Gooseberry, Grapefruit, Grass,... \n", "3 France 10.1 Peach, Apricot, Floral, Lemon \n", "4 USA 10.1 Smoke, Black Cherry, Cedar, Raspberry, Red Fruit \n", "... ... ... ... \n", "1285 France 9.4 Quince, Green Apple, Lemon, Pear \n", "1286 New Zealand 9.8 Floral, Raspberry, Red Cherry, Strawberry \n", "1287 Australia 10.1 Vanilla, Blackberry, Blackcurrant, Cedar, Euca... \n", "1288 Australia 10.9 Black Plum, Blackcurrant, Vanilla \n", "1289 Australia 10.1 Peach, Apricot, Floral, Honeysuckle \n", "\n", " Per bottle / case / each Type ABV Region \\\n", "0 per bottle Red ABV 14.00% NaN \n", "1 per bottle White ABV 13.50% California \n", "2 per bottle White ABV 13.00% Marlborough \n", "3 per bottle White ABV 13.50% Burgundy \n", "4 per bottle Red ABV 13.50% California \n", "... ... ... ... ... \n", "1285 per bottle White ABV 12.50% Loire \n", "1286 per bottle Rosé ABV 13.00% Nelson \n", "1287 per bottle Red ABV 13.50% South Australia \n", "1288 per bottle Red ABV 14.50% South Australia \n", "1289 per bottle White ABV 13.50% NaN \n", "\n", " Style Vintage Appellation \n", "0 Rich & Juicy NV NaN \n", "1 Rich & Toasty 2021 Napa Valley \n", "2 Crisp & Zesty 2022 NaN \n", "3 Ripe & Rounded 2022 Macon \n", "4 Smooth & Mellow 2021 Napa Valley \n", "... ... ... ... \n", "1285 Aromatic & Floral 2021 Vouvray \n", "1286 Delicate & Dry 2021 NaN \n", "1287 Savoury & Full Bodied 2010 Clare Valley \n", "1288 Savoury & Full Bodied 2016 Coonawarra \n", "1289 Aromatic & Floral 2017 Eden Valley \n", "\n", "[1290 rows x 17 columns]" ] }, "execution_count": 2, "metadata": {}, "output_type": "execute_result" } ], "source": [ "df_full = pd.read_csv(\"data/WineDataset.csv\")\n", "df_full " ] }, { "cell_type": "code", "execution_count": 3, "id": "6891866a", "metadata": {}, "outputs": [ { "name": "stdout", "output_type": "stream", "text": [ "\n", "RangeIndex: 1290 entries, 0 to 1289\n", "Data columns (total 17 columns):\n", " # Column Non-Null Count Dtype \n", "--- ------ -------------- ----- \n", " 0 Title 1290 non-null object \n", " 1 Description 1286 non-null object \n", " 2 Price 1290 non-null object \n", " 3 Capacity 1290 non-null object \n", " 4 Grape 1275 non-null object \n", " 5 Secondary Grape Varieties 488 non-null object \n", " 6 Closure 1279 non-null object \n", " 7 Country 1284 non-null object \n", " 8 Unit 1281 non-null float64\n", " 9 Characteristics 1253 non-null object \n", " 10 Per bottle / case / each 1290 non-null object \n", " 11 Type 1285 non-null object \n", " 12 ABV 1281 non-null object \n", " 13 Region 1124 non-null object \n", " 14 Style 1212 non-null object \n", " 15 Vintage 1283 non-null object \n", " 16 Appellation 644 non-null object \n", "dtypes: float64(1), object(16)\n", "memory usage: 171.5+ KB\n" ] } ], "source": [ "df_full.info()" ] }, { "cell_type": "code", "execution_count": 4, "id": "9104e088", "metadata": {}, "outputs": [ { "data": { "text/html": [ "
\n", "\n", "\n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", "
DescriptionType
0We asked some of our most prized winemakers wo...Red
1This really does what it says on the tin. It’s...White
2Oyster Bay has been an award-winning gold-stan...White
3We’ve sold this wine for thirty years – and fo...White
4Bread & Butter is that thing that you can coun...Red
.........
1285Although Denis Meunier is only in his late 20s...White
1286Waimea’s vineyards are less than 1km from the ...Rosé
1287Made from the very best fruit nourished by 40-...Red
1288In 1961, Sir Robert Menzies said that Yalumba'...Red
1289Yalumba's Virgilius is often regarded as Austr...White
\n", "

1290 rows × 2 columns

\n", "
" ], "text/plain": [ " Description Type\n", "0 We asked some of our most prized winemakers wo... Red\n", "1 This really does what it says on the tin. It’s... White\n", "2 Oyster Bay has been an award-winning gold-stan... White\n", "3 We’ve sold this wine for thirty years – and fo... White\n", "4 Bread & Butter is that thing that you can coun... Red\n", "... ... ...\n", "1285 Although Denis Meunier is only in his late 20s... White\n", "1286 Waimea’s vineyards are less than 1km from the ... Rosé\n", "1287 Made from the very best fruit nourished by 40-... Red\n", "1288 In 1961, Sir Robert Menzies said that Yalumba'... Red\n", "1289 Yalumba's Virgilius is often regarded as Austr... White\n", "\n", "[1290 rows x 2 columns]" ] }, "execution_count": 4, "metadata": {}, "output_type": "execute_result" } ], "source": [ "df = df_full[['Description', 'Type']]\n", "df" ] }, { "cell_type": "code", "execution_count": 5, "id": "bcd9a741", "metadata": {}, "outputs": [ { "name": "stderr", "output_type": "stream", "text": [ "/var/folders/l1/t3rml9dj3c5_2blf8jgn4xg00000gn/T/ipykernel_9698/3607499803.py:1: FutureWarning: Series.__getitem__ treating keys as positions is deprecated. In a future version, integer keys will always be treated as labels (consistent with DataFrame behavior). To access a value by position, use `ser.iloc[pos]`\n", " df.iloc[296][0]\n" ] }, { "data": { "text/plain": [ "'Just because it’s inexpensive, doesn’t mean it’s not great quality. This wine is actually made for us by one of the most prestigious winemaking families in Chile, the Arestis. They know how to get the very best peachy, citrussy flavours out of their grapes, while still keeping prices wallet-friendly. The Central Valley, where this wine is from, is cool but has loads of sunshine – so this is more like a ripe white Bordeaux than a full, tropically fruity Kiwi Sauvignon. The fact is, Chilean wines are just incredible value – a wine of this quality from another country could cost far more. In our book, that makes it the ideal party wine. It’s got a real citrussy zing and great tropical flavours. It’s incredibly refreshing, so your guests will be happy to drink it alone, or with finger foods. It’s hard working bottle which won’t disappoint.'" ] }, "execution_count": 5, "metadata": {}, "output_type": "execute_result" } ], "source": [ "df.iloc[296][0]" ] }, { "cell_type": "code", "execution_count": 6, "id": "850300d7", "metadata": {}, "outputs": [ { "data": { "text/plain": [ "array(['Red', 'White', 'Brown', 'Rosé', nan, 'Mixed', 'Orange', 'Tawny'],\n", " dtype=object)" ] }, "execution_count": 6, "metadata": {}, "output_type": "execute_result" } ], "source": [ "df['Type'].unique()" ] }, { "cell_type": "code", "execution_count": 7, "id": "2a6fbb28", "metadata": {}, "outputs": [ { "data": { "text/plain": [ "" ] }, "execution_count": 7, "metadata": {}, "output_type": "execute_result" }, { "data": { "image/png": "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", "text/plain": [ "
" ] }, "metadata": {}, "output_type": "display_data" } ], "source": [ "sns.histplot(x='Type', data=df, hue='Type', multiple='stack')" ] }, { "cell_type": "markdown", "id": "2fc3a46b", "metadata": {}, "source": [ "### Очистка данных" ] }, { "cell_type": "code", "execution_count": 8, "id": "ae317d02", "metadata": {}, "outputs": [ { "data": { "text/plain": [ "Description 4\n", "Type 5\n", "dtype: int64" ] }, "execution_count": 8, "metadata": {}, "output_type": "execute_result" } ], "source": [ "df.isnull().sum()" ] }, { "cell_type": "code", "execution_count": 9, "id": "fef53c5a", "metadata": {}, "outputs": [ { "name": "stdout", "output_type": "stream", "text": [ "[ 252 295 350 631 638 667 960 1187 1250]\n", "[1 0 1 1 0 1 1 0 0]\n" ] } ], "source": [ "null_rows, null_cols = np.where(df.isnull())\n", "print(null_rows)\n", "print(null_cols)" ] }, { "cell_type": "code", "execution_count": 10, "id": "ad14c21e", "metadata": {}, "outputs": [ { "name": "stderr", "output_type": "stream", "text": [ "/var/folders/l1/t3rml9dj3c5_2blf8jgn4xg00000gn/T/ipykernel_9698/1850752272.py:1: FutureWarning: Series.__getitem__ treating keys as positions is deprecated. In a future version, integer keys will always be treated as labels (consistent with DataFrame behavior). To access a value by position, use `ser.iloc[pos]`\n", " df.iloc[667][1]\n" ] }, { "data": { "text/plain": [ "nan" ] }, "execution_count": 10, "metadata": {}, "output_type": "execute_result" } ], "source": [ "df.iloc[667][1]" ] }, { "cell_type": "code", "execution_count": 11, "id": "c8bdedb8", "metadata": {}, "outputs": [], "source": [ "def fill_missing_type(df, df_full, row_idx, value):\n", " if row_idx < 0 or row_idx >= len(df):\n", " print(f\"Индекс {row_idx} вне диапазона\")\n", " return\n", " \n", " after = df_full.iloc[row_idx]['Type']\n", " print(f\"Было: '{after}' → Стало: '{value}'\")\n", " \n", " df.loc[df.index[row_idx], 'Type'] = value" ] }, { "cell_type": "code", "execution_count": 12, "id": "da5957f5", "metadata": {}, "outputs": [ { "name": "stdout", "output_type": "stream", "text": [ "Было: 'nan' → Стало: 'White'\n" ] }, { "name": "stderr", "output_type": "stream", "text": [ "/var/folders/l1/t3rml9dj3c5_2blf8jgn4xg00000gn/T/ipykernel_9698/2357629074.py:9: SettingWithCopyWarning: \n", "A value is trying to be set on a copy of a slice from a DataFrame\n", "\n", "See the caveats in the documentation: https://pandas.pydata.org/pandas-docs/stable/user_guide/indexing.html#returning-a-view-versus-a-copy\n", " df.loc[df.index[row_idx], 'Type'] = value\n" ] } ], "source": [ "fill_missing_type (df, df_full, 667, 'White')" ] }, { "cell_type": "code", "execution_count": 13, "id": "dc0514b2", "metadata": {}, "outputs": [ { "data": { "text/plain": [ "Description 4\n", "Type 4\n", "dtype: int64" ] }, "execution_count": 13, "metadata": {}, "output_type": "execute_result" } ], "source": [ "df.isna().sum()" ] }, { "cell_type": "code", "execution_count": 14, "id": "ff877ce3", "metadata": {}, "outputs": [ { "name": "stderr", "output_type": "stream", "text": [ "/var/folders/l1/t3rml9dj3c5_2blf8jgn4xg00000gn/T/ipykernel_9698/1379821321.py:1: SettingWithCopyWarning: \n", "A value is trying to be set on a copy of a slice from a DataFrame\n", "\n", "See the caveats in the documentation: https://pandas.pydata.org/pandas-docs/stable/user_guide/indexing.html#returning-a-view-versus-a-copy\n", " df.dropna(inplace=True)\n" ] } ], "source": [ "df.dropna(inplace=True)" ] }, { "cell_type": "code", "execution_count": 15, "id": "60d5bfb4", "metadata": {}, "outputs": [ { "name": "stdout", "output_type": "stream", "text": [ "Type\n", "White 584\n", "Red 566\n", "Rosé 124\n", "Name: count, dtype: int64\n", " Description label\n", "659 the rhone valley was the birthplace of syrah a... 0\n", "660 this delicious rose was produced by union des ... 2\n", "661 white rioja fell out of fashion but its made a... 1\n", "662 the gentle rolling hills of the maconnais regi... 0\n", "663 the ardeche is a bit of an insider secret only... 1\n", "664 of all the sonoma valleys many prized grapegro... 0\n" ] }, { "name": "stderr", "output_type": "stream", "text": [ "/var/folders/l1/t3rml9dj3c5_2blf8jgn4xg00000gn/T/ipykernel_9698/766881011.py:11: SettingWithCopyWarning: \n", "A value is trying to be set on a copy of a slice from a DataFrame.\n", "Try using .loc[row_indexer,col_indexer] = value instead\n", "\n", "See the caveats in the documentation: https://pandas.pydata.org/pandas-docs/stable/user_guide/indexing.html#returning-a-view-versus-a-copy\n", " df['Description'] = df['Description'].apply(normalize_text)\n", "/var/folders/l1/t3rml9dj3c5_2blf8jgn4xg00000gn/T/ipykernel_9698/766881011.py:14: SettingWithCopyWarning: \n", "A value is trying to be set on a copy of a slice from a DataFrame.\n", "Try using .loc[row_indexer,col_indexer] = value instead\n", "\n", "See the caveats in the documentation: https://pandas.pydata.org/pandas-docs/stable/user_guide/indexing.html#returning-a-view-versus-a-copy\n", " df['label'] = df['Type'].map(label_map)\n" ] } ], "source": [ "valid_types = ['Red', 'White', 'Rosé']\n", "df = df[df['Type'].isin(valid_types)]\n", "\n", "def normalize_text(text):\n", " text = unicodedata.normalize('NFKD', text)\n", " text = text.encode('ascii', 'ignore').decode('utf-8')\n", " text = text.lower()\n", " text = re.sub(r'[^a-z\\s]', '', text)\n", " return text.strip()\n", "\n", "df['Description'] = df['Description'].apply(normalize_text)\n", "\n", "label_map = {'Red': 0, 'White': 1, 'Rosé': 2}\n", "df['label'] = df['Type'].map(label_map)\n", "\n", "print(df['Type'].value_counts())\n", "print(df[['Description', 'label']][649:655])" ] }, { "cell_type": "code", "execution_count": 16, "id": "d2e27ebc", "metadata": {}, "outputs": [ { "data": { "text/plain": [ "label\n", "1 584\n", "0 566\n", "2 124\n", "Name: count, dtype: int64" ] }, "execution_count": 16, "metadata": {}, "output_type": "execute_result" } ], "source": [ "df['label'].value_counts()" ] }, { "cell_type": "markdown", "id": "499618ce", "metadata": {}, "source": [ "### Написание классов и методов " ] }, { "cell_type": "code", "execution_count": 17, "id": "ae540262", "metadata": {}, "outputs": [], "source": [ "DOMAIN_WORDS = ['malbec','tannins','oak','champagne','citrus','sauvignon','pinot','merlot','barrique','fermentation']\n", "MODEL_NAME = 'bert-base-uncased'\n", "MAX_LEN = 128\n", "BATCH_SIZE = 16\n", "NUM_EPOCHS = 3\n", "LR = 2e-5\n", "NUM_CLASSES = 3" ] }, { "cell_type": "code", "execution_count": 18, "id": "2504f90d", "metadata": {}, "outputs": [], "source": [ "def extract_domain_features(text):\n", " text_lower = text.lower()\n", " return [1.0 if word in text_lower else 0.0 for word in DOMAIN_WORDS]" ] }, { "cell_type": "code", "execution_count": 19, "id": "51bb10c4", "metadata": {}, "outputs": [], "source": [ "class WineDataset(Dataset):\n", " def __init__(self, texts, labels, tokenizer):\n", " self.texts = texts\n", " self.labels = labels\n", " self.tokenizer = tokenizer\n", " self.domain_features = [extract_domain_features(t) for t in texts]\n", "\n", " def __len__(self):\n", " return len(self.texts)\n", "\n", " def __getitem__(self, idx):\n", " encoding = self.tokenizer(\n", " self.texts[idx],\n", " max_length=MAX_LEN,\n", " padding='max_length',\n", " truncation=True,\n", " return_tensors='pt'\n", " )\n", " return {\n", " 'input_ids': encoding['input_ids'].squeeze(0),\n", " 'attention_mask': encoding['attention_mask'].squeeze(0),\n", " 'domain_features': torch.tensor(self.domain_features[idx], dtype=torch.float32),\n", " 'label': torch.tensor(self.labels[idx], dtype=torch.long)\n", " }" ] }, { "cell_type": "code", "execution_count": 20, "id": "621caf5e", "metadata": {}, "outputs": [], "source": [ "class BertWineClassifier(nn.Module):\n", " def __init__(self, num_classes, domain_feature_dim):\n", " super().__init__()\n", " self.bert = AutoModel.from_pretrained(MODEL_NAME)\n", " hidden_size = self.bert.config.hidden_size\n", " self.classifier = nn.Sequential(\n", " nn.Linear(hidden_size + domain_feature_dim, 256),\n", " nn.BatchNorm1d(256),\n", " nn.ReLU(),\n", " nn.Dropout(0.3),\n", " nn.Linear(256, 64),\n", " nn.ReLU(),\n", " nn.Linear(64, num_classes)\n", " )\n", "\n", " def forward(self, input_ids, attention_mask, domain_features):\n", " outputs = self.bert(input_ids=input_ids, attention_mask=attention_mask)\n", " cls_embedding = outputs.last_hidden_state[:, 0, :]\n", " combined = torch.cat([cls_embedding, domain_features], dim=1)\n", " return self.classifier(combined)" ] }, { "cell_type": "code", "execution_count": 21, "id": "3b4762d3", "metadata": {}, "outputs": [], "source": [ "def make_weighted_sampler(labels):\n", " class_counts = np.bincount(labels)\n", " weights = 1.0 / class_counts\n", " sample_weights = torch.tensor([weights[l] for l in labels], dtype=torch.float32)\n", " return WeightedRandomSampler(sample_weights, num_samples=len(sample_weights), replacement=True)\n" ] }, { "cell_type": "code", "execution_count": 22, "id": "c10454ce", "metadata": {}, "outputs": [], "source": [ "def train_epoch(model, loader, optimizer, criterion, device):\n", " model.train()\n", " total_loss = 0\n", " for batch in loader:\n", " input_ids = batch['input_ids'].to(device)\n", " attention_mask = batch['attention_mask'].to(device)\n", " domain_features = batch['domain_features'].to(device)\n", " labels = batch['label'].to(device)\n", "\n", " optimizer.zero_grad()\n", " outputs = model(input_ids, attention_mask, domain_features)\n", " loss = criterion(outputs, labels)\n", " loss.backward()\n", " optimizer.step()\n", " total_loss += loss.item()\n", " return total_loss / len(loader)\n" ] }, { "cell_type": "code", "execution_count": 23, "id": "b09bf213", "metadata": {}, "outputs": [], "source": [ "def evaluate(model, loader, device):\n", " model.eval()\n", " all_preds = []\n", " all_labels = []\n", " with torch.no_grad():\n", " for batch in loader:\n", " input_ids = batch['input_ids'].to(device)\n", " attention_mask = batch['attention_mask'].to(device)\n", " domain_features = batch['domain_features'].to(device)\n", " labels = batch['label'].to(device)\n", "\n", " outputs = model(input_ids, attention_mask, domain_features)\n", " preds = outputs.argmax(dim=1).cpu().numpy()\n", " all_preds.extend(preds)\n", " all_labels.extend(labels.cpu().numpy())\n", "\n", " acc = accuracy_score(all_labels, all_preds)\n", " f1 = f1_score(all_labels, all_preds, average='macro')\n", " return acc, f1\n" ] }, { "cell_type": "markdown", "id": "7586badf", "metadata": {}, "source": [ "### Обучение" ] }, { "cell_type": "code", "execution_count": 24, "id": "d6b421e5", "metadata": {}, "outputs": [], "source": [ "device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')\n", "\n", "texts = df['Description'].tolist()\n", "labels = df['label'].tolist()\n", "\n", "X_train, X_test, y_train, y_test = train_test_split(\n", " texts, labels, test_size=0.2, random_state=42, stratify=labels\n", ")\n" ] }, { "cell_type": "code", "execution_count": 25, "id": "abf20bb1", "metadata": {}, "outputs": [ { "name": "stderr", "output_type": "stream", "text": [ "Downloading tokenizer_config.json: 100%|██████████| 48.0/48.0 [00:00<00:00, 101kB/s]\n", "Downloading config.json: 100%|██████████| 570/570 [00:00<00:00, 4.79MB/s]\n", "Downloading vocab.txt: 232kB [00:00, 1.39MB/s]\n", "Downloading tokenizer.json: 466kB [00:00, 2.61MB/s]\n" ] } ], "source": [ "tokenizer = AutoTokenizer.from_pretrained(MODEL_NAME)\n", "\n", "train_dataset = WineDataset(X_train, y_train, tokenizer)\n", "test_dataset = WineDataset(X_test, y_test, tokenizer)\n", "\n", "sampler = make_weighted_sampler(y_train)\n", "train_loader = DataLoader(train_dataset, batch_size=BATCH_SIZE, sampler=sampler)\n", "test_loader = DataLoader(test_dataset, batch_size=BATCH_SIZE, shuffle=False)\n" ] }, { "cell_type": "code", "execution_count": 26, "id": "92f527b6", "metadata": {}, "outputs": [], "source": [ "class_weights = compute_class_weight('balanced', classes=np.unique(y_train), y=y_train)\n", "class_weights_tensor = torch.tensor(class_weights, dtype=torch.float32).to(device)\n", "criterion = nn.CrossEntropyLoss(weight=class_weights_tensor)" ] }, { "cell_type": "code", "execution_count": 27, "id": "83c664a8", "metadata": {}, "outputs": [ { "name": "stderr", "output_type": "stream", "text": [ "/Users/affiliation/Data_science/Wine_guide/.venv/lib/python3.13/site-packages/transformers/utils/generic.py:309: FutureWarning: `torch.utils._pytree._register_pytree_node` is deprecated. Please use `torch.utils._pytree.register_pytree_node` instead.\n", " _torch_pytree._register_pytree_node(\n", "Downloading model.safetensors: 100%|██████████| 440M/440M [05:43<00:00, 1.28MB/s] \n" ] }, { "name": "stdout", "output_type": "stream", "text": [ "Epoch 1/3 | Loss: 0.8716 | Accuracy: 54.51% | Macro F1: 0.5303\n", "Epoch 2/3 | Loss: 0.4873 | Accuracy: 72.55% | Macro F1: 0.7405\n", "Epoch 3/3 | Loss: 0.3512 | Accuracy: 91.37% | Macro F1: 0.9256\n" ] } ], "source": [ "model = BertWineClassifier(num_classes=NUM_CLASSES, domain_feature_dim=len(DOMAIN_WORDS)).to(device)\n", "optimizer = torch.optim.AdamW(model.parameters(), lr=LR)\n", "\n", "for epoch in range(NUM_EPOCHS):\n", " loss = train_epoch(model, train_loader, optimizer, criterion, device)\n", " acc, f1 = evaluate(model, test_loader, device)\n", " print(f\"Epoch {epoch+1}/{NUM_EPOCHS} | Loss: {loss:.4f} | Accuracy: {acc:.2%} | Macro F1: {f1:.4f}\")" ] }, { "cell_type": "code", "execution_count": 29, "id": "a8500bc0", "metadata": {}, "outputs": [ { "name": "stdout", "output_type": "stream", "text": [ "Тип: Red\n", "Red: 45.01% | White: 25.88% | Rosé: 29.11%\n", "Тип: White\n", "Red: 25.39% | White: 45.21% | Rosé: 29.40%\n" ] } ], "source": [ "def predict(text):\n", " model.eval()\n", " label_map = {0: 'Red', 1: 'White', 2: 'Rosé'}\n", " encoding = tokenizer(text, max_length=MAX_LEN, padding='max_length', truncation=True, return_tensors='pt')\n", " domain = torch.tensor([extract_domain_features(text)], dtype=torch.float32).to(device)\n", " with torch.no_grad():\n", " output = model(encoding['input_ids'].to(device), encoding['attention_mask'].to(device), domain)\n", " probs = torch.softmax(output, dim=1)[0]\n", " pred = output.argmax(dim=1).item()\n", " print(f\"Тип: {label_map[pred]}\")\n", " print(f\"Red: {probs[0]:.2%} | White: {probs[1]:.2%} | Rosé: {probs[2]:.2%}\")\n", "\n", "predict(\"Bold and fruity with hints of blackberry and vanilla oak\")\n", "predict(\"Crisp and refreshing with citrus and green apple aromas\")" ] }, { "cell_type": "code", "execution_count": 31, "id": "808e5aa8", "metadata": {}, "outputs": [ { "data": { "text/plain": [ "('wine_model/tokenizer/tokenizer_config.json',\n", " 'wine_model/tokenizer/special_tokens_map.json',\n", " 'wine_model/tokenizer/vocab.txt',\n", " 'wine_model/tokenizer/added_tokens.json',\n", " 'wine_model/tokenizer/tokenizer.json')" ] }, "execution_count": 31, "metadata": {}, "output_type": "execute_result" } ], "source": [ "os.makedirs('wine_model', exist_ok=True)\n", "\n", "torch.save(model.state_dict(), 'wine_model/model.pt')\n", "tokenizer.save_pretrained('wine_model/tokenizer')" ] } ], "metadata": { "kernelspec": { "display_name": ".venv (3.13.3)", "language": "python", "name": "python3" }, "language_info": { "codemirror_mode": { "name": "ipython", "version": 3 }, "file_extension": ".py", "mimetype": "text/x-python", "name": "python", "nbconvert_exporter": "python", "pygments_lexer": "ipython3", "version": "3.13.3" } }, "nbformat": 4, "nbformat_minor": 5 }