{"cells":[{"metadata":{"_uuid":"8f2839f25d086af736a60e9eeb907d3b93b6e0e5","_cell_guid":"b1076dfc-b9ad-4769-8c92-a6c4dae69d19","trusted":true},"cell_type":"code","source":"# This Python 3 environment comes with many helpful analytics libraries installed\n# It is defined by the kaggle/python docker image: https://github.com/kaggle/docker-python\n# For example, here's several helpful packages to load in \n\nimport numpy as np # linear algebra\nimport pandas as pd # data processing, CSV file I/O (e.g. pd.read_csv)\n\n# Input data files are available in the \"../input/\" directory.\n# For example, running this (by clicking run or pressing Shift+Enter) will list all files under the input directory\n\nimport os\nfor dirname, _, filenames in os.walk('/kaggle/input'):\n    for filename in filenames:\n        print(os.path.join(dirname, filename))\n\n# Any results you write to the current directory are saved as output.","execution_count":null,"outputs":[]},{"metadata":{"trusted":true},"cell_type":"code","source":"train = pd.read_csv('/kaggle/input/deepnlp-hse-course/train.csv')\ntest = pd.read_csv('/kaggle/input/deepnlp-hse-course/test.csv')","execution_count":null,"outputs":[]},{"metadata":{"trusted":true},"cell_type":"code","source":"import torch\nfrom torch.utils.data import Dataset, DataLoader\n\nfrom nltk.tokenize import word_tokenize, wordpunct_tokenize\nfrom tqdm import tqdm\n\nfrom sklearn.model_selection import train_test_split\nfrom sklearn.metrics import f1_score","execution_count":null,"outputs":[]},{"metadata":{"trusted":true},"cell_type":"code","source":"def process_text(text):\n    \n    words = wordpunct_tokenize(text.lower())\n    \n    return words","execution_count":null,"outputs":[]},{"metadata":{"trusted":true},"cell_type":"code","source":"word2freq = {}\n\nfor question in tqdm(train.question):\n    \n    words = process_text(question)\n    \n    for word in words:\n        \n        if word in word2freq:\n            word2freq[word] += 1\n        else:\n            word2freq[word] = 1","execution_count":null,"outputs":[]},{"metadata":{"trusted":true},"cell_type":"code","source":"word2index = {'PAD': 0}\nvectors = []\n    \nword2vec_file = open('/kaggle/input/fasttest-common-crawl-russian/cc.ru.300.vec')\n    \nn_words, embedding_dim = word2vec_file.readline().split()\nn_words, embedding_dim = int(n_words), int(embedding_dim)\n\n# Zero vector for PAD\nvectors.append(np.zeros((1, embedding_dim)))\n\nprogress_bar = tqdm(desc='Read word2vec', total=n_words)\n\nwhile True:\n\n    line = word2vec_file.readline().strip()\n\n    if not line:\n        break\n        \n    current_parts = line.split()\n\n    current_word = ' '.join(current_parts[:-embedding_dim])\n\n    if current_word in word2freq:\n\n        word2index[current_word] = len(word2index)\n\n        current_vectors = current_parts[-embedding_dim:]\n        current_vectors = np.array(list(map(float, current_vectors)))\n        current_vectors = np.expand_dims(current_vectors, 0)\n\n        vectors.append(current_vectors)\n\n    progress_bar.update(1)\n\nprogress_bar.close()\n\nword2vec_file.close()\n\nvectors = np.concatenate(vectors)","execution_count":null,"outputs":[]},{"metadata":{"trusted":true},"cell_type":"code","source":"unk_words = [word for word in word2freq if word not in word2index]\nunk_counts = [word2freq[word] for word in unk_words]\nn_unk = sum(unk_counts) * 100 / sum(list(word2freq.values()))\n\nsub_sample_unk_words = {word: word2freq[word] for word in unk_words}\nsorted_unk_words = list(sorted(sub_sample_unk_words, key=lambda x: sub_sample_unk_words[x], reverse=True))\n\nprint('Мы не знаем {:.2f} % слов в датасете'.format(n_unk))\nprint('Количество неизвестных слов {} из {}, то есть {:.2f} % уникальных слов в словаре'.format(\n    len(unk_words), len(word2freq), len(unk_words) * 100 / len(word2freq)))\nprint('В среднем каждое встречается {:.2f} раз'.format(np.mean(unk_counts)))\nprint()\nprint('Топ 5 невошедших слов:')\n\nfor i in range(5):\n    print(sorted_unk_words[i], 'с количеством вхождениий -', word2freq[sorted_unk_words[i]])","execution_count":null,"outputs":[]},{"metadata":{"trusted":true},"cell_type":"code","source":"class WordData(Dataset):\n    \n    def __init__(self, x_data, y_data, word2index, sequence_length=32, pad_token='PAD', verbose=True):\n        \n        super().__init__()\n        \n        self.x_data = []\n        self.y_data = y_data\n        \n        self.word2index = word2index\n        self.sequence_length = sequence_length\n        \n        self.pad_token = pad_token\n        self.pad_index = self.word2index[self.pad_token]\n        \n        self.load(x_data, verbose=verbose)\n        \n    @staticmethod\n    def process_text(text):\n        \n        # Место для вашей предобработки\n    \n        words = wordpunct_tokenize(text.lower())\n\n        return words\n        \n    def load(self, data, verbose=True):\n        \n        data_iterator = tqdm(data, desc='Loading data', disable=not verbose)\n        \n        for text in data_iterator:\n            words = self.process_text(text)\n            indexed_words = self.indexing(words)\n            self.x_data.append(indexed_words)\n    \n    def indexing(self, tokenized_text):\n\n        # здесь мы не используем токен UNK, потому что мы мы его специально не учили\n        # становится непонятно какой же эмбеддинг присвоить неизвестному слову,\n        # поэтому просто выбрасываем наши неизветсные слова\n        \n        ### CODE ###\n\n        return [self.word2index[token] for token in tokenized_text if token in self.word2index ]\n    \n    def padding(self, sequence):\n        \n        # Ограничить длину self.sequence_length\n        # если длина меньше максимально - западить\n        \n        ### CODE ###\n\n        if len(sequence) > self.sequence_length:\n            sequence = sequence[:self.sequence_length]\n        elif len(sequence) < self.sequence_length:\n            sequence = sequence + [self.pad_index] * (self.sequence_length - len(sequence))\n\n        return sequence\n    \n    def __len__(self):\n        \n        return len(self.x_data)\n    \n    def __getitem__(self, idx):\n        \n        x = self.x_data[idx]\n        x = self.padding(x)\n        x = torch.Tensor(x).long()\n        \n        y = self.y_data[idx]\n        \n        return x, y","execution_count":null,"outputs":[]},{"metadata":{"trusted":true},"cell_type":"code","source":"x_train, x_validation, y_train, y_validation = train_test_split(train.question, train.main_category, test_size=0.1)\n\ntrain_dataset = WordData(list(x_train), list(y_train), word2index)\ntrain_loader = DataLoader(train_dataset, batch_size=64)\n\nvalidation_dataset = WordData(list(x_validation), list(y_validation), word2index)\nvalidation_loader = DataLoader(validation_dataset, batch_size=64)\n\ntest_dataset = WordData(list(test.question), np.zeros((test.shape[0])), word2index)\ntest_loader = DataLoader(test_dataset, batch_size=64)","execution_count":null,"outputs":[]},{"metadata":{"trusted":true},"cell_type":"code","source":"for x, y in test_loader:\n    break","execution_count":null,"outputs":[]},{"metadata":{"trusted":true},"cell_type":"code","source":"x.shape","execution_count":null,"outputs":[]},{"metadata":{"trusted":true},"cell_type":"code","source":"y","execution_count":null,"outputs":[]},{"metadata":{"trusted":true},"cell_type":"code","source":"n_classes = train.main_category.unique().shape[0]","execution_count":null,"outputs":[]},{"metadata":{"trusted":true},"cell_type":"code","source":"class DeepAverageNetwork(torch.nn.Module):\n    \n    def __init__(self, embedding_matrix, n_classes):\n        \n        super().__init__()\n        \n        self.embedding_layer = torch.nn.Embedding.from_pretrained(torch.Tensor(embedding_matrix))\n        \n        self.layers = torch.nn.Sequential(torch.nn.Linear(300, 256),\n                                          torch.nn.ReLU(), \n                                          torch.nn.Linear(256, 128),\n                                          torch.nn.ReLU(),\n                                          torch.nn.Linear(128, n_classes))\n    def forward(self, x):\n        \n        sequence_lengths = (x > 0).sum(dim=1)\n        sequence_lengths[sequence_lengths == 0.] = 1\n        \n        x = self.embedding_layer(x)\n        \n        x = x.mean(dim=-2)\n        \n        lengths_scaling = sequence_lengths.float() / x.size(1)\n        lengths_scaling = lengths_scaling.unsqueeze(1).repeat((1, x.size(-1)))\n        x /= lengths_scaling.to(x.device)\n        \n        x = self.layers(x)\n        \n        return x","execution_count":null,"outputs":[]},{"metadata":{"trusted":true},"cell_type":"code","source":"model = DeepAverageNetwork(embedding_matrix=vectors, n_classes=n_classes)","execution_count":null,"outputs":[]},{"metadata":{"trusted":true},"cell_type":"code","source":"with torch.no_grad():\n    pred = model(x)","execution_count":null,"outputs":[]},{"metadata":{"trusted":true},"cell_type":"code","source":"pred.shape","execution_count":null,"outputs":[]},{"metadata":{"trusted":true},"cell_type":"code","source":"device = torch.device('cuda') if torch.cuda.is_available() else torch.device('cpu')","execution_count":null,"outputs":[]},{"metadata":{"trusted":true},"cell_type":"code","source":"criterion = torch.nn.CrossEntropyLoss()\noptimizer = torch.optim.Adam(params=model.parameters())\n\nmodel = model.to(device)\ncriterion = criterion.to(device)","execution_count":null,"outputs":[]},{"metadata":{"trusted":true},"cell_type":"code","source":"epochs = 10\nlosses = []\nbest_test_loss = 10.\n\ntest_f1 = []\n\nfor n_epoch in range(epochs):\n    \n    train_losses = []\n    test_losses = []\n    test_targets = []\n    test_pred_class = []\n    \n    progress_bar = tqdm(total=len(train_loader.dataset), desc='Epoch {}'.format(n_epoch + 1))\n    \n    model.train()\n    \n    for x, y in train_loader:\n\n        x = x.to(device)\n        y = y.to(device)\n        \n        optimizer.zero_grad()\n        \n        pred = model(x)\n        loss = criterion(pred, y)\n        \n        loss.backward()\n        \n        optimizer.step()\n        \n        train_losses.append(loss.item())\n        losses.append(loss.item())\n        \n        progress_bar.set_postfix(train_loss = np.mean(losses[-500:]))\n\n        progress_bar.update(x.shape[0])\n        \n    progress_bar.close()\n    \n    model.eval()\n    \n    for x, y in validation_loader:\n        \n        x = x.to(device)\n        y = y.to(device)\n\n        with torch.no_grad():\n\n            pred = model(x)\n\n            pred = pred.cpu()\n            y = y.cpu()\n\n            test_targets.append(y.numpy())\n            test_pred_class.append(np.argmax(pred, axis=1))\n\n            loss = criterion(pred, y)\n\n            test_losses.append(loss.item())\n        \n    mean_test_loss = np.mean(test_losses)\n\n    test_targets = np.concatenate(test_targets).squeeze()\n    test_pred_class = np.concatenate(test_pred_class).squeeze()\n\n    f1 = f1_score(test_targets, test_pred_class, average='micro')\n\n    test_f1.append(f1)\n    \n    print()\n    print('Losses: train - {:.3f}, test - {:.3f}'.format(np.mean(train_losses), mean_test_loss))\n\n    print('F1 test - {:.3f}'.format(f1))\n        \n    # Early stopping:\n    if mean_test_loss < best_test_loss:\n        best_test_loss = mean_test_loss\n    else:\n        print('Early stopping')\n        break","execution_count":null,"outputs":[]},{"metadata":{"trusted":true},"cell_type":"code","source":"model.eval()\n\npredictions = []\n\nfor x, _ in test_loader:\n\n    x = x.to(device)\n\n    with torch.no_grad():\n\n        pred = model(x)\n\n        pred = pred.cpu()\n        \n        predictions.append(np.argmax(pred, axis=1))\n        \npredictions = np.concatenate(predictions).squeeze()","execution_count":null,"outputs":[]},{"metadata":{"trusted":true},"cell_type":"code","source":"test['main_category'] = predictions","execution_count":null,"outputs":[]},{"metadata":{"trusted":true},"cell_type":"code","source":"test = test[['index', 'main_category']]","execution_count":null,"outputs":[]},{"metadata":{"trusted":true},"cell_type":"code","source":"test.head()","execution_count":null,"outputs":[]},{"metadata":{"trusted":true},"cell_type":"code","source":"test.to_csv('submission.csv', index=False)","execution_count":null,"outputs":[]},{"metadata":{"trusted":true},"cell_type":"code","source":"","execution_count":null,"outputs":[]}],"metadata":{"kernelspec":{"language":"python","display_name":"Python 3","name":"python3"},"language_info":{"pygments_lexer":"ipython3","nbconvert_exporter":"python","version":"3.6.4","file_extension":".py","codemirror_mode":{"name":"ipython","version":3},"name":"python","mimetype":"text/x-python"}},"nbformat":4,"nbformat_minor":1}