chapter_09/chapter_09.ipynb

{
 "cells": [
  {
   "cell_type": "markdown",
   "id": "d51cfabc",
   "metadata": {},
   "source": [
    "# Advanced Regularization in Natural Language Processing"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "83367475",
   "metadata": {},
   "source": [
    "## Regularization using a word2vec embedding"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 1,
   "id": "0695c352",
   "metadata": {
    "scrolled": false
   },
   "outputs": [
    {
     "name": "stderr",
     "output_type": "stream",
     "text": [
      "[nltk_data] Downloading package punkt to /home/vince/nltk_data...\n",
      "[nltk_data]   Package punkt is already up-to-date!\n"
     ]
    },
    {
     "data": {
      "text/plain": [
       "True"
      ]
     },
     "execution_count": 1,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "import torch\n",
    "import torch.nn as nn\n",
    "import torch.optim as optim\n",
    "from torch.utils.data import DataLoader, Dataset\n",
    "from sklearn.model_selection import train_test_split\n",
    "from sklearn.preprocessing import LabelEncoder\n",
    "import pandas as pd\n",
    "import matplotlib.pyplot as plt\n",
    "import numpy as np\n",
    "import gensim.downloader\n",
    "import nltk\n",
    "\n",
    "# If running for the first time\n",
    "nltk.download('punkt')"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 2,
   "id": "9320c41f",
   "metadata": {},
   "outputs": [],
   "source": [
    "# Will take a while the first time, need to download about 1.6GB of the model\n",
    "word2vec_model = gensim.downloader.load('word2vec-google-news-300')"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 4,
   "id": "668dd64f",
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "Downloading imdb-dataset-of-50k-movie-reviews.zip to /home/vince/Documents/Freelance/packt/The-Regularization-Cookbook/chapter_09\n",
      " 93%|███████████████████████████████████▍  | 24.0M/25.7M [00:01<00:00, 36.5MB/s]\n",
      "100%|██████████████████████████████████████| 25.7M/25.7M [00:01<00:00, 23.8MB/s]\n"
     ]
    }
   ],
   "source": [
    "!kaggle datasets download -d lakshmi25npathi/imdb-dataset-of-50k-movie-reviews --unzip"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 3,
   "id": "a16d2aa9",
   "metadata": {},
   "outputs": [
    {
     "data": {
      "text/html": [
       "<div>\n",
       "<style scoped>\n",
       "    .dataframe tbody tr th:only-of-type {\n",
       "        vertical-align: middle;\n",
       "    }\n",
       "\n",
       "    .dataframe tbody tr th {\n",
       "        vertical-align: top;\n",
       "    }\n",
       "\n",
       "    .dataframe thead th {\n",
       "        text-align: right;\n",
       "    }\n",
       "</style>\n",
       "<table border=\"1\" class=\"dataframe\">\n",
       "  <thead>\n",
       "    <tr style=\"text-align: right;\">\n",
       "      <th></th>\n",
       "      <th>review</th>\n",
       "      <th>sentiment</th>\n",
       "    </tr>\n",
       "  </thead>\n",
       "  <tbody>\n",
       "    <tr>\n",
       "      <th>0</th>\n",
       "      <td>One of the other reviewers has mentioned that ...</td>\n",
       "      <td>positive</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>1</th>\n",
       "      <td>A wonderful little production. &lt;br /&gt;&lt;br /&gt;The...</td>\n",
       "      <td>positive</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>2</th>\n",
       "      <td>I thought this was a wonderful way to spend ti...</td>\n",
       "      <td>positive</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>3</th>\n",
       "      <td>Basically there's a family where a little boy ...</td>\n",
       "      <td>negative</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>4</th>\n",
       "      <td>Petter Mattei's \"Love in the Time of Money\" is...</td>\n",
       "      <td>positive</td>\n",
       "    </tr>\n",
       "  </tbody>\n",
       "</table>\n",
       "</div>"
      ],
      "text/plain": [
       "                                              review sentiment\n",
       "0  One of the other reviewers has mentioned that ...  positive\n",
       "1  A wonderful little production. <br /><br />The...  positive\n",
       "2  I thought this was a wonderful way to spend ti...  positive\n",
       "3  Basically there's a family where a little boy ...  negative\n",
       "4  Petter Mattei's \"Love in the Time of Money\" is...  positive"
      ]
     },
     "execution_count": 3,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "# Load data\n",
    "data = pd.read_csv('IMDB Dataset.csv')\n",
    "data.head()"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 4,
   "id": "34cd0861",
   "metadata": {},
   "outputs": [],
   "source": [
    "# Split data into train and test sets\n",
    "train_data, test_data = train_test_split(data, test_size=0.2, random_state=0)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 5,
   "id": "c938a135",
   "metadata": {},
   "outputs": [],
   "source": [
    "# Define dataset class\n",
    "class TextClassificationDataset(Dataset):\n",
    "    def __init__(self, data, word2vec_model, max_words):\n",
    "        self.data = data\n",
    "        self.word2vec_model = word2vec_model\n",
    "        self.max_words = max_words\n",
    "        self.embeddings = data['review'].apply(self.embed)\n",
    "        le = LabelEncoder()\n",
    "        self.labels = torch.tensor(le.fit_transform(data['sentiment']).astype(np.float32))\n",
    "\n",
    "    def __len__(self):\n",
    "        return len(self.data)\n",
    "    \n",
    "    def __getitem__(self, index):\n",
    "        return self.embeddings.iloc[index], self.labels[index]\n",
    "    \n",
    "    def embed(self, text):\n",
    "        tokens = nltk.word_tokenize(text)\n",
    "        return self.tokens_to_embeddings(tokens)\n",
    "    \n",
    "    def tokens_to_embeddings(self, tokens):\n",
    "        embeddings = []\n",
    "        for i, token in enumerate(tokens):\n",
    "            if i >= self.max_words:\n",
    "                break\n",
    "            if token not in self.word2vec_model:\n",
    "                continue\n",
    "            embeddings.append(self.word2vec_model[token])\n",
    "\n",
    "        while len(embeddings) < self.max_words:\n",
    "            embeddings.append(np.zeros((300, )))\n",
    "        \n",
    "        return np.array(embeddings, dtype=np.float32)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 8,
   "id": "849fd4a0",
   "metadata": {},
   "outputs": [],
   "source": [
    "batch_size = 64\n",
    "max_words = 64\n",
    "\n",
    "# Initialize datasets and dataloaders\n",
    "train_dataset = TextClassificationDataset(train_data, word2vec_model, max_words)\n",
    "test_dataset = TextClassificationDataset(test_data, word2vec_model, max_words)\n",
    "train_dataloader = DataLoader(train_dataset, batch_size=batch_size, shuffle=True)\n",
    "test_dataloader = DataLoader(test_dataset, batch_size=batch_size, shuffle=True)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 9,
   "id": "b3d9f421",
   "metadata": {},
   "outputs": [],
   "source": [
    "# Define RNN model\n",
    "class GRUClassifier(nn.Module):\n",
    "    def __init__(self, embedding_dim, hidden_size, output_size, num_layers=3):\n",
    "        super(GRUClassifier, self).__init__()\n",
    "        self.hidden_size = hidden_size\n",
    "        self.num_layers = num_layers\n",
    "        self.gru = nn.GRU(input_size=embedding_dim, hidden_size=hidden_size, num_layers=num_layers, batch_first=True)\n",
    "        self.fc = nn.Linear(hidden_size, output_size)\n",
    "    \n",
    "    def forward(self, inputs):\n",
    "        batch_size = inputs.size(0)\n",
    "        zero_hidden = torch.zeros(self.num_layers, batch_size, self.hidden_size).to(device)\n",
    "        output, hidden = self.gru(inputs, zero_hidden)\n",
    "        output = torch.sigmoid(self.fc(output[:, -1]))\n",
    "        return output"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 10,
   "id": "80fc6276",
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "torch.Size([64, 64, 300])\n",
      "Resulting output tensor: torch.Size([64, 1])\n",
      "Sum of the output tensor: tensor(33.9811, device='cuda:0', grad_fn=<SumBackward0>)\n"
     ]
    }
   ],
   "source": [
    "embedding_dim = 300\n",
    "hidden_dim = 32\n",
    "output_size = 1\n",
    "\n",
    "# Optionally, set the device to GPU if you have one\n",
    "device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')\n",
    "\n",
    "model = GRUClassifier(\n",
    "    embedding_dim=embedding_dim,\n",
    "    hidden_size=hidden_dim, \n",
    "    output_size=output_size,\n",
    ").to(device)\n",
    "\n",
    "random_data = torch.rand((batch_size, max_words, embedding_dim)).to(device)\n",
    "print(random_data.shape)\n",
    "result = model(random_data)\n",
    "print('Resulting output tensor:', result.shape)\n",
    "print('Sum of the output tensor:', result.sum())"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 11,
   "id": "d99640b3",
   "metadata": {},
   "outputs": [],
   "source": [
    "optimizer = optim.Adam(model.parameters(), lr=0.001)\n",
    "criterion = nn.BCELoss()"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "id": "d02c681f",
   "metadata": {},
   "outputs": [],
   "source": [
    "def train_model(model, train_dataloader, test_dataloader, criterion, optimizer, device, epochs: int = 20):\n",
    "    # Train the model\n",
    "    train_losses = []\n",
    "    test_losses = []\n",
    "    train_accuracy = []\n",
    "    test_accuracy = []\n",
    "\n",
    "    for epoch in range(20):\n",
    "        running_train_loss = 0.\n",
    "        correct = 0.\n",
    "        model.train()\n",
    "        for inputs, labels in train_dataloader:\n",
    "            inputs = inputs.to(device)\n",
    "            labels = labels.unsqueeze(1).to(device)\n",
    "            outputs = model(inputs)\n",
    "            loss = criterion(outputs, labels)\n",
    "            optimizer.zero_grad()\n",
    "            loss.backward()\n",
    "            optimizer.step()\n",
    "            # Add correct predictions for this batch\n",
    "            correct += ((outputs > 0.5) == labels).float().sum()\n",
    "            # Compute loss for this batch\n",
    "            running_train_loss += loss.item()\n",
    "\n",
    "        # Compute and store loss and accuracy for this epoch\n",
    "        train_epoch_loss = running_train_loss / len(train_dataloader)\n",
    "        train_losses.append(train_epoch_loss)\n",
    "        train_epoch_accuracy = correct / len(train_dataset)\n",
    "        train_accuracy.append(train_epoch_accuracy.cpu().numpy())\n",
    "\n",
    "        ## Evaluate the model on the test set\n",
    "        running_test_loss = 0.\n",
    "        correct = 0.\n",
    "        model.eval()\n",
    "        with torch.no_grad():\n",
    "            for i, data in enumerate(test_dataloader, 0):\n",
    "                inputs, labels = data\n",
    "                inputs = inputs.to(device)\n",
    "                labels = labels.unsqueeze(1).to(device)\n",
    "                outputs = model(inputs)\n",
    "                loss = criterion(outputs, labels)\n",
    "                running_test_loss += loss.item()\n",
    "                correct += ((outputs > 0.5) == labels).float().sum()\n",
    "\n",
    "            test_epoch_loss = running_test_loss / len(test_dataloader)\n",
    "            test_losses.append(test_epoch_loss)\n",
    "            test_epoch_accuracy = correct / len(test_dataset)\n",
    "            test_accuracy.append(test_epoch_accuracy.cpu().numpy())\n",
    "\n",
    "        # Print stats\n",
    "        print(f'[epoch {epoch + 1}] Training: loss={train_epoch_loss:.3f} accuracy={train_epoch_accuracy:.3f} |\\\n",
    "    \\t Test: loss={test_epoch_loss:.3f} accuracy={test_epoch_accuracy:.3f}')\n",
    "        \n",
    "    return train_losses, test_losses, train_accuracy, test_accuracy"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 12,
   "id": "1f47d0cd",
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "[epoch 1] Training: loss=0.572 accuracy=0.674 |\t Test: loss=0.478 accuracy=0.770\n",
      "[epoch 2] Training: loss=0.451 accuracy=0.783 |\t Test: loss=0.450 accuracy=0.783\n",
      "[epoch 3] Training: loss=0.429 accuracy=0.796 |\t Test: loss=0.433 accuracy=0.797\n",
      "[epoch 4] Training: loss=0.413 accuracy=0.806 |\t Test: loss=0.431 accuracy=0.797\n",
      "[epoch 5] Training: loss=0.399 accuracy=0.813 |\t Test: loss=0.425 accuracy=0.801\n",
      "[epoch 6] Training: loss=0.384 accuracy=0.823 |\t Test: loss=0.417 accuracy=0.804\n",
      "[epoch 7] Training: loss=0.371 accuracy=0.829 |\t Test: loss=0.430 accuracy=0.807\n",
      "[epoch 8] Training: loss=0.360 accuracy=0.835 |\t Test: loss=0.424 accuracy=0.805\n",
      "[epoch 9] Training: loss=0.346 accuracy=0.844 |\t Test: loss=0.429 accuracy=0.804\n",
      "[epoch 10] Training: loss=0.335 accuracy=0.850 |\t Test: loss=0.455 accuracy=0.802\n",
      "[epoch 11] Training: loss=0.320 accuracy=0.859 |\t Test: loss=0.433 accuracy=0.803\n",
      "[epoch 12] Training: loss=0.305 accuracy=0.867 |\t Test: loss=0.428 accuracy=0.803\n",
      "[epoch 13] Training: loss=0.294 accuracy=0.873 |\t Test: loss=0.468 accuracy=0.796\n",
      "[epoch 14] Training: loss=0.279 accuracy=0.881 |\t Test: loss=0.445 accuracy=0.801\n",
      "[epoch 15] Training: loss=0.269 accuracy=0.889 |\t Test: loss=0.474 accuracy=0.801\n",
      "[epoch 16] Training: loss=0.253 accuracy=0.897 |\t Test: loss=0.512 accuracy=0.796\n",
      "[epoch 17] Training: loss=0.242 accuracy=0.902 |\t Test: loss=0.586 accuracy=0.775\n",
      "[epoch 18] Training: loss=0.229 accuracy=0.908 |\t Test: loss=0.485 accuracy=0.793\n",
      "[epoch 19] Training: loss=0.220 accuracy=0.914 |\t Test: loss=0.541 accuracy=0.796\n",
      "[epoch 20] Training: loss=0.207 accuracy=0.917 |\t Test: loss=0.533 accuracy=0.790\n"
     ]
    }
   ],
   "source": [
    "train_losses, test_losses, train_accuracy, test_accuracy = train_model(\n",
    "    model, train_dataloader, test_dataloader, criterion, optimizer, device, epochs=20\n",
    ")"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 13,
   "id": "81efe9b7",
   "metadata": {
    "scrolled": false
   },
   "outputs": [
    {
     "data": {
      "image/png": "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\n",
      "text/plain": [
       "<Figure size 640x480 with 1 Axes>"
      ]
     },
     "metadata": {},
     "output_type": "display_data"
    }
   ],
   "source": [
    "plt.plot(train_losses, label='train')\n",
    "plt.plot(test_losses, label='test')\n",
    "plt.xlabel('epoch')\n",
    "plt.ylabel('loss (BCE)')\n",
    "plt.legend()\n",
    "plt.show()"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 14,
   "id": "032388ca",
   "metadata": {
    "scrolled": false
   },
   "outputs": [
    {
     "data": {
      "image/png": "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\n",
      "text/plain": [
       "<Figure size 640x480 with 1 Axes>"
      ]
     },
     "metadata": {},
     "output_type": "display_data"
    }
   ],
   "source": [
    "plt.plot(train_accuracy, label='train')\n",
    "plt.plot(test_accuracy, label='test')\n",
    "plt.xlabel('epoch')\n",
    "plt.ylabel('Accuracy')\n",
    "plt.legend()\n",
    "plt.show()"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "46f28c93",
   "metadata": {},
   "source": [
    "## Regularize with data augmentation using word2vec"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 1,
   "id": "28e03ccd",
   "metadata": {},
   "outputs": [],
   "source": [
    "import numpy as np\n",
    "import gensim.downloader"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 2,
   "id": "458c1264",
   "metadata": {},
   "outputs": [],
   "source": [
    "# Load the Word2Vec model\n",
    "word2vec_model = gensim.downloader.load('word2vec-google-news-300')"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 46,
   "id": "952d0b0b",
   "metadata": {},
   "outputs": [],
   "source": [
    "def replace_words_with_similar(text, model, sim_threshold: float = 0.5, probability: float = 0.5, top_similar: int = 3, stop_words: list[str] = []):\n",
    "    # Split in words\n",
    "    words = text.split()\n",
    "    # Create an empty list of the output words\n",
    "    new_words = []\n",
    "    # Loop over the words\n",
    "    for word in words:\n",
    "        added = False\n",
    "        # If the word is in the vocab, not in stop words, and above probability, then...\n",
    "        if word in model and word not in stop_words and np.random.uniform(0, 1) > probability:\n",
    "            # Get the top_similar most similar words \n",
    "            similar_words = model.most_similar(word, topn=top_similar)\n",
    "            # Randomly pick one of those words\n",
    "            idx = np.random.randint(len(similar_words))\n",
    "            # Get the similar word and similarity score\n",
    "            sim_word, sim_score = similar_words[idx]\n",
    "            # If the similary score is above threshold, add the word\n",
    "            if sim_score > sim_threshold:\n",
    "                new_words.append(sim_word)\n",
    "                added = True\n",
    "        if not added:\n",
    "            # If no similar word is added, add the original word\n",
    "            new_words.append(word)\n",
    "    # Return the list as a string\n",
    "    return ' '.join(new_words)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 52,
   "id": "1084f610",
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "Original text: The quick brown fox jumps over the lazy dog\n",
      "New text: This quick brown squirrel jumps Over the lazy puppy\n"
     ]
    }
   ],
   "source": [
    "original_text = \"The quick brown fox jumps over the lazy dog\"\n",
    "generated_text = replace_words_with_similar(original_text, word2vec_model, stop_words=['the'])\n",
    "print(\"Original text: {}\".format(original_text))\n",
    "print(\"New text: {}\".format(generated_text))"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "d4bae049",
   "metadata": {},
   "source": [
    "### There's more"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "id": "9b3f9524",
   "metadata": {},
   "outputs": [],
   "source": [
    "class TextClassificationDatasetGeneration(Dataset):\n",
    "    def __init__(self, data, max_length):\n",
    "        self.data = data\n",
    "        self.max_length = max_length\n",
    "        self.tokenizer = AutoTokenizer.from_pretrained('bert-base-uncased')\n",
    "        self.tokens = self.tokenizer(data['review'].to_list(), padding=True, truncation=True, max_length=max_length, return_tensors='pt')['input_ids']\n",
    "        le = LabelEncoder()\n",
    "        self.labels = torch.tensor(le.fit_transform(data['sentiment']).astype(np.float32))\n",
    "\n",
    "    def __len__(self):\n",
    "        return len(self.data)\n",
    "    \n",
    "    def __getitem__(self, index):\n",
    "        # Generate a new text\n",
    "        text = replace_words_with_similar(self.data['review'].iloc[index])\n",
    "        # Tokenize it\n",
    "        tokens = self.tokenizer(text, padding=True, truncation=True, max_length=self.max_length, return_tensors='pt')['input_ids']\n",
    "        return self.tokens[index], self.labels[index]"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "6c0d40cd",
   "metadata": {},
   "source": [
    "## Zero-shot inference with pre-trained models"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 1,
   "id": "c849750f",
   "metadata": {
    "id": "c849750f",
    "scrolled": true
   },
   "outputs": [],
   "source": [
    "import numpy as np\n",
    "import pandas as pd\n",
    "from sklearn.model_selection import train_test_split\n",
    "from sklearn.metrics import accuracy_score\n",
    "from transformers import pipeline"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 2,
   "id": "85b6ff6f",
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "Downloading twitter-airline-sentiment.zip to /home/vince/Documents/Freelance/packt/The-Regularization-Cookbook/chapter_09\n",
      "100%|██████████████████████████████████████| 2.55M/2.55M [00:00<00:00, 5.54MB/s]\n",
      "100%|██████████████████████████████████████| 2.55M/2.55M [00:00<00:00, 4.73MB/s]\n"
     ]
    }
   ],
   "source": [
    "!kaggle datasets download -d crowdflower/twitter-airline-sentiment --unzip"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 3,
   "id": "d0c2b8e8",
   "metadata": {
    "colab": {
     "base_uri": "https://localhost:8080/",
     "height": 530
    },
    "id": "d0c2b8e8",
    "outputId": "b5e55652-e021-46d0-ea6e-be8703eeb558"
   },
   "outputs": [
    {
     "data": {
      "text/html": [
       "<div>\n",
       "<style scoped>\n",
       "    .dataframe tbody tr th:only-of-type {\n",
       "        vertical-align: middle;\n",
       "    }\n",
       "\n",
       "    .dataframe tbody tr th {\n",
       "        vertical-align: top;\n",
       "    }\n",
       "\n",
       "    .dataframe thead th {\n",
       "        text-align: right;\n",
       "    }\n",
       "</style>\n",
       "<table border=\"1\" class=\"dataframe\">\n",
       "  <thead>\n",
       "    <tr style=\"text-align: right;\">\n",
       "      <th></th>\n",
       "      <th>airline_sentiment</th>\n",
       "      <th>text</th>\n",
       "    </tr>\n",
       "  </thead>\n",
       "  <tbody>\n",
       "    <tr>\n",
       "      <th>0</th>\n",
       "      <td>neutral</td>\n",
       "      <td>@VirginAmerica What @dhepburn said.</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>1</th>\n",
       "      <td>positive</td>\n",
       "      <td>@VirginAmerica plus you've added commercials t...</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>2</th>\n",
       "      <td>neutral</td>\n",
       "      <td>@VirginAmerica I didn't today... Must mean I n...</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>3</th>\n",
       "      <td>negative</td>\n",
       "      <td>@VirginAmerica it's really aggressive to blast...</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>4</th>\n",
       "      <td>negative</td>\n",
       "      <td>@VirginAmerica and it's a really big bad thing...</td>\n",
       "    </tr>\n",
       "  </tbody>\n",
       "</table>\n",
       "</div>"
      ],
      "text/plain": [
       "  airline_sentiment                                               text\n",
       "0           neutral                @VirginAmerica What @dhepburn said.\n",
       "1          positive  @VirginAmerica plus you've added commercials t...\n",
       "2           neutral  @VirginAmerica I didn't today... Must mean I n...\n",
       "3          negative  @VirginAmerica it's really aggressive to blast...\n",
       "4          negative  @VirginAmerica and it's a really big bad thing..."
      ]
     },
     "execution_count": 3,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "# Load data\n",
    "data = pd.read_csv('Tweets.csv')\n",
    "data[['airline_sentiment', 'text']].head()"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 4,
   "id": "35055788",
   "metadata": {
    "id": "35055788"
   },
   "outputs": [],
   "source": [
    "# Split data into train and test sets\n",
    "train_data, test_data = train_test_split(data, test_size=0.2, random_state=0)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 5,
   "id": "0b422a4e",
   "metadata": {
    "colab": {
     "base_uri": "https://localhost:8080/",
     "height": 377,
     "referenced_widgets": [
      "ce8a201ae86a405aa6726d48cb1626ac",
      "e400036bac92481a80db36bc8d86a61d",
      "93579496e38e4328b887ce6dca84e829",
      "b2e01b02bb5d442a8f560e0909ad2482",
      "afc9b62fa6444b15870fb69ff9073ce4",
      "51fe07b5142a436087c899eadb5576ca",
      "805c9b73a2b54f8d8a861e9f41a612b6",
      "a30ba9b8f88243a4bbb7485f5e4a6c8e",
      "af1d28b4b45742f4879d31c88461e878",
      "b31b55bb93e449ab81af275eeed874da",
      "2d10559b896a48cabbd3d34ffeae3289",
      "f8b91773391a41da9fda3debd6fd0cb8",
      "1d3ce5c9a0a3463fb0351efc4e781734",
      "c5c69dbef1e745a79b835c49b6635066",
      "5459bb30f3e44612ae375adc029f9a09",
      "8ba51d39837f456fa92e13b20d68fabf",
      "0cced4560f974262bdd95c081a71c122",
      "57f07cf9571d4c4e930cb7f92a46ab28",
      "ea0d1a70429646da9a28a2e0b670b039",
      "bb22fcbb73b0416ea03156da2bf673d5",
      "084d5efe03404e949535e01ca99ee458",
      "cb4da034e13e4421a452447eb90bfc87",
      "1b719d9cbac442868827abc996c2cd6e",
      "c4f3e420f01f4a34a7090bb465392b3b",
      "00c54bd573eb43e3804cc3671b6867d0",
      "8facba96b54b49629493a4fea8150c83",
      "ec2fb8af08b041ec958dac8b9861a896",
      "bc9c40ddff5c40babb2863ff60cd2ab3",
      "f46feb41afd24b0791a6e15558041bdb",
      "e4d64d3f21e4489f86548c9cd00b02f3",
      "25639ef7dc65437daef1cd177b78fb79",
      "cfde87b5f1704148bc5a45206867a349",
      "15e729eb20a34f0e837eb3777fc98f92",
      "13b056d0a68a4fba9a8c1891208175ee",
      "ae64d55c395847a3ac4fe46daf7365d7",
      "0d9f9c9c81124c9395bc71cf64ebcbea",
      "4765b8540324472cac2958e110179157",
      "b1083442bb504ad4a049e962daeeee29",
      "bb86ecf340aa41168e152aa0c0074c1f",
      "7149a449be1548e6a310380b6d52dc97",
      "a0432ff359bf4f0fb7976ac48762e960",
      "16bbeb845e2347a9977418a0a3f04f60",
      "e3ca02980b244ae8879848530870fe4e",
      "4c59a6dd90ab4f26afb97b8a773fba00",
      "f2696ad0d0064ccab2ac7d3971cc3e55",
      "46d769de8879433c9b4af8bb0d48e7d3",
      "59bee06a8ddf44e392a0f489a328f858",
      "d43ff8a425de49efb62f1b7771dc10d6",
      "cefbb174ffdf49339df4324427d8e4da",
      "0a62c452141d4a16843d4d70c6287391",
      "bc07a8ec40bd4a08a64fe50e2c76d5c4",
      "4fecc3041fb146e5b50e58281d1e45b5",
      "94394a74a5004a9ea0fd524920b57f2c",
      "dbfba22bc95d4263ac30abba9348310f",
      "817a1f47797741eebd49afc65fc05dc4",
      "3df627c6dd08489ca5de662b8973dc52",
      "07d97b8b0ef5458ab9cceeb900fc1d64",
      "71b26e8cbd964ca1b94916444eaad56f",
      "ec5040647c2146ac919ff349b5e185fb",
      "bff37eaacb8e43608c1f91e03ac96a9d",
      "d5f46a10fbe44010ae59fb5b1b6e449e",
      "781641889b284724bb9cda04adb9e72d",
      "878a69f8ff314df98dc303995c898d48",
      "6f4b0a1693e34a99bf231c58f1b3ea29",
      "5bda0b1420b04876a66051bb296f4ec6",
      "f18aa9b01c314179910f11272db814f1"
     ]
    },
    "id": "0b422a4e",
    "outputId": "1b735a4a-c4b2-4484-d5bb-40ea7de3deac",
    "scrolled": true
   },
   "outputs": [],
   "source": [
    "# Taking a long time the first time for downloading model...\n",
    "classifier = pipeline(task=\"zero-shot-classification\", model=\"facebook/bart-large-mnli\")"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 6,
   "id": "59598fac",
   "metadata": {
    "colab": {
     "base_uri": "https://localhost:8080/"
    },
    "id": "59598fac",
    "outputId": "60114ee3-ea28-414b-fa7f-81fde7178a35"
   },
   "outputs": [
    {
     "data": {
      "text/plain": [
       "array(['neutral', 'positive', 'negative'], dtype=object)"
      ]
     },
     "execution_count": 6,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "candidate_labels = data['airline_sentiment'].unique()\n",
    "candidate_labels"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 10,
   "id": "c8dd248b",
   "metadata": {
    "id": "c8dd248b"
   },
   "outputs": [],
   "source": [
    "# Create an empty list to store the predictions\n",
    "preds = []\n",
    "# Loop over the data\n",
    "for i in range(len(test_data)):\n",
    "    # Compute the classifier results\n",
    "    res = classifier(\n",
    "        test_data['text'].iloc[i],\n",
    "        candidate_labels=candidate_labels,\n",
    "    )\n",
    "    # Apply softmax to the results to get the predicted class\n",
    "    pred = np.array(res['scores']).argmax()\n",
    "    labels = res['labels']\n",
    "    # Store the results in the list\n",
    "    preds.append(labels[pred])"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 12,
   "id": "21a7442d",
   "metadata": {
    "colab": {
     "base_uri": "https://localhost:8080/"
    },
    "id": "21a7442d",
    "outputId": "aeeca5cd-cd48-4dc6-dec7-659af0f54dad"
   },
   "outputs": [
    {
     "data": {
      "text/plain": [
       "0.7452725250278087"
      ]
     },
     "execution_count": 12,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "accuracy_score(test_data['airline_sentiment'], preds)"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "e64caf81",
   "metadata": {},
   "source": [
    "### There's more"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 15,
   "id": "7806ba23",
   "metadata": {},
   "outputs": [
    {
     "data": {
      "text/plain": [
       "{'sequence': 'I love to learn about regularization',\n",
       " 'labels': ['positive', 'neutral', 'negative'],\n",
       " 'scores': [0.6277033686637878, 0.27620458602905273, 0.09609206020832062]}"
      ]
     },
     "execution_count": 15,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "res = classifier(\n",
    "    'I love to learn about regularization',\n",
    "    candidate_labels=['positive', 'negative', 'neutral'],\n",
    ")\n",
    "res"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 17,
   "id": "0d733bc5",
   "metadata": {},
   "outputs": [
    {
     "data": {
      "text/plain": [
       "'positive'"
      ]
     },
     "execution_count": 17,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "res['labels'][np.array(res['scores']).argmax()]"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "80dbec5a",
   "metadata": {},
   "source": [
    "## Regularization with BERT embeddings"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 1,
   "id": "ad7bc09f",
   "metadata": {
    "id": "ad7bc09f",
    "scrolled": true
   },
   "outputs": [],
   "source": [
    "import torch\n",
    "from sklearn.model_selection import train_test_split\n",
    "from sklearn.linear_model import LogisticRegression\n",
    "from transformers import BertConfig, BertModel, BertTokenizer\n",
    "import pandas as pd"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 2,
   "id": "a2a86104",
   "metadata": {},
   "outputs": [
    {
     "data": {
      "text/html": [
       "<div>\n",
       "<style scoped>\n",
       "    .dataframe tbody tr th:only-of-type {\n",
       "        vertical-align: middle;\n",
       "    }\n",
       "\n",
       "    .dataframe tbody tr th {\n",
       "        vertical-align: top;\n",
       "    }\n",
       "\n",
       "    .dataframe thead th {\n",
       "        text-align: right;\n",
       "    }\n",
       "</style>\n",
       "<table border=\"1\" class=\"dataframe\">\n",
       "  <thead>\n",
       "    <tr style=\"text-align: right;\">\n",
       "      <th></th>\n",
       "      <th>tweet_id</th>\n",
       "      <th>airline_sentiment</th>\n",
       "      <th>airline_sentiment_confidence</th>\n",
       "      <th>negativereason</th>\n",
       "      <th>negativereason_confidence</th>\n",
       "      <th>airline</th>\n",
       "      <th>airline_sentiment_gold</th>\n",
       "      <th>name</th>\n",
       "      <th>negativereason_gold</th>\n",
       "      <th>retweet_count</th>\n",
       "      <th>text</th>\n",
       "      <th>tweet_coord</th>\n",
       "      <th>tweet_created</th>\n",
       "      <th>tweet_location</th>\n",
       "      <th>user_timezone</th>\n",
       "    </tr>\n",
       "  </thead>\n",
       "  <tbody>\n",
       "    <tr>\n",
       "      <th>0</th>\n",
       "      <td>570306133677760513</td>\n",
       "      <td>neutral</td>\n",
       "      <td>1.0000</td>\n",
       "      <td>NaN</td>\n",
       "      <td>NaN</td>\n",
       "      <td>Virgin America</td>\n",
       "      <td>NaN</td>\n",
       "      <td>cairdin</td>\n",
       "      <td>NaN</td>\n",
       "      <td>0</td>\n",
       "      <td>@VirginAmerica What @dhepburn said.</td>\n",
       "      <td>NaN</td>\n",
       "      <td>2015-02-24 11:35:52 -0800</td>\n",
       "      <td>NaN</td>\n",
       "      <td>Eastern Time (US &amp; Canada)</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>1</th>\n",
       "      <td>570301130888122368</td>\n",
       "      <td>positive</td>\n",
       "      <td>0.3486</td>\n",
       "      <td>NaN</td>\n",
       "      <td>0.0000</td>\n",
       "      <td>Virgin America</td>\n",
       "      <td>NaN</td>\n",
       "      <td>jnardino</td>\n",
       "      <td>NaN</td>\n",
       "      <td>0</td>\n",
       "      <td>@VirginAmerica plus you've added commercials t...</td>\n",
       "      <td>NaN</td>\n",
       "      <td>2015-02-24 11:15:59 -0800</td>\n",
       "      <td>NaN</td>\n",
       "      <td>Pacific Time (US &amp; Canada)</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>2</th>\n",
       "      <td>570301083672813571</td>\n",
       "      <td>neutral</td>\n",
       "      <td>0.6837</td>\n",
       "      <td>NaN</td>\n",
       "      <td>NaN</td>\n",
       "      <td>Virgin America</td>\n",
       "      <td>NaN</td>\n",
       "      <td>yvonnalynn</td>\n",
       "      <td>NaN</td>\n",
       "      <td>0</td>\n",
       "      <td>@VirginAmerica I didn't today... Must mean I n...</td>\n",
       "      <td>NaN</td>\n",
       "      <td>2015-02-24 11:15:48 -0800</td>\n",
       "      <td>Lets Play</td>\n",
       "      <td>Central Time (US &amp; Canada)</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>3</th>\n",
       "      <td>570301031407624196</td>\n",
       "      <td>negative</td>\n",
       "      <td>1.0000</td>\n",
       "      <td>Bad Flight</td>\n",
       "      <td>0.7033</td>\n",
       "      <td>Virgin America</td>\n",
       "      <td>NaN</td>\n",
       "      <td>jnardino</td>\n",
       "      <td>NaN</td>\n",
       "      <td>0</td>\n",
       "      <td>@VirginAmerica it's really aggressive to blast...</td>\n",
       "      <td>NaN</td>\n",
       "      <td>2015-02-24 11:15:36 -0800</td>\n",
       "      <td>NaN</td>\n",
       "      <td>Pacific Time (US &amp; Canada)</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>4</th>\n",
       "      <td>570300817074462722</td>\n",
       "      <td>negative</td>\n",
       "      <td>1.0000</td>\n",
       "      <td>Can't Tell</td>\n",
       "      <td>1.0000</td>\n",
       "      <td>Virgin America</td>\n",
       "      <td>NaN</td>\n",
       "      <td>jnardino</td>\n",
       "      <td>NaN</td>\n",
       "      <td>0</td>\n",
       "      <td>@VirginAmerica and it's a really big bad thing...</td>\n",
       "      <td>NaN</td>\n",
       "      <td>2015-02-24 11:14:45 -0800</td>\n",
       "      <td>NaN</td>\n",
       "      <td>Pacific Time (US &amp; Canada)</td>\n",
       "    </tr>\n",
       "  </tbody>\n",
       "</table>\n",
       "</div>"
      ],
      "text/plain": [
       "             tweet_id airline_sentiment  airline_sentiment_confidence  \\\n",
       "0  570306133677760513           neutral                        1.0000   \n",
       "1  570301130888122368          positive                        0.3486   \n",
       "2  570301083672813571           neutral                        0.6837   \n",
       "3  570301031407624196          negative                        1.0000   \n",
       "4  570300817074462722          negative                        1.0000   \n",
       "\n",
       "  negativereason  negativereason_confidence         airline  \\\n",
       "0            NaN                        NaN  Virgin America   \n",
       "1            NaN                     0.0000  Virgin America   \n",
       "2            NaN                        NaN  Virgin America   \n",
       "3     Bad Flight                     0.7033  Virgin America   \n",
       "4     Can't Tell                     1.0000  Virgin America   \n",
       "\n",
       "  airline_sentiment_gold        name negativereason_gold  retweet_count  \\\n",
       "0                    NaN     cairdin                 NaN              0   \n",
       "1                    NaN    jnardino                 NaN              0   \n",
       "2                    NaN  yvonnalynn                 NaN              0   \n",
       "3                    NaN    jnardino                 NaN              0   \n",
       "4                    NaN    jnardino                 NaN              0   \n",
       "\n",
       "                                                text tweet_coord  \\\n",
       "0                @VirginAmerica What @dhepburn said.         NaN   \n",
       "1  @VirginAmerica plus you've added commercials t...         NaN   \n",
       "2  @VirginAmerica I didn't today... Must mean I n...         NaN   \n",
       "3  @VirginAmerica it's really aggressive to blast...         NaN   \n",
       "4  @VirginAmerica and it's a really big bad thing...         NaN   \n",
       "\n",
       "               tweet_created tweet_location               user_timezone  \n",
       "0  2015-02-24 11:35:52 -0800            NaN  Eastern Time (US & Canada)  \n",
       "1  2015-02-24 11:15:59 -0800            NaN  Pacific Time (US & Canada)  \n",
       "2  2015-02-24 11:15:48 -0800      Lets Play  Central Time (US & Canada)  \n",
       "3  2015-02-24 11:15:36 -0800            NaN  Pacific Time (US & Canada)  \n",
       "4  2015-02-24 11:14:45 -0800            NaN  Pacific Time (US & Canada)  "
      ]
     },
     "execution_count": 2,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "# Load data\n",
    "data = pd.read_csv('Tweets.csv')\n",
    "data.head()"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 3,
   "id": "eb3299d5",
   "metadata": {},
   "outputs": [],
   "source": [
    "# Split data into train and test sets\n",
    "train_data, test_data = train_test_split(data, test_size=0.2, random_state=0)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 4,
   "id": "e163830b",
   "metadata": {
    "scrolled": true
   },
   "outputs": [
    {
     "name": "stderr",
     "output_type": "stream",
     "text": [
      "Some weights of the model checkpoint at bert-base-uncased were not used when initializing BertModel: ['cls.predictions.decoder.weight', 'cls.predictions.bias', 'cls.seq_relationship.weight', 'cls.seq_relationship.bias', 'cls.predictions.transform.LayerNorm.weight', 'cls.predictions.transform.LayerNorm.bias', 'cls.predictions.transform.dense.weight', 'cls.predictions.transform.dense.bias']\n",
      "- This IS expected if you are initializing BertModel from the checkpoint of a model trained on another task or with another architecture (e.g. initializing a BertForSequenceClassification model from a BertForPreTraining model).\n",
      "- This IS NOT expected if you are initializing BertModel from the checkpoint of a model that you expect to be exactly identical (initializing a BertForSequenceClassification model from a BertForSequenceClassification model).\n"
     ]
    },
    {
     "data": {
      "text/plain": [
       "BertModel(\n",
       "  (embeddings): BertEmbeddings(\n",
       "    (word_embeddings): Embedding(30522, 768, padding_idx=0)\n",
       "    (position_embeddings): Embedding(512, 768)\n",
       "    (token_type_embeddings): Embedding(2, 768)\n",
       "    (LayerNorm): LayerNorm((768,), eps=1e-12, elementwise_affine=True)\n",
       "    (dropout): Dropout(p=0.1, inplace=False)\n",
       "  )\n",
       "  (encoder): BertEncoder(\n",
       "    (layer): ModuleList(\n",
       "      (0-11): 12 x BertLayer(\n",
       "        (attention): BertAttention(\n",
       "          (self): BertSelfAttention(\n",
       "            (query): Linear(in_features=768, out_features=768, bias=True)\n",
       "            (key): Linear(in_features=768, out_features=768, bias=True)\n",
       "            (value): Linear(in_features=768, out_features=768, bias=True)\n",
       "            (dropout): Dropout(p=0.1, inplace=False)\n",
       "          )\n",
       "          (output): BertSelfOutput(\n",
       "            (dense): Linear(in_features=768, out_features=768, bias=True)\n",
       "            (LayerNorm): LayerNorm((768,), eps=1e-12, elementwise_affine=True)\n",
       "            (dropout): Dropout(p=0.1, inplace=False)\n",
       "          )\n",
       "        )\n",
       "        (intermediate): BertIntermediate(\n",
       "          (dense): Linear(in_features=768, out_features=3072, bias=True)\n",
       "          (intermediate_act_fn): GELUActivation()\n",
       "        )\n",
       "        (output): BertOutput(\n",
       "          (dense): Linear(in_features=3072, out_features=768, bias=True)\n",
       "          (LayerNorm): LayerNorm((768,), eps=1e-12, elementwise_affine=True)\n",
       "          (dropout): Dropout(p=0.1, inplace=False)\n",
       "        )\n",
       "      )\n",
       "    )\n",
       "  )\n",
       "  (pooler): BertPooler(\n",
       "    (dense): Linear(in_features=768, out_features=768, bias=True)\n",
       "    (activation): Tanh()\n",
       "  )\n",
       ")"
      ]
     },
     "execution_count": 4,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "# Instantiate the tokenizer\n",
    "tokenizer = BertTokenizer.from_pretrained('bert-base-uncased')\n",
    "# Initializing a BERT configuration\n",
    "configuration = BertConfig()\n",
    "# Initializing a BERt model with random weights\n",
    "bert = BertModel(configuration)\n",
    "# Loading pre-trained weights\n",
    "bert = bert.from_pretrained('bert-base-uncased')\n",
    "# Load the model on the GPU if any\n",
    "device = torch.device(\"cuda\" if torch.cuda.is_available() else \"cpu\")\n",
    "bert.to(device)\n",
    "# Set the model to eval mode\n",
    "bert.eval()"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 5,
   "id": "a5175421",
   "metadata": {},
   "outputs": [],
   "source": [
    "max_length = 24\n",
    "# Compute the embeddings for the train set\n",
    "train_tokens = tokenizer(train_data['text'].values.tolist(), add_special_tokens=True, padding='max_length', truncation=True, max_length=max_length, return_tensors='pt')\n",
    "train_tokens = {k: v.to(device) for k, v in train_tokens.items()}\n",
    "with torch.no_grad():\n",
    "    train_embeddings = bert(**train_tokens).pooler_output\n",
    "# Compute the embeddings for the test set\n",
    "test_tokens = tokenizer(test_data['text'].values.tolist(), add_special_tokens=True, padding='max_length', truncation=True, max_length=max_length, return_tensors='pt')\n",
    "test_tokens = {k: v.to(device) for k, v in test_tokens.items()}\n",
    "with torch.no_grad():\n",
    "    test_embeddings = bert(**test_tokens).pooler_output"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "id": "ec69b44c",
   "metadata": {},
   "outputs": [],
   "source": [
    "lr = LogisticRegression(C=0.5, max_iter=10000)\n",
    "lr.fit(train_embeddings.cpu(), train_data['airline_sentiment'])"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 7,
   "id": "deb58ea3",
   "metadata": {
    "scrolled": false
   },
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "train accuracy: 0.8035348360655737\n",
      "test accuracy: 0.7882513661202186\n"
     ]
    }
   ],
   "source": [
    "print('train accuracy:', lr.score(train_embeddings.cpu(), train_data['airline_sentiment']))\n",
    "print('test accuracy:', lr.score(test_embeddings.cpu(), test_data['airline_sentiment']))"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "28c7dba2",
   "metadata": {},
   "source": [
    "### There's more"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 16,
   "id": "734c4a73",
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "{'input_ids': tensor([[  101,  2054,  2003,  1037, 19204, 17629,  1029,  2023,  2003,  1037,\n",
      "          2204,  3160,  1012,   102,     0,     0,     0,     0,     0,     0,\n",
      "             0,     0,     0,     0]]), 'token_type_ids': tensor([[0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0]]), 'attention_mask': tensor([[1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0]])}\n"
     ]
    }
   ],
   "source": [
    "tokens = tokenizer('What is a tokenizer?', add_special_tokens=True, padding='max_length', truncation=True, max_length=max_length, return_tensors='pt')\n",
    "print(tokens)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 23,
   "id": "dc1f8dac",
   "metadata": {
    "scrolled": true
   },
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "tensor([[-0.8947, -0.3769, -0.8185,  0.7429,  0.5408, -0.1211,  0.8675,  0.2659,\n",
      "         -0.6507, -1.0000, -0.2761,  0.8524,  0.9738,  0.4190,  0.9171, -0.7309,\n",
      "         -0.1398, -0.5986,  0.2661, -0.5249,  0.6231,  0.9999,  0.1622,  0.3898,\n",
      "          0.3964,  0.9409, -0.7071,  0.9133,  0.9513,  0.7316, -0.6867,  0.1575,\n",
      "         -0.9853, -0.1422, -0.7964, -0.9872,  0.3429, -0.6807,  0.0680,  0.0643,\n",
      "         -0.8723,  0.2593,  0.9999, -0.5637,  0.3893, -0.3236, -1.0000,  0.2309,\n",
      "         -0.8693,  0.8439,  0.8033,  0.7725,  0.1262,  0.4381,  0.4909, -0.2119,\n",
      "         -0.1303,  0.1261, -0.2855, -0.5235, -0.6414,  0.3183, -0.6871, -0.8881,\n",
      "          0.7282,  0.7444, -0.0875, -0.2863, -0.0955, -0.0733,  0.8767,  0.1984,\n",
      "         -0.2358, -0.8437,  0.5129,  0.1521, -0.6840,  1.0000, -0.5588, -0.9722,\n",
      "          0.7326,  0.7103,  0.6082, -0.2197,  0.4593, -1.0000,  0.5995,  0.0387,\n",
      "         -0.9855,  0.1585,  0.5374, -0.2143,  0.6246,  0.6323, -0.5971, -0.4695,\n",
      "         -0.3326, -0.6912, -0.1741, -0.2946,  0.0922, -0.2478, -0.3453, -0.3638,\n",
      "          0.2821, -0.5113, -0.5566,  0.6299,  0.1535,  0.6737,  0.4999, -0.3053,\n",
      "          0.3765, -0.9379,  0.6018, -0.3739, -0.9827, -0.6631, -0.9816,  0.6305,\n",
      "         -0.0697, -0.1893,  0.9470, -0.1557,  0.4019, -0.1564, -0.8164, -1.0000,\n",
      "         -0.5143, -0.5879, -0.2293, -0.2650, -0.9721, -0.9569,  0.6309,  0.9597,\n",
      "          0.2255,  0.9997, -0.2128,  0.9151, -0.2301, -0.5306,  0.4005, -0.4125,\n",
      "          0.7363,  0.3229, -0.6048,  0.1894, -0.3140,  0.2744, -0.6307, -0.2526,\n",
      "         -0.7327, -0.9034, -0.3566,  0.9254, -0.3734, -0.8414,  0.1034, -0.2369,\n",
      "         -0.3529,  0.8411,  0.7034,  0.3951, -0.3648,  0.4487,  0.3167,  0.5252,\n",
      "         -0.8508,  0.0873,  0.4024, -0.3523, -0.7956, -0.9688, -0.3644,  0.5094,\n",
      "          0.9846,  0.7169,  0.2729,  0.7034, -0.3327,  0.5995, -0.9368,  0.9751,\n",
      "         -0.0682,  0.2862, -0.4909,  0.3629, -0.8810,  0.1158,  0.8153, -0.5005,\n",
      "         -0.8082, -0.0057, -0.5074, -0.3723, -0.6882,  0.4987, -0.3249, -0.3577,\n",
      "         -0.0855,  0.9213,  0.9703,  0.7038,  0.0470,  0.5744, -0.8554, -0.4701,\n",
      "          0.0841,  0.2506,  0.0161,  0.9896, -0.6853, -0.1093, -0.9241, -0.9757,\n",
      "         -0.0907, -0.8820, -0.1457, -0.6577,  0.5822, -0.2220,  0.4585,  0.4749,\n",
      "         -0.9592, -0.8094,  0.3428, -0.4746,  0.4076, -0.2833,  0.8737,  0.8842,\n",
      "         -0.6067,  0.6472,  0.9295, -0.7992, -0.7293,  0.7619, -0.2958,  0.8867,\n",
      "         -0.5881,  0.9860,  0.8170,  0.6390, -0.8725, -0.6492, -0.8787, -0.6318,\n",
      "         -0.1354,  0.0285,  0.8141,  0.7019,  0.3632,  0.3339, -0.4959,  0.9951,\n",
      "         -0.9013, -0.9441, -0.3405, -0.0793, -0.9867,  0.7706,  0.2813,  0.2518,\n",
      "         -0.4130, -0.6318, -0.9419,  0.7883,  0.0983,  0.9790, -0.4074, -0.8963,\n",
      "         -0.5877, -0.8863, -0.1886, -0.2099, -0.2642, -0.1223, -0.9388,  0.5592,\n",
      "          0.5687,  0.4984, -0.7116,  0.9972,  1.0000,  0.9639,  0.8753,  0.8682,\n",
      "         -0.9993, -0.5748,  1.0000, -0.9697, -1.0000, -0.9219, -0.6573,  0.4017,\n",
      "         -1.0000, -0.2632, -0.0212, -0.8547,  0.5267,  0.9667,  0.9880, -1.0000,\n",
      "          0.8252,  0.9172, -0.7073,  0.8607, -0.4014,  0.9624,  0.5866,  0.5798,\n",
      "         -0.1943,  0.3552, -0.8762, -0.8275, -0.4475, -0.6115,  0.9943,  0.1557,\n",
      "         -0.8239, -0.8470,  0.4815, -0.1449, -0.1796, -0.9595, -0.2141,  0.3020,\n",
      "          0.7732,  0.1330,  0.3232, -0.6856,  0.2551,  0.0999,  0.2367,  0.6977,\n",
      "         -0.9288, -0.4959,  0.0299, -0.2896, -0.5265, -0.9591,  0.9528, -0.3486,\n",
      "          0.6735,  1.0000,  0.1925, -0.8301,  0.5841,  0.2572, -0.6304,  1.0000,\n",
      "          0.7709, -0.9724, -0.6401,  0.6061, -0.5220, -0.5927,  0.9991, -0.2314,\n",
      "         -0.5429, -0.3000,  0.9617, -0.9855,  0.9893, -0.9118, -0.9611,  0.9605,\n",
      "          0.9284, -0.4544, -0.6762,  0.0581, -0.6756,  0.2627, -0.9340,  0.6592,\n",
      "          0.5003, -0.1164,  0.8721, -0.7947, -0.6451,  0.3659, -0.5297,  0.1116,\n",
      "          0.8988,  0.4350, -0.2993, -0.0211, -0.2713, -0.6482, -0.9653,  0.5329,\n",
      "          1.0000, -0.1551,  0.5720, -0.3886, -0.0822, -0.1859,  0.5531,  0.5531,\n",
      "         -0.3017, -0.8531,  0.6327, -0.9338, -0.9813,  0.6658,  0.1512, -0.3391,\n",
      "          1.0000,  0.3559,  0.1471,  0.2495,  0.9235, -0.0098,  0.5311,  0.7620,\n",
      "          0.9788, -0.2938,  0.6553,  0.8280, -0.8292, -0.3138, -0.7050, -0.0311,\n",
      "         -0.9088,  0.0376, -0.9426,  0.9577,  0.7560,  0.4054,  0.1999,  0.6327,\n",
      "          1.0000, -0.6974,  0.6180, -0.0285,  0.7446, -0.9994, -0.7894, -0.3298,\n",
      "         -0.0648, -0.7457, -0.3454,  0.2781, -0.9515,  0.7126,  0.4429, -0.9747,\n",
      "         -0.9809, -0.0160,  0.7993,  0.0287, -0.9652, -0.6557, -0.6020,  0.5154,\n",
      "         -0.3711, -0.9077, -0.0022, -0.2758,  0.4638, -0.3446,  0.6690,  0.7998,\n",
      "          0.7117, -0.6802, -0.2564, -0.1060, -0.7328,  0.8233, -0.8352, -0.8514,\n",
      "         -0.1433,  1.0000, -0.5575,  0.8778,  0.7400,  0.7145, -0.1530,  0.2101,\n",
      "          0.8714,  0.2496, -0.6848, -0.6819, -0.2509, -0.3841,  0.6909,  0.4829,\n",
      "          0.7089,  0.7667,  0.7341,  0.0613, -0.0711,  0.0056,  0.9988, -0.0290,\n",
      "         -0.2140, -0.5516, -0.0291, -0.3421, -0.0619,  1.0000,  0.3065,  0.2100,\n",
      "         -0.9829, -0.7141, -0.8791,  1.0000,  0.8104, -0.7823,  0.6749,  0.5129,\n",
      "         -0.1432,  0.7348, -0.1558, -0.2354,  0.1185,  0.1094,  0.9297, -0.5428,\n",
      "         -0.9598, -0.5726,  0.3695, -0.9496,  0.9996, -0.5794, -0.2116, -0.3305,\n",
      "         -0.1448,  0.0914, -0.1652, -0.9745, -0.2282,  0.1648,  0.9423,  0.2230,\n",
      "         -0.6485, -0.8670,  0.6265,  0.6776, -0.7930, -0.8944,  0.9473, -0.9683,\n",
      "          0.5913,  1.0000,  0.3661, -0.1340,  0.1897, -0.5366,  0.3233, -0.1290,\n",
      "          0.7142, -0.9482, -0.3540, -0.1540,  0.2018, -0.1680, -0.3709,  0.5342,\n",
      "          0.1531, -0.6077, -0.5333, -0.1322,  0.4422,  0.7941, -0.1733, -0.0844,\n",
      "          0.1784, -0.1516, -0.8728, -0.3284, -0.4418, -1.0000,  0.6855, -1.0000,\n",
      "          0.3434,  0.1393, -0.1857,  0.8185,  0.5661,  0.6260, -0.6892, -0.7183,\n",
      "          0.5309,  0.7308, -0.2424, -0.1655, -0.6115,  0.2995,  0.0176,  0.1481,\n",
      "         -0.5469,  0.7604, -0.2286,  1.0000,  0.1108, -0.7093, -0.9608,  0.2316,\n",
      "         -0.2862,  1.0000, -0.8963, -0.9332,  0.3587, -0.7341, -0.8044,  0.3398,\n",
      "          0.0268, -0.6696, -0.8682,  0.9343,  0.8191, -0.6639,  0.4493, -0.2659,\n",
      "         -0.5172, -0.0198,  0.7330,  0.9808,  0.5112,  0.9119,  0.1258, -0.2332,\n",
      "          0.9637,  0.2772,  0.3838,  0.0049,  1.0000,  0.2962, -0.9244,  0.3074,\n",
      "         -0.9806, -0.1029, -0.9497,  0.2481,  0.2148,  0.8832, -0.2284,  0.9320,\n",
      "         -0.5969, -0.0138, -0.3953, -0.3560,  0.3457, -0.9007, -0.9788, -0.9739,\n",
      "          0.5209, -0.3962, -0.0713,  0.2502,  0.0882,  0.4073,  0.3575, -1.0000,\n",
      "          0.9139,  0.4312,  0.8522,  0.9362,  0.6475,  0.4919,  0.2331, -0.9776,\n",
      "         -0.9618, -0.2772, -0.2651,  0.7398,  0.6012,  0.8911,  0.3578, -0.4987,\n",
      "         -0.5231, -0.5632, -0.7339, -0.9895,  0.4238, -0.5237, -0.9412,  0.9536,\n",
      "         -0.2304, -0.0958,  0.0714, -0.7483,  0.8384,  0.7466,  0.3269,  0.0669,\n",
      "          0.5152,  0.8469,  0.9335,  0.9781, -0.7427,  0.7679, -0.5253,  0.4792,\n",
      "          0.7947, -0.9279,  0.0877,  0.4830, -0.3713,  0.2607, -0.2351, -0.9471,\n",
      "          0.6708, -0.2534,  0.5377, -0.4108,  0.0444, -0.3616, -0.1510, -0.7593,\n",
      "         -0.6190,  0.6931,  0.4030,  0.8776,  0.7584, -0.0188, -0.6706, -0.0731,\n",
      "         -0.7119, -0.9010,  0.9063, -0.0302, -0.3186,  0.5539, -0.0551,  0.8991,\n",
      "          0.1510, -0.3858, -0.2562, -0.7680,  0.8355, -0.6208, -0.5459, -0.5147,\n",
      "          0.6378,  0.2897,  0.9999, -0.7117, -0.8168, -0.4287, -0.3122,  0.3756,\n",
      "         -0.4129, -1.0000,  0.3355, -0.3500,  0.6304, -0.5393,  0.7149, -0.4586,\n",
      "         -0.9639, -0.2225,  0.6004,  0.6894, -0.4942, -0.4236,  0.6473,  0.0515,\n",
      "          0.9040,  0.8568, -0.5347,  0.1156,  0.7070, -0.7850, -0.6372,  0.8980]],\n",
      "       device='cuda:0', grad_fn=<TanhBackward0>)\n"
     ]
    }
   ],
   "source": [
    "tokens = {k: v.to(device) for k, v in tokens.items()}\n",
    "output = bert(**tokens)\n",
    "print(output.pooler_output)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 13,
   "id": "7953c923",
   "metadata": {},
   "outputs": [],
   "source": [
    "import openai\n",
    "# Give your \n",
    "openai.api_key = 'xx-xxx'\n",
    "# Query the API\n",
    "input_text = 'This is a test sentence'\n",
    "model = 'text-embedding-ada-002'\n",
    "embeddings = openai.Embedding.create(input = [input_text], model=model)['data'][0]['embedding']"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "c64e31f5",
   "metadata": {},
   "source": [
    "## Data augmentation using GPT-3"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 2,
   "id": "030aaeba",
   "metadata": {},
   "outputs": [],
   "source": [
    "import openai"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 3,
   "id": "b816b188",
   "metadata": {},
   "outputs": [],
   "source": [
    "openai.api_key = 'xx-xxx'"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 5,
   "id": "a03efb7e",
   "metadata": {},
   "outputs": [],
   "source": [
    "positive_examples = openai.ChatCompletion.create(\n",
    "    model=\"gpt-3.5-turbo\",\n",
    "    messages=[\n",
    "        {\"role\": \"system\", \"content\": \"You watched a movie you loved.\"},\n",
    "        {\"role\": \"user\", \"content\": \"Write a short, 100-words review about this movie\"},\n",
    "    ],\n",
    "    max_tokens=128,\n",
    "    temperature=0.5,\n",
    "    n=3,\n",
    ")"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 6,
   "id": "1827e1b8",
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "\n",
      "\n",
      "Generated sentence 1: \n",
      "\n",
      "I recently watched the movie \"Inception\" and was blown away by its intricate plot and stunning visuals. The film follows a team of skilled thieves who enter people's dreams to steal their secrets. The concept of dream-sharing is fascinating and the execution of the idea is flawless. The cast, led by Leonardo DiCaprio, delivers outstanding performances that add depth to the characters. The action scenes are thrilling and the special effects are mind-bending. The film's score by Hans Zimmer is also noteworthy, adding to the overall immersive experience. \"Inception\" is a masterpiece that will leave you pondering its themes long after the credits roll.\n",
      "\n",
      "\n",
      "Generated sentence 2: \n",
      "\n",
      "I recently watched the movie \"The Shawshank Redemption\" and absolutely loved it. The story follows the life of a man named Andy Dufresne, who is wrongfully convicted of murder and sent to Shawshank prison. The movie beautifully portrays the struggles and hardships faced by prisoners, and the importance of hope and friendship in such a harsh environment. The acting by Tim Robbins and Morgan Freeman is outstanding, and the plot twists keep you engaged throughout the movie. Overall, \"The Shawshank Redemption\" is a must-watch for anyone who loves a good drama and a heartwarming story about the power of the human spirit.\n",
      "\n",
      "\n",
      "Generated sentence 3: \n",
      "\n",
      "I recently watched the movie \"Parasite\" and it blew me away. The story revolves around a poor family who slowly infiltrates the lives of a wealthy family, but things take a dark turn. The movie is a masterclass in storytelling, with each scene building tension and adding layers to the plot. The acting is superb, with standout performances from the entire cast. The cinematography is also stunning, with each shot expertly crafted to enhance the mood and atmosphere of the film. \"Parasite\" is a must-watch for anyone who loves a good thriller with a twist.\n"
     ]
    }
   ],
   "source": [
    "for i in range(len(positive_examples['choices'])):\n",
    "    print(f'\\n\\nGenerated sentence {i+1}: \\n')\n",
    "    print(positive_examples['choices'][i]['message']['content'])"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 4,
   "id": "b287a6f6",
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "\n",
      "\n",
      "Generated sentence 1: \n",
      "\n",
      "I recently watched a movie that left me feeling disappointed and frustrated. The plot was weak and predictable, and the characters were one-dimensional and unrelatable. The acting was subpar, with wooden performances and lackluster chemistry between the cast. The special effects were underwhelming and failed to add any excitement or visual interest to the film. Overall, I found myself checking the time and counting down the minutes until the end. I wouldn't recommend this movie to anyone looking for a compelling and engaging cinematic experience.\n",
      "\n",
      "\n",
      "Generated sentence 2: \n",
      "\n",
      "I recently watched a movie that left me feeling disappointed and underwhelmed. The plot was predictable and lacked any real depth or complexity. The characters were one-dimensional and unrelatable, making it hard to invest in their stories. The pacing was slow and dragged on unnecessarily, making the already dull plot even more tedious to sit through. The acting was subpar, with even the most talented actors failing to bring any life to their roles. Overall, I found this movie to be a complete waste of time and would not recommend it to anyone looking for an engaging and entertaining film experience.\n",
      "\n",
      "\n",
      "Generated sentence 3: \n",
      "\n",
      "I recently watched a movie that I absolutely hated - \"The Roommate\". The plot was predictable and the acting was subpar at best. The characters were one-dimensional and lacked any depth or development throughout the film. The dialogue was cringe-worthy and the attempts at suspense fell flat. Overall, I found the movie to be a waste of time and would not recommend it to anyone. If you're looking for a thrilling and well-crafted thriller, \"The Roommate\" is definitely not the movie for you.\n"
     ]
    }
   ],
   "source": [
    "# Generate the generated examples\n",
    "negative_examples = openai.ChatCompletion.create(\n",
    "    model=\"gpt-3.5-turbo\",\n",
    "    messages=[\n",
    "        {\"role\": \"system\", \"content\": \"You watched a movie you hated.\"},\n",
    "        {\"role\": \"user\", \"content\": \"Write a short, 100-words review about this movie\"},\n",
    "    ],\n",
    "    max_tokens=128,\n",
    "    temperature=0.5,\n",
    "    n=3,\n",
    ")\n",
    "# Display the generated examples\n",
    "for i in range(len(negative_examples['choices'])):\n",
    "    print(f'\\n\\nGenerated sentence {i+1}: \\n')\n",
    "    print(negative_examples['choices'][i]['message']['content'])"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "3ac1e08a",
   "metadata": {},
   "source": [
    "### There's more"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 22,
   "id": "eca062c9",
   "metadata": {},
   "outputs": [],
   "source": [
    "# Import the pipeline from transformers\n",
    "from transformers import pipeline\n",
    "# Instantiate a text-generation pipeline based on GPT-2 model\n",
    "generator = pipeline('text-generation', model='gpt2')\n",
    "#generator = pipeline('text2text-generation')#, model = 'facebook/mbart-large-50')"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 28,
   "id": "d04dbee9",
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "\n",
      "\n",
      "Generated sentence 1: \n",
      "\n",
      "I loved everything about this movie,  the movie itself was great and I didn't want to stop after the first movie because I was so happy for it. But I did like the first movie and I loved the second one and so much of what I did with it and other things with how I handled the other films.\n",
      "What were your thoughts as fans?\n",
      "After the opening credits, I kept thinking, \"This will be my big, special moment, I can't do this alone\". I've watched the original 2 episodes, they do have a big difference in the ending. But really, it's just my dream to put in movies you love, and I'm always looking over every minute on how good they were and looking forward.\n",
      "I think because everyone is talking about it, and so much of the movie is just it of my dream, the movie that I am to get this, the movie that I want to watch or maybe I'm doing my dreams.\n",
      "Also, it's more to a big than your dream. I think if I get enough reviews, people will start to see it. There will even be a few who will see it in the summer.\n",
      "I know I did it in a way when I was in high school. I never read the reviews again because I'm not a gamer and I never read the reviews because I love that. But by watching the movies, I feel that I am better than that. So, my dream is going to be to give up on the game I started out with in high school and focus on the movies I like. Not just the one I played, but the one you all enjoy watching!\n",
      "I really hope you guys give it more attention and like for this one, just keep going or stay up for the next movie for when all the things you said can be true.\n",
      "Thanks in advance, and happy movie watching!\n",
      "\n",
      "\n",
      "Generated sentence 2: \n",
      "\n",
      "I loved everything about this movie,  It was a surprise to see. I want to say thank to the cast of the film, but don't call me the original star. I love that I have to keep myself on top of the world in other things. (laughs) I was excited about the ending and I was thinking about how much fun it would be to watch that ending. At the end of the day it was all for me. The movie was a shock to watch. It was all about the fact that he and her father can all die. It was so exciting.\n",
      "Says a fan, \"I've been waiting for this movie since childhood, and this is the first time I've seen it.\"\n",
      "\n",
      "\n",
      "Generated sentence 3: \n",
      "\n",
      "I loved everything about this movie,  so I made the only mistake I have ever made because for once it felt like this movie was happening. It's always exciting to see a feature that gives the fans something to feel. It's a truly beautiful world in which life isn't a game; life is a process. But it's fun to be forced to watch something that tells you some great things about our environment, even when only one person actually is there, who cares about it. This film was not just another film, it was a true movie. And while I'm still looking forward to seeing more amazing, unique movies from the history of cinema, I can guarantee you that there's more we'll be hearing about from our friends at AMC and others who care about our history, the history of film making, and the history of art-design in general...\n"
     ]
    }
   ],
   "source": [
    "# Generate the positive examples\n",
    "positive_example = generator(\n",
    "    \"I loved everything about this movie, \",\n",
    "    max_length = 512,\n",
    "    num_return_sequences=3,\n",
    ")\n",
    "# Display the generated examples\n",
    "for i in range(len(positive_example)):\n",
    "    print(f'\\n\\nGenerated sentence {i+1}: \\n')\n",
    "    print(positive_example[i]['generated_text'])"
   ]
  }
 ],
 "metadata": {
  "kernelspec": {
   "display_name": "Python 3 (ipykernel)",
   "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": "3.9.16"
  },
  "varInspector": {
   "cols": {
    "lenName": 16,
    "lenType": 16,
    "lenVar": 40
   },
   "kernels_config": {
    "python": {
     "delete_cmd_postfix": "",
     "delete_cmd_prefix": "del ",
     "library": "var_list.py",
     "varRefreshCmd": "print(var_dic_list())"
    },
    "r": {
     "delete_cmd_postfix": ") ",
     "delete_cmd_prefix": "rm(",
     "library": "var_list.r",
     "varRefreshCmd": "cat(var_dic_list()) "
    }
   },
   "types_to_exclude": [
    "module",
    "function",
    "builtin_function_or_method",
    "instance",
    "_Feature"
   ],
   "window_display": false
  }
 },
 "nbformat": 4,
 "nbformat_minor": 5
}