{
 "cells": [
  {
   "cell_type": "markdown",
   "id": "f0dc396f",
   "metadata": {},
   "source": [
    "### TSDAE: Fine-tune sentence transformers using unsupervised learning with Pytorch\n",
    "https://www.sbert.net/examples/unsupervised_learning/TSDAE/README.html"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 2,
   "id": "34329058",
   "metadata": {
    "_cell_guid": "b1076dfc-b9ad-4769-8c92-a6c4dae69d19",
    "_uuid": "8f2839f25d086af736a60e9eeb907d3b93b6e0e5",
    "execution": {
     "iopub.execute_input": "2024-01-13T16:51:37.887254Z",
     "iopub.status.busy": "2024-01-13T16:51:37.886390Z",
     "iopub.status.idle": "2024-01-13T16:51:37.891827Z",
     "shell.execute_reply": "2024-01-13T16:51:37.890706Z",
     "shell.execute_reply.started": "2024-01-13T16:51:37.887212Z"
    }
   },
   "outputs": [],
   "source": [
    "# !pip install sentence_transformers==2.2.2"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 4,
   "id": "ebd138d8",
   "metadata": {
    "execution": {
     "iopub.execute_input": "2024-01-13T16:51:50.310221Z",
     "iopub.status.busy": "2024-01-13T16:51:50.309586Z",
     "iopub.status.idle": "2024-01-13T16:51:50.315850Z",
     "shell.execute_reply": "2024-01-13T16:51:50.314927Z",
     "shell.execute_reply.started": "2024-01-13T16:51:50.310185Z"
    }
   },
   "outputs": [],
   "source": [
    "import pandas as pd\n",
    "import numpy as np\n",
    "import string\n",
    "from tqdm import tqdm\n",
    "from numpy.linalg import norm\n",
    "from sentence_transformers import SentenceTransformer, LoggingHandler\n",
    "from sentence_transformers import models, util, datasets, evaluation, losses\n",
    "from torch.utils.data import DataLoader"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 6,
   "id": "7ef2d063",
   "metadata": {
    "execution": {
     "iopub.execute_input": "2024-01-13T13:24:44.770211Z",
     "iopub.status.busy": "2024-01-13T13:24:44.769806Z",
     "iopub.status.idle": "2024-01-13T13:24:44.775042Z",
     "shell.execute_reply": "2024-01-13T13:24:44.773860Z",
     "shell.execute_reply.started": "2024-01-13T13:24:44.770177Z"
    }
   },
   "outputs": [],
   "source": [
    "# import nltk\n",
    "# nltk.download('punkt')"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 5,
   "id": "453c1add",
   "metadata": {
    "execution": {
     "iopub.execute_input": "2024-01-13T16:51:54.070689Z",
     "iopub.status.busy": "2024-01-13T16:51:54.069945Z",
     "iopub.status.idle": "2024-01-13T16:51:54.809726Z",
     "shell.execute_reply": "2024-01-13T16:51:54.808920Z",
     "shell.execute_reply.started": "2024-01-13T16:51:54.070657Z"
    }
   },
   "outputs": [],
   "source": [
    "data = pd.read_csv('news_processed.csv', usecols=['short_description'])"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 6,
   "id": "61629b79",
   "metadata": {
    "execution": {
     "iopub.execute_input": "2024-01-13T16:51:55.125758Z",
     "iopub.status.busy": "2024-01-13T16:51:55.124990Z",
     "iopub.status.idle": "2024-01-13T16:51:55.180470Z",
     "shell.execute_reply": "2024-01-13T16:51:55.179559Z",
     "shell.execute_reply.started": "2024-01-13T16:51:55.125716Z"
    }
   },
   "outputs": [],
   "source": [
    "data.dropna(inplace=True)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 7,
   "id": "14162e2c",
   "metadata": {
    "execution": {
     "iopub.execute_input": "2024-01-13T16:51:55.817535Z",
     "iopub.status.busy": "2024-01-13T16:51:55.816764Z",
     "iopub.status.idle": "2024-01-13T16:51:55.836578Z",
     "shell.execute_reply": "2024-01-13T16:51:55.835697Z",
     "shell.execute_reply.started": "2024-01-13T16:51:55.817499Z"
    }
   },
   "outputs": [
    {
     "data": {
      "text/plain": [
       "'Experimental coronavirus vaccine prevents severe disease in mice'"
      ]
     },
     "execution_count": 7,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "data['short_description'][1000]"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 14,
   "id": "8fd1a801",
   "metadata": {
    "execution": {
     "iopub.execute_input": "2024-01-13T17:08:09.906826Z",
     "iopub.status.busy": "2024-01-13T17:08:09.906182Z",
     "iopub.status.idle": "2024-01-13T17:08:09.914834Z",
     "shell.execute_reply": "2024-01-13T17:08:09.913737Z",
     "shell.execute_reply.started": "2024-01-13T17:08:09.906795Z"
    }
   },
   "outputs": [],
   "source": [
    "def finetune_model(data: pd.DataFrame, col_to_use: str='short_description', \n",
    "                   model_id: str=\"sentence-transformers/multi-qa-distilbert-cos-v1\", \n",
    "                   batch_size: int=8, epochs: int=2):\n",
    "    \n",
    "#     https://www.sbert.net/examples/unsupervised_learning/TSDAE/README.html\n",
    "    \n",
    "    word_embedding_model = models.Transformer(model_id)\n",
    "    pooling_model = models.Pooling(word_embedding_model.get_word_embedding_dimension(), 'cls')\n",
    "    model = SentenceTransformer(modules=[word_embedding_model, pooling_model])\n",
    "    \n",
    "    train_examples = data[col_to_use].tolist()\n",
    "    train_dataset = datasets.DenoisingAutoEncoderDataset(train_examples)\n",
    "    train_dataloader = DataLoader(train_dataset, batch_size=batch_size, shuffle=True)\n",
    "    train_loss = losses.DenoisingAutoEncoderLoss(model, decoder_name_or_path=\"sentence-transformers/paraphrase-distilroberta-base-v2\", tie_encoder_decoder=False)\n",
    "#     train_loss = losses.DenoisingAutoEncoderLoss(model, decoder_name_or_path=model_id, tie_encoder_decoder=True)\n",
    "\n",
    "    \n",
    "    model.fit(\n",
    "    train_objectives=[(train_dataloader, train_loss)],\n",
    "    epochs=epochs,\n",
    "    weight_decay=0,\n",
    "    scheduler='constantlr',\n",
    "    optimizer_params={'lr': 3e-5},\n",
    "    show_progress_bar=True\n",
    "    )\n",
    "    model_save_path = model_id + '_finetuned'\n",
    "    model.save(model_save_path)\n",
    "    return model_save_path"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "id": "b17a9779",
   "metadata": {
    "scrolled": true
   },
   "outputs": [],
   "source": [
    "# fine-tune sentence transformer\n",
    "finetuned_model_id = finetune_model(data=data)\n",
    "finetuned_model = SentenceTransformer(finetuned_model_id)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "id": "4d87c128",
   "metadata": {},
   "outputs": [],
   "source": []
  }
 ],
 "metadata": {
  "kaggle": {
   "accelerator": "gpu",
   "dataSources": [
    {
     "datasetId": 4298708,
     "sourceId": 7394110,
     "sourceType": "datasetVersion"
    }
   ],
   "dockerImageVersionId": 30636,
   "isGpuEnabled": true,
   "isInternetEnabled": true,
   "language": "python",
   "sourceType": "notebook"
  },
  "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": "3.10.12"
  }
 },
 "nbformat": 4,
 "nbformat_minor": 5
}