demo.ipynb

{
 "cells": [
  {
   "cell_type": "code",
   "execution_count": 1,
   "id": "efdf9c5b-db65-4477-9c37-754be71cefe0",
   "metadata": {
    "tags": []
   },
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "\n",
      "\u001b[1m[\u001b[0m\u001b[34;49mnotice\u001b[0m\u001b[1;39;49m]\u001b[0m\u001b[39;49m A new release of pip available: \u001b[0m\u001b[31;49m22.2.2\u001b[0m\u001b[39;49m -> \u001b[0m\u001b[32;49m23.3.2\u001b[0m\n",
      "\u001b[1m[\u001b[0m\u001b[34;49mnotice\u001b[0m\u001b[1;39;49m]\u001b[0m\u001b[39;49m To update, run: \u001b[0m\u001b[32;49mpip install --upgrade pip\u001b[0m\n"
     ]
    }
   ],
   "source": [
    "# install required libraries and packages\n",
    "!pip install -qr requirements.txt"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 2,
   "id": "d41600ce-9dcb-4c9c-adf6-aa6bc8035877",
   "metadata": {
    "tags": []
   },
   "outputs": [],
   "source": [
    "# common\n",
    "import os\n",
    "import random\n",
    "import numpy as np\n",
    "\n",
    "# viz\n",
    "from IPython.display import display\n",
    "\n",
    "# local\n",
    "from ig_enabler import (\n",
    "    ig,\n",
    "    visualize\n",
    ")\n",
    "\n",
    "# third party \n",
    "from transformers import (\n",
    "    AutoModelForSequenceClassification, \n",
    "    AutoTokenizer\n",
    ")\n",
    "from datasets import load_dataset"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "b30fcbaf-5f18-4d91-9c96-695ae89557e3",
   "metadata": {},
   "source": [
    "## Objective\n",
    "\n",
    "Determine which words in an input sentence contribute the most to a LLM sentiment classification prediction using Integrated Gradients."
   ]
  },
  {
   "cell_type": "markdown",
   "id": "c3ecbca0-a956-4b71-ad23-780291a5c52c",
   "metadata": {},
   "source": [
    "## Load dataset and sample"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 3,
   "id": "317c1836-dce6-4d0d-adac-144e6bef578c",
   "metadata": {
    "tags": []
   },
   "outputs": [],
   "source": [
    "def get_rand_sample(dataset):\n",
    "    sample_idx = random.randint(0, len(dataset))\n",
    "    sample = dataset.select([sample_idx])\n",
    "    review = sample['text'][0]\n",
    "    label = sample['label'][0]\n",
    "    return review, label"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 4,
   "id": "5999b8e2-ac98-4151-97f4-d412e073c3b7",
   "metadata": {
    "tags": []
   },
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "Review: Wow. What a wonderful film. The script is nearly perfect it appears this is the only film written by Minglun Wei,I hope he has more stories in him.<br /><br />The acting is sublime. Renying Zhou as Doggie was amazing -- very natural talent, and Xu Zhu was a delight - very believable as the jaded old traditionalist. <br /><br />The soundtrack was very effective, guiding without being overwhelming. <br /><br />If only more movies like this were made whether in Hollywood or Hong Kong- a family friendly, well acted, well written, well directed, near perfect gem.\n",
      "Label: 1\n"
     ]
    }
   ],
   "source": [
    "dataset = load_dataset('imdb')\n",
    "review, label = get_rand_sample(dataset['train'])\n",
    "print(f'Review: {review}\\nLabel: {label}')"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "fbcd8735-3b23-49ba-bcfc-a746f4403d74",
   "metadata": {
    "tags": []
   },
   "source": [
    "## Load finetuned model and tokenizer"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 6,
   "id": "7f0a123b-933e-4d3f-89c9-8b797133925d",
   "metadata": {
    "tags": []
   },
   "outputs": [],
   "source": [
    "tokenizer = AutoTokenizer.from_pretrained(\"lvwerra/distilbert-imdb\")\n",
    "model = AutoModelForSequenceClassification.from_pretrained(\"lvwerra/distilbert-imdb\")"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 5,
   "id": "b9fd28dc-1f77-452b-b149-d7b97506818a",
   "metadata": {
    "tags": []
   },
   "outputs": [],
   "source": [
    "def sigmoid(logits):\n",
    "    return 1.0 / (1.0 + np.exp(-logits))"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 7,
   "id": "3274e21f-4ebb-4c4a-9ade-173b1e39f007",
   "metadata": {
    "tags": []
   },
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "label: POSITIVE, prob: 0.9452095627784729\n"
     ]
    }
   ],
   "source": [
    "model_inputs = tokenizer(review, return_tensors=\"pt\", truncation=True)\n",
    "model_outputs = model(**model_inputs)\n",
    "\n",
    "# extract logits\n",
    "logits = model_outputs[\"logits\"][0]\n",
    "logits = logits.detach().numpy()\n",
    "output = sigmoid(logits)\n",
    "pred, prob = np.argmax(output), max(output)\n",
    "\n",
    "label = \"POSITIVE\" if pred==1 else \"NEGATIVE\"\n",
    "print(f\"label: {label}, prob: {prob}\")"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "8e9926ac-1a0a-4e70-934f-879314ffadd4",
   "metadata": {},
   "source": [
    "## Integrated gradients\n",
    "\n",
    "Allows us to examine how much each input feature contributes to a LLMS's output by assigning importance scores to each input feature. \n",
    "\n",
    "1. Start with a baseline or a \"neutral\" input\n",
    "2. Define a certain number of steps that we need to take in order to get from baseline to the actual input of the model\n",
    "3. In each step, we have the model output a prediction\n",
    "4. Take the gradient of the prediction w.r.t the input\n",
    "5. Accumlate the gradients and take the integral w.r.t the output"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 8,
   "id": "fba6f7d7-d0d0-404f-8a26-2166b264f6d4",
   "metadata": {
    "tags": []
   },
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "step: 0/5\n",
      "input embedding: tensor([[[0., 0., 0.,  ..., 0., 0., 0.],\n",
      "         [0., 0., 0.,  ..., 0., 0., 0.],\n",
      "         [0., 0., 0.,  ..., 0., 0., 0.],\n",
      "         ...,\n",
      "         [0., 0., 0.,  ..., 0., 0., 0.],\n",
      "         [0., 0., 0.,  ..., 0., 0., 0.],\n",
      "         [0., 0., 0.,  ..., 0., 0., 0.]]])\n",
      "step: 1/5\n",
      "input embedding: tensor([[[ 0.0082, -0.0025, -0.0044,  ...,  0.0126,  0.0043,  0.0051],\n",
      "         [-0.0042, -0.0035, -0.0221,  ..., -0.0016,  0.0020, -0.0112],\n",
      "         [-0.0044, -0.0027, -0.0014,  ...,  0.0016,  0.0006, -0.0005],\n",
      "         ...,\n",
      "         [-0.0032, -0.0146, -0.0074,  ..., -0.0235, -0.0099, -0.0182],\n",
      "         [-0.0044, -0.0027, -0.0014,  ...,  0.0016,  0.0006, -0.0005],\n",
      "         [-0.0034, -0.0021, -0.0024,  ..., -0.0045,  0.0011, -0.0015]]])\n",
      "step: 2/5\n",
      "input embedding: tensor([[[ 0.0165, -0.0050, -0.0088,  ...,  0.0251,  0.0085,  0.0101],\n",
      "         [-0.0084, -0.0071, -0.0442,  ..., -0.0033,  0.0040, -0.0225],\n",
      "         [-0.0087, -0.0055, -0.0029,  ...,  0.0031,  0.0012, -0.0009],\n",
      "         ...,\n",
      "         [-0.0063, -0.0292, -0.0148,  ..., -0.0471, -0.0197, -0.0364],\n",
      "         [-0.0087, -0.0055, -0.0029,  ...,  0.0031,  0.0012, -0.0009],\n",
      "         [-0.0069, -0.0043, -0.0048,  ..., -0.0091,  0.0022, -0.0029]]])\n",
      "step: 3/5\n",
      "input embedding: tensor([[[ 0.0247, -0.0075, -0.0132,  ...,  0.0377,  0.0128,  0.0152],\n",
      "         [-0.0126, -0.0106, -0.0663,  ..., -0.0049,  0.0059, -0.0337],\n",
      "         [-0.0131, -0.0082, -0.0043,  ...,  0.0047,  0.0018, -0.0014],\n",
      "         ...,\n",
      "         [-0.0095, -0.0438, -0.0221,  ..., -0.0706, -0.0296, -0.0546],\n",
      "         [-0.0131, -0.0082, -0.0043,  ...,  0.0047,  0.0018, -0.0014],\n",
      "         [-0.0103, -0.0064, -0.0072,  ..., -0.0136,  0.0033, -0.0044]]])\n",
      "step: 4/5\n",
      "input embedding: tensor([[[ 0.0329, -0.0100, -0.0176,  ...,  0.0502,  0.0170,  0.0203],\n",
      "         [-0.0168, -0.0142, -0.0883,  ..., -0.0065,  0.0079, -0.0450],\n",
      "         [-0.0175, -0.0109, -0.0058,  ...,  0.0062,  0.0024, -0.0019],\n",
      "         ...,\n",
      "         [-0.0126, -0.0584, -0.0295,  ..., -0.0941, -0.0395, -0.0728],\n",
      "         [-0.0175, -0.0109, -0.0058,  ...,  0.0062,  0.0024, -0.0019],\n",
      "         [-0.0138, -0.0086, -0.0096,  ..., -0.0181,  0.0044, -0.0059]]])\n",
      "step: 5/5\n",
      "input embedding: tensor([[[ 0.0412, -0.0125, -0.0220,  ...,  0.0628,  0.0213,  0.0253],\n",
      "         [-0.0210, -0.0177, -0.1104,  ..., -0.0081,  0.0099, -0.0562],\n",
      "         [-0.0218, -0.0137, -0.0072,  ...,  0.0078,  0.0031, -0.0023],\n",
      "         ...,\n",
      "         [-0.0158, -0.0730, -0.0369,  ..., -0.1177, -0.0493, -0.0910],\n",
      "         [-0.0218, -0.0137, -0.0072,  ...,  0.0078,  0.0031, -0.0023],\n",
      "         [-0.0172, -0.0107, -0.0120,  ..., -0.0227,  0.0054, -0.0073]]])\n"
     ]
    }
   ],
   "source": [
    "ig_enabled_model = ig.IGWrapper(model)\n",
    "\n",
    "scores = ig_enabled_model(model_inputs['input_ids'],\n",
    "                          model_inputs['attention_mask'],\n",
    "                          baseline=None, \n",
    "                          num_steps=5)\n",
    "\n",
    "scores = scores.tolist()[0]"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "11739813-810f-40df-a063-74ccad0f98de",
   "metadata": {},
   "source": [
    "## Visualize top tokens"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 9,
   "id": "735497da-951b-4ad3-9bd8-f020a2b5daa4",
   "metadata": {
    "tags": []
   },
   "outputs": [
    {
     "data": {
      "text/html": [
       " <span style='color:rgb(110,163,110)'>[CLS]</span> <span style='color:rgb(95,193,95)'>wow</span> <span style='color:rgb(123,138,123)'>.</span> <span style='color:rgb(116,151,116)'>what</span> <span style='color:rgb(123,138,123)'>a</span> <span style='color:rgb(92,200,92)'>wonderful</span> <span style='color:rgb(115,153,115)'>film</span> <span style='color:rgb(125,134,125)'>.</span> <span style='color:rgb(126,132,126)'>the</span> <span style='color:rgb(108,168,108)'>script</span> <span style='color:rgb(121,142,121)'>is</span> <span style='color:rgb(115,154,115)'>nearly</span> <span style='color:rgb(89,206,89)'>perfect</span> <span style='color:rgb(113,157,113)'>it</span> <span style='color:rgb(101,181,101)'>appears</span> <span style='color:rgb(117,149,117)'>this</span> <span style='color:rgb(121,141,121)'>is</span> <span style='color:rgb(124,135,124)'>the</span> <span style='color:rgb(113,157,113)'>only</span> <span style='color:rgb(119,145,119)'>film</span> <span style='color:rgb(122,140,122)'>written</span> <span style='color:rgb(125,134,125)'>by</span> <span style='color:rgb(122,139,122)'>ming</span> <span style='color:rgb(122,140,122)'>##lun</span> <span style='color:rgb(119,145,119)'>wei</span> <span style='color:rgb(126,131,126)'>,</span> <span style='color:rgb(124,135,124)'>i</span> <span style='color:rgb(120,143,120)'>hope</span> <span style='color:rgb(125,133,125)'>he</span> <span style='color:rgb(125,134,125)'>has</span> <span style='color:rgb(125,134,125)'>more</span> <span style='color:rgb(120,144,120)'>stories</span> <span style='color:rgb(127,130,127)'>in</span> <span style='color:rgb(126,132,126)'>him</span> <span style='color:rgb(126,131,126)'>.</span> <span style='color:rgb(126,131,126)'><</span> <span style='color:rgb(128,128,128)'>br</span> <span style='color:rgb(128,128,128)'>/</span> <span style='color:rgb(127,130,127)'>></span> <span style='color:rgb(127,130,127)'><</span> <span style='color:rgb(128,128,128)'>br</span> <span style='color:rgb(128,128,128)'>/</span> <span style='color:rgb(125,134,125)'>></span> <span style='color:rgb(126,131,126)'>the</span> <span style='color:rgb(112,160,112)'>acting</span> <span style='color:rgb(116,151,116)'>is</span> <span style='color:rgb(64,255,64)'>sublime</span> <span style='color:rgb(121,142,121)'>.</span> <span style='color:rgb(119,145,119)'>ren</span> <span style='color:rgb(117,149,117)'>##ying</span> <span style='color:rgb(103,178,103)'>zhou</span> <span style='color:rgb(124,136,124)'>as</span> <span style='color:rgb(120,144,120)'>dogg</span> <span style='color:rgb(121,141,121)'>##ie</span> <span style='color:rgb(123,138,123)'>was</span> <span style='color:rgb(111,162,111)'>amazing</span> <span style='color:rgb(125,133,125)'>-</span> <span style='color:rgb(125,133,125)'>-</span> <span style='color:rgb(122,139,122)'>very</span> <span style='color:rgb(116,151,116)'>natural</span> <span style='color:rgb(113,157,113)'>talent</span> <span style='color:rgb(126,131,126)'>,</span> <span style='color:rgb(124,135,124)'>and</span> <span style='color:rgb(116,151,116)'>xu</span> <span style='color:rgb(117,150,117)'>zhu</span> <span style='color:rgb(124,136,124)'>was</span> <span style='color:rgb(125,134,125)'>a</span> <span style='color:rgb(100,184,100)'>delight</span> <span style='color:rgb(125,133,125)'>-</span> <span style='color:rgb(123,138,123)'>very</span> <span style='color:rgb(122,140,122)'>bel</span> <span style='color:rgb(124,135,124)'>##ie</span> <span style='color:rgb(116,152,116)'>##vable</span> <span style='color:rgb(125,134,125)'>as</span> <span style='color:rgb(127,129,127)'>the</span> <span style='color:rgb(121,141,121)'>jade</span> <span style='color:rgb(126,131,126)'>##d</span> <span style='color:rgb(124,136,124)'>old</span> <span style='color:rgb(122,140,122)'>traditional</span> <span style='color:rgb(124,135,124)'>##ist</span> <span style='color:rgb(126,132,126)'>.</span> <span style='color:rgb(127,130,127)'><</span> <span style='color:rgb(128,128,128)'>br</span> <span style='color:rgb(127,128,128)'>/</span> <span style='color:rgb(127,129,127)'>></span> <span style='color:rgb(126,131,126)'><</span> <span style='color:rgb(128,128,128)'>br</span> <span style='color:rgb(128,128,128)'>/</span> <span style='color:rgb(125,133,125)'>></span> <span style='color:rgb(125,133,125)'>the</span> <span style='color:rgb(110,163,110)'>soundtrack</span> <span style='color:rgb(114,155,114)'>was</span> <span style='color:rgb(118,147,118)'>very</span> <span style='color:rgb(92,200,92)'>effective</span> <span style='color:rgb(122,139,122)'>,</span> <span style='color:rgb(105,173,105)'>guiding</span> <span style='color:rgb(112,159,112)'>without</span> <span style='color:rgb(122,140,122)'>being</span> <span style='color:rgb(106,172,106)'>overwhelming</span> <span style='color:rgb(122,139,122)'>.</span> <span style='color:rgb(125,134,125)'><</span> <span style='color:rgb(127,129,127)'>br</span> <span style='color:rgb(128,128,128)'>/</span> <span style='color:rgb(126,131,126)'>></span> <span style='color:rgb(126,131,126)'><</span> <span style='color:rgb(128,128,128)'>br</span> <span style='color:rgb(128,128,128)'>/</span> <span style='color:rgb(123,138,123)'>></span> <span style='color:rgb(112,160,112)'>if</span> <span style='color:rgb(115,154,115)'>only</span> <span style='color:rgb(120,143,120)'>more</span> <span style='color:rgb(117,149,117)'>movies</span> <span style='color:rgb(115,153,115)'>like</span> <span style='color:rgb(121,142,121)'>this</span> <span style='color:rgb(120,144,120)'>were</span> <span style='color:rgb(122,140,122)'>made</span> <span style='color:rgb(110,164,110)'>whether</span> <span style='color:rgb(123,138,123)'>in</span> <span style='color:rgb(116,151,116)'>hollywood</span> <span style='color:rgb(120,144,120)'>or</span> <span style='color:rgb(112,159,112)'>hong</span> <span style='color:rgb(112,159,112)'>kong</span> <span style='color:rgb(120,144,120)'>-</span> <span style='color:rgb(120,144,120)'>a</span> <span style='color:rgb(118,148,118)'>family</span> <span style='color:rgb(112,160,112)'>friendly</span> <span style='color:rgb(124,136,124)'>,</span> <span style='color:rgb(113,157,113)'>well</span> <span style='color:rgb(109,166,109)'>acted</span> <span style='color:rgb(122,139,122)'>,</span> <span style='color:rgb(113,157,113)'>well</span> <span style='color:rgb(117,149,117)'>written</span> <span style='color:rgb(123,137,123)'>,</span> <span style='color:rgb(116,152,116)'>well</span> <span style='color:rgb(114,155,114)'>directed</span> <span style='color:rgb(119,146,119)'>,</span> <span style='color:rgb(107,169,107)'>near</span> <span style='color:rgb(100,184,100)'>perfect</span> <span style='color:rgb(81,221,81)'>gem</span> <span style='color:rgb(117,149,117)'>.</span> <span style='color:rgb(121,141,121)'>[SEP]</span>"
      ],
      "text/plain": [
       "<IPython.core.display.HTML object>"
      ]
     },
     "metadata": {},
     "output_type": "display_data"
    },
    {
     "data": {
      "image/png": "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",
      "text/plain": [
       "<Figure size 2100x300 with 1 Axes>"
      ]
     },
     "metadata": {},
     "output_type": "display_data"
    }
   ],
   "source": [
    "tokens = tokenizer.convert_ids_to_tokens(model_inputs['input_ids'][0])\n",
    "\n",
    "display(visualize.visualize_token_scores(tokens, scores))\n",
    "\n",
    "visualize.plot_topk_scores(tokens, scores, 10)"
   ]
  }
 ],
 "metadata": {
  "kernelspec": {
   "display_name": "Python 3.9",
   "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": "3.9.18"
  }
 },
 "nbformat": 4,
 "nbformat_minor": 5
}