SentimentClassificationSSTAnalysis.ipynb

{
 "cells": [
  {
   "cell_type": "code",
   "execution_count": 1,
   "metadata": {},
   "outputs": [],
   "source": [
    "%load_ext autoreload\n",
    "%autoreload 2"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 2,
   "metadata": {},
   "outputs": [],
   "source": [
    "import os\n",
    "import argparse\n",
    "import torch\n",
    "import spacy\n",
    "import pickle\n",
    "import json\n",
    "from transformers import BertTokenizerFast\n",
    "\n",
    "import matplotlib.pyplot as plt\n",
    "from tqdm.notebook import tqdm, trange\n",
    "\n",
    "from diffmask.models.sentiment_classification_sst import (\n",
    "    BertSentimentClassificationSST,\n",
    "    RecurrentSentimentClassificationSST,\n",
    ")\n",
    "\n",
    "from diffmask.attributions.integrated_gradient import (\n",
    "    sst_bert_integrated_gradient,\n",
    "    sst_gru_integrated_gradient,\n",
    ")\n",
    "from diffmask.attributions.schulz import (\n",
    "    bert_hidden_states_statistics,\n",
    "    sst_gru_hidden_states_statistics,\n",
    "    sst_bert_schulz_explainer,\n",
    "    sst_gru_schulz_explainer,\n",
    ")\n",
    "from diffmask.attributions.guan import (\n",
    "    sst_bert_guan_explainer,\n",
    "    sst_gru_guan_explainer,\n",
    ")\n",
    "\n",
    "from diffmask.utils.getter_setter import gru_getter, gru_setter\n",
    "from diffmask.utils.plot import plot_sst_attributions"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 3,
   "metadata": {},
   "outputs": [],
   "source": [
    "if __name__ == \"__main__\":\n",
    "    parser = argparse.ArgumentParser()\n",
    "    parser.add_argument(\"--gpu\", type=str, default=\"0\")\n",
    "    parser.add_argument(\"--model\", type=str, default=\"bert-base-uncased\")\n",
    "    parser.add_argument(\"--train_filename\", type=str, default=\"./datasets/sst/train.txt\")\n",
    "    parser.add_argument(\"--val_filename\", type=str, default=\"./datasets/sst/dev.txt\")\n",
    "    parser.add_argument(\"--batch_size\", type=int, default=64)\n",
    "    parser.add_argument(\"--seed\", type=int, default=0)\n",
    "    parser.add_argument(\"--architecture\", type=str, default=\"bert\", choices=[\"gru\", \"bert\"])\n",
    "    parser.add_argument(\n",
    "        \"--model_path\",\n",
    "        type=str,\n",
    "        default=\"models/sst.ckpt\",\n",
    "    )\n",
    "    parser.add_argument(\"--eps_valid\", type=int, default=0.05)\n",
    "    \n",
    "    hparams, _ = parser.parse_known_args()\n",
    "    \n",
    "    torch.manual_seed(hparams.seed)\n",
    "    \n",
    "    os.environ[\"CUDA_VISIBLE_DEVICES\"] = hparams.gpu"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "# Loading model"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 4,
   "metadata": {
    "scrolled": false
   },
   "outputs": [],
   "source": [
    "device = \"cuda:0\"\n",
    "\n",
    "if hparams.architecture == \"bert\":\n",
    "    model = BertSentimentClassificationSST.load_from_checkpoint(hparams.model_path).to(device)\n",
    "    integrated_gradient = sst_bert_integrated_gradient\n",
    "    hidden_states_statistics = bert_hidden_states_statistics\n",
    "    schulz_explainer = sst_bert_schulz_explainer\n",
    "    guan_explainer = sst_bert_guan_explainer\n",
    "    \n",
    "elif hparams.architecture == \"gru\":\n",
    "    model = RecurrentSentimentClassificationSST.load_from_checkpoint(hparams.model_path).to(device)\n",
    "    integrated_gradient = sst_gru_integrated_gradient\n",
    "    hidden_states_statistics = sst_gru_hidden_states_statistics\n",
    "    schulz_explainer = sst_gru_schulz_explainer\n",
    "    guan_explainer = sst_gru_guan_explainer\n",
    "    \n",
    "model.freeze()\n",
    "model.prepare_data()"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "# Create example"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 5,
   "metadata": {},
   "outputs": [],
   "source": [
    "source = \"A fast , funny , highly enjoyable movie .\"\n",
    "\n",
    "inputs_dict = {\n",
    "    k: v.to(device)\n",
    "    for k, v in model.tokenizer.encode_plus(\n",
    "        source,\n",
    "        max_length=84,\n",
    "        pad_to_max_length=True,\n",
    "        return_tensors=\"pt\",\n",
    "    ).items()\n",
    "}\n",
    "inputs_dict[\"mask\"] = inputs_dict[\"attention_mask\"]\n",
    "inputs_dict[\"labels\"] = torch.tensor([4]).to(device) # very positive\n",
    "del inputs_dict[\"attention_mask\"]\n",
    "del inputs_dict[\"token_type_ids\"]\n",
    "\n",
    "source = model.tokenizer.tokenize(source)\n",
    "tokens = [\"[CLS]\"]  + source + [\"[SEP]\"]"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "# Sundararajan et al. (2017) - Integrated Gradient"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 6,
   "metadata": {
    "scrolled": true
   },
   "outputs": [
    {
     "data": {
      "application/vnd.jupyter.widget-view+json": {
       "model_id": "49b685b5e9b34645b10ae751fde4846c",
       "version_major": 2,
       "version_minor": 0
      },
      "text/plain": [
       "HBox(children=(FloatProgress(value=0.0, max=50.0), HTML(value='')))"
      ]
     },
     "metadata": {},
     "output_type": "display_data"
    },
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "\n"
     ]
    },
    {
     "data": {
      "application/vnd.jupyter.widget-view+json": {
       "model_id": "4f21d196ce774296937377f476e718b4",
       "version_major": 2,
       "version_minor": 0
      },
      "text/plain": [
       "HBox(children=(FloatProgress(value=0.0, max=50.0), HTML(value='')))"
      ]
     },
     "metadata": {},
     "output_type": "display_data"
    },
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "\n"
     ]
    },
    {
     "data": {
      "application/vnd.jupyter.widget-view+json": {
       "model_id": "1bb7b898f33642f285bafceeb36a5e33",
       "version_major": 2,
       "version_minor": 0
      },
      "text/plain": [
       "HBox(children=(FloatProgress(value=0.0, max=50.0), HTML(value='')))"
      ]
     },
     "metadata": {},
     "output_type": "display_data"
    },
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "\n"
     ]
    },
    {
     "data": {
      "application/vnd.jupyter.widget-view+json": {
       "model_id": "36c76f24949c41509f1fbc553eb9a7ba",
       "version_major": 2,
       "version_minor": 0
      },
      "text/plain": [
       "HBox(children=(FloatProgress(value=0.0, max=50.0), HTML(value='')))"
      ]
     },
     "metadata": {},
     "output_type": "display_data"
    },
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "\n"
     ]
    },
    {
     "data": {
      "application/vnd.jupyter.widget-view+json": {
       "model_id": "8e4606fdc1434ad0ba1260eb0dd1d63a",
       "version_major": 2,
       "version_minor": 0
      },
      "text/plain": [
       "HBox(children=(FloatProgress(value=0.0, max=50.0), HTML(value='')))"
      ]
     },
     "metadata": {},
     "output_type": "display_data"
    },
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "\n"
     ]
    },
    {
     "data": {
      "application/vnd.jupyter.widget-view+json": {
       "model_id": "30733f270c12462088a890dce4e610a8",
       "version_major": 2,
       "version_minor": 0
      },
      "text/plain": [
       "HBox(children=(FloatProgress(value=0.0, max=50.0), HTML(value='')))"
      ]
     },
     "metadata": {},
     "output_type": "display_data"
    },
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "\n"
     ]
    },
    {
     "data": {
      "application/vnd.jupyter.widget-view+json": {
       "model_id": "7a65e3b8c5a041b785d56ec192e42e32",
       "version_major": 2,
       "version_minor": 0
      },
      "text/plain": [
       "HBox(children=(FloatProgress(value=0.0, max=50.0), HTML(value='')))"
      ]
     },
     "metadata": {},
     "output_type": "display_data"
    },
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "\n"
     ]
    },
    {
     "data": {
      "application/vnd.jupyter.widget-view+json": {
       "model_id": "c36bf781719b4aec8b365a76ac473996",
       "version_major": 2,
       "version_minor": 0
      },
      "text/plain": [
       "HBox(children=(FloatProgress(value=0.0, max=50.0), HTML(value='')))"
      ]
     },
     "metadata": {},
     "output_type": "display_data"
    },
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "\n"
     ]
    },
    {
     "data": {
      "application/vnd.jupyter.widget-view+json": {
       "model_id": "0ba182ef60524a19b5a049a1ac5466d5",
       "version_major": 2,
       "version_minor": 0
      },
      "text/plain": [
       "HBox(children=(FloatProgress(value=0.0, max=50.0), HTML(value='')))"
      ]
     },
     "metadata": {},
     "output_type": "display_data"
    },
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "\n"
     ]
    },
    {
     "data": {
      "application/vnd.jupyter.widget-view+json": {
       "model_id": "6ef90e0834294d41b339f1d925d06c8f",
       "version_major": 2,
       "version_minor": 0
      },
      "text/plain": [
       "HBox(children=(FloatProgress(value=0.0, max=50.0), HTML(value='')))"
      ]
     },
     "metadata": {},
     "output_type": "display_data"
    },
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "\n"
     ]
    },
    {
     "data": {
      "application/vnd.jupyter.widget-view+json": {
       "model_id": "74d6eacd52024f41b4e682e86eafc67c",
       "version_major": 2,
       "version_minor": 0
      },
      "text/plain": [
       "HBox(children=(FloatProgress(value=0.0, max=50.0), HTML(value='')))"
      ]
     },
     "metadata": {},
     "output_type": "display_data"
    },
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "\n"
     ]
    },
    {
     "data": {
      "application/vnd.jupyter.widget-view+json": {
       "model_id": "ab1a51ea97274f479e603f0764135ef2",
       "version_major": 2,
       "version_minor": 0
      },
      "text/plain": [
       "HBox(children=(FloatProgress(value=0.0, max=50.0), HTML(value='')))"
      ]
     },
     "metadata": {},
     "output_type": "display_data"
    },
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "\n"
     ]
    },
    {
     "data": {
      "application/vnd.jupyter.widget-view+json": {
       "model_id": "7a4adc99141540449783e0a1ffdc0f96",
       "version_major": 2,
       "version_minor": 0
      },
      "text/plain": [
       "HBox(children=(FloatProgress(value=0.0, max=50.0), HTML(value='')))"
      ]
     },
     "metadata": {},
     "output_type": "display_data"
    },
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "\n"
     ]
    },
    {
     "data": {
      "application/vnd.jupyter.widget-view+json": {
       "model_id": "30573cb5598c442fa84cf9f5ccd571bf",
       "version_major": 2,
       "version_minor": 0
      },
      "text/plain": [
       "HBox(children=(FloatProgress(value=0.0, max=50.0), HTML(value='')))"
      ]
     },
     "metadata": {},
     "output_type": "display_data"
    },
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "\n"
     ]
    }
   ],
   "source": [
    "attributions_ig = torch.cat([integrated_gradient(\n",
    "    model,\n",
    "    inputs_dict,\n",
    "    hidden_state_idx=i,\n",
    "    steps=50,\n",
    ") for i in range(14)], 0).sum(-1).abs().T"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 7,
   "metadata": {},
   "outputs": [
    {
     "data": {
      "image/png": "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\n",
      "text/plain": [
       "<Figure size 648x264 with 1 Axes>"
      ]
     },
     "metadata": {},
     "output_type": "display_data"
    }
   ],
   "source": [
    "plot_sst_attributions(\n",
    "    attributions_ig[:inputs_dict[\"mask\"].sum(-1).item()].cpu(),\n",
    "    tokens,\n",
    ")"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "# Schulz et al. (2020)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 8,
   "metadata": {},
   "outputs": [
    {
     "data": {
      "application/vnd.jupyter.widget-view+json": {
       "model_id": "71f89a0201ad41da99ab01e6e7ca40b3",
       "version_major": 2,
       "version_minor": 0
      },
      "text/plain": [
       "HBox(children=(FloatProgress(value=0.0, max=134.0), HTML(value='')))"
      ]
     },
     "metadata": {},
     "output_type": "display_data"
    },
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "\n"
     ]
    }
   ],
   "source": [
    "all_q_z_loc, all_q_z_scale = hidden_states_statistics(model, input_only=False)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 9,
   "metadata": {
    "scrolled": true
   },
   "outputs": [
    {
     "data": {
      "application/vnd.jupyter.widget-view+json": {
       "model_id": "474275ab7b164795b2d7e3e85b71057f",
       "version_major": 2,
       "version_minor": 0
      },
      "text/plain": [
       "HBox(children=(FloatProgress(value=0.0, max=50.0), HTML(value='')))"
      ]
     },
     "metadata": {},
     "output_type": "display_data"
    },
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "\n"
     ]
    },
    {
     "data": {
      "application/vnd.jupyter.widget-view+json": {
       "model_id": "45d7667beadb4302af22d8d1ce6c25a0",
       "version_major": 2,
       "version_minor": 0
      },
      "text/plain": [
       "HBox(children=(FloatProgress(value=0.0, max=50.0), HTML(value='')))"
      ]
     },
     "metadata": {},
     "output_type": "display_data"
    },
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "\n"
     ]
    },
    {
     "data": {
      "application/vnd.jupyter.widget-view+json": {
       "model_id": "397fc79c6ad441a78f14642cdc8ba3ad",
       "version_major": 2,
       "version_minor": 0
      },
      "text/plain": [
       "HBox(children=(FloatProgress(value=0.0, max=50.0), HTML(value='')))"
      ]
     },
     "metadata": {},
     "output_type": "display_data"
    },
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "\n"
     ]
    },
    {
     "data": {
      "application/vnd.jupyter.widget-view+json": {
       "model_id": "3c56cc664d7d436ea9d34507e4743ded",
       "version_major": 2,
       "version_minor": 0
      },
      "text/plain": [
       "HBox(children=(FloatProgress(value=0.0, max=50.0), HTML(value='')))"
      ]
     },
     "metadata": {},
     "output_type": "display_data"
    },
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "\n"
     ]
    },
    {
     "data": {
      "application/vnd.jupyter.widget-view+json": {
       "model_id": "764c426137184e0e926f71a32a640b53",
       "version_major": 2,
       "version_minor": 0
      },
      "text/plain": [
       "HBox(children=(FloatProgress(value=0.0, max=50.0), HTML(value='')))"
      ]
     },
     "metadata": {},
     "output_type": "display_data"
    },
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "\n"
     ]
    },
    {
     "data": {
      "application/vnd.jupyter.widget-view+json": {
       "model_id": "1bb3e788ee3840918243fe8be948ab1a",
       "version_major": 2,
       "version_minor": 0
      },
      "text/plain": [
       "HBox(children=(FloatProgress(value=0.0, max=50.0), HTML(value='')))"
      ]
     },
     "metadata": {},
     "output_type": "display_data"
    },
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "\n"
     ]
    },
    {
     "data": {
      "application/vnd.jupyter.widget-view+json": {
       "model_id": "176025aaf7f940f2a41a5c511268c940",
       "version_major": 2,
       "version_minor": 0
      },
      "text/plain": [
       "HBox(children=(FloatProgress(value=0.0, max=50.0), HTML(value='')))"
      ]
     },
     "metadata": {},
     "output_type": "display_data"
    },
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "\n"
     ]
    },
    {
     "data": {
      "application/vnd.jupyter.widget-view+json": {
       "model_id": "4296cf70237f4962bdf5eb726f3ad566",
       "version_major": 2,
       "version_minor": 0
      },
      "text/plain": [
       "HBox(children=(FloatProgress(value=0.0, max=50.0), HTML(value='')))"
      ]
     },
     "metadata": {},
     "output_type": "display_data"
    },
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "\n"
     ]
    },
    {
     "data": {
      "application/vnd.jupyter.widget-view+json": {
       "model_id": "5738ff3abc2d4dce8cdb0e36d3f6949e",
       "version_major": 2,
       "version_minor": 0
      },
      "text/plain": [
       "HBox(children=(FloatProgress(value=0.0, max=50.0), HTML(value='')))"
      ]
     },
     "metadata": {},
     "output_type": "display_data"
    },
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "\n"
     ]
    },
    {
     "data": {
      "application/vnd.jupyter.widget-view+json": {
       "model_id": "755b6a6632874cdd83d3b95c603e4dac",
       "version_major": 2,
       "version_minor": 0
      },
      "text/plain": [
       "HBox(children=(FloatProgress(value=0.0, max=50.0), HTML(value='')))"
      ]
     },
     "metadata": {},
     "output_type": "display_data"
    },
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "\n"
     ]
    },
    {
     "data": {
      "application/vnd.jupyter.widget-view+json": {
       "model_id": "c4dc737a34334bf895b6c28a8ae11832",
       "version_major": 2,
       "version_minor": 0
      },
      "text/plain": [
       "HBox(children=(FloatProgress(value=0.0, max=50.0), HTML(value='')))"
      ]
     },
     "metadata": {},
     "output_type": "display_data"
    },
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "\n"
     ]
    },
    {
     "data": {
      "application/vnd.jupyter.widget-view+json": {
       "model_id": "b080839e9ff84943b54491e395f02ee7",
       "version_major": 2,
       "version_minor": 0
      },
      "text/plain": [
       "HBox(children=(FloatProgress(value=0.0, max=50.0), HTML(value='')))"
      ]
     },
     "metadata": {},
     "output_type": "display_data"
    },
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "\n"
     ]
    },
    {
     "data": {
      "application/vnd.jupyter.widget-view+json": {
       "model_id": "b52abefdb74c4446b64d1fa26fd27650",
       "version_major": 2,
       "version_minor": 0
      },
      "text/plain": [
       "HBox(children=(FloatProgress(value=0.0, max=50.0), HTML(value='')))"
      ]
     },
     "metadata": {},
     "output_type": "display_data"
    },
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "\n"
     ]
    },
    {
     "data": {
      "application/vnd.jupyter.widget-view+json": {
       "model_id": "ef7ee6f1809143cf85e5ccee1262909a",
       "version_major": 2,
       "version_minor": 0
      },
      "text/plain": [
       "HBox(children=(FloatProgress(value=0.0, max=50.0), HTML(value='')))"
      ]
     },
     "metadata": {},
     "output_type": "display_data"
    },
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "\n"
     ]
    }
   ],
   "source": [
    "attributions_schulz = torch.cat([schulz_explainer(\n",
    "    model,\n",
    "    inputs_dict,\n",
    "    all_q_z_loc,\n",
    "    all_q_z_scale,\n",
    "    hidden_state_idx=i,\n",
    "    steps=50,\n",
    "    lr=1e-1,\n",
    "    la=10,\n",
    ") for i in range(14)], 0).T"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 10,
   "metadata": {},
   "outputs": [
    {
     "data": {
      "image/png": "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\n",
      "text/plain": [
       "<Figure size 648x264 with 1 Axes>"
      ]
     },
     "metadata": {},
     "output_type": "display_data"
    }
   ],
   "source": [
    "plot_sst_attributions(\n",
    "    attributions_schulz[:inputs_dict[\"mask\"].sum(-1).item()].cpu(),\n",
    "    tokens,\n",
    ")"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "# Guan et al. (2019)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 11,
   "metadata": {
    "scrolled": true
   },
   "outputs": [
    {
     "data": {
      "application/vnd.jupyter.widget-view+json": {
       "model_id": "8294ca001d2b4e72a3e1019dbe5b314d",
       "version_major": 2,
       "version_minor": 0
      },
      "text/plain": [
       "HBox(children=(FloatProgress(value=0.0, max=50.0), HTML(value='')))"
      ]
     },
     "metadata": {},
     "output_type": "display_data"
    },
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "\n"
     ]
    },
    {
     "data": {
      "application/vnd.jupyter.widget-view+json": {
       "model_id": "8308f82c85c24f8898e7c7f76c71bac5",
       "version_major": 2,
       "version_minor": 0
      },
      "text/plain": [
       "HBox(children=(FloatProgress(value=0.0, max=50.0), HTML(value='')))"
      ]
     },
     "metadata": {},
     "output_type": "display_data"
    },
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "\n"
     ]
    },
    {
     "data": {
      "application/vnd.jupyter.widget-view+json": {
       "model_id": "16e9eea813b346feba9266022e2a46cc",
       "version_major": 2,
       "version_minor": 0
      },
      "text/plain": [
       "HBox(children=(FloatProgress(value=0.0, max=50.0), HTML(value='')))"
      ]
     },
     "metadata": {},
     "output_type": "display_data"
    },
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "\n"
     ]
    },
    {
     "data": {
      "application/vnd.jupyter.widget-view+json": {
       "model_id": "98dfbd34c6bf479f8009251e1ca1e298",
       "version_major": 2,
       "version_minor": 0
      },
      "text/plain": [
       "HBox(children=(FloatProgress(value=0.0, max=50.0), HTML(value='')))"
      ]
     },
     "metadata": {},
     "output_type": "display_data"
    },
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "\n"
     ]
    },
    {
     "data": {
      "application/vnd.jupyter.widget-view+json": {
       "model_id": "967f0b199c374446b70420798226bdf1",
       "version_major": 2,
       "version_minor": 0
      },
      "text/plain": [
       "HBox(children=(FloatProgress(value=0.0, max=50.0), HTML(value='')))"
      ]
     },
     "metadata": {},
     "output_type": "display_data"
    },
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "\n"
     ]
    },
    {
     "data": {
      "application/vnd.jupyter.widget-view+json": {
       "model_id": "85902e60505748f584c8735cb9c76368",
       "version_major": 2,
       "version_minor": 0
      },
      "text/plain": [
       "HBox(children=(FloatProgress(value=0.0, max=50.0), HTML(value='')))"
      ]
     },
     "metadata": {},
     "output_type": "display_data"
    },
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "\n"
     ]
    },
    {
     "data": {
      "application/vnd.jupyter.widget-view+json": {
       "model_id": "2b51b09cc0f940c1b0cd829742f99f38",
       "version_major": 2,
       "version_minor": 0
      },
      "text/plain": [
       "HBox(children=(FloatProgress(value=0.0, max=50.0), HTML(value='')))"
      ]
     },
     "metadata": {},
     "output_type": "display_data"
    },
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "\n"
     ]
    },
    {
     "data": {
      "application/vnd.jupyter.widget-view+json": {
       "model_id": "12867abc98364663a85f7e14c791c669",
       "version_major": 2,
       "version_minor": 0
      },
      "text/plain": [
       "HBox(children=(FloatProgress(value=0.0, max=50.0), HTML(value='')))"
      ]
     },
     "metadata": {},
     "output_type": "display_data"
    },
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "\n"
     ]
    },
    {
     "data": {
      "application/vnd.jupyter.widget-view+json": {
       "model_id": "eab975d6cc3a4c1690264370b8f925c3",
       "version_major": 2,
       "version_minor": 0
      },
      "text/plain": [
       "HBox(children=(FloatProgress(value=0.0, max=50.0), HTML(value='')))"
      ]
     },
     "metadata": {},
     "output_type": "display_data"
    },
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "\n"
     ]
    },
    {
     "data": {
      "application/vnd.jupyter.widget-view+json": {
       "model_id": "78c3da43118a4879acb167300e88998d",
       "version_major": 2,
       "version_minor": 0
      },
      "text/plain": [
       "HBox(children=(FloatProgress(value=0.0, max=50.0), HTML(value='')))"
      ]
     },
     "metadata": {},
     "output_type": "display_data"
    },
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "\n"
     ]
    },
    {
     "data": {
      "application/vnd.jupyter.widget-view+json": {
       "model_id": "d6391fd129a84390ad4403fd08c986f8",
       "version_major": 2,
       "version_minor": 0
      },
      "text/plain": [
       "HBox(children=(FloatProgress(value=0.0, max=50.0), HTML(value='')))"
      ]
     },
     "metadata": {},
     "output_type": "display_data"
    },
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "\n"
     ]
    },
    {
     "data": {
      "application/vnd.jupyter.widget-view+json": {
       "model_id": "623a492808204c3fbfacf1515884e828",
       "version_major": 2,
       "version_minor": 0
      },
      "text/plain": [
       "HBox(children=(FloatProgress(value=0.0, max=50.0), HTML(value='')))"
      ]
     },
     "metadata": {},
     "output_type": "display_data"
    },
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "\n"
     ]
    },
    {
     "data": {
      "application/vnd.jupyter.widget-view+json": {
       "model_id": "7e2bec98694e40a891f6d85fb1c45c69",
       "version_major": 2,
       "version_minor": 0
      },
      "text/plain": [
       "HBox(children=(FloatProgress(value=0.0, max=50.0), HTML(value='')))"
      ]
     },
     "metadata": {},
     "output_type": "display_data"
    },
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "\n"
     ]
    },
    {
     "data": {
      "application/vnd.jupyter.widget-view+json": {
       "model_id": "a060b008f096411183e1d0c88d639b57",
       "version_major": 2,
       "version_minor": 0
      },
      "text/plain": [
       "HBox(children=(FloatProgress(value=0.0, max=50.0), HTML(value='')))"
      ]
     },
     "metadata": {},
     "output_type": "display_data"
    },
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "\n"
     ]
    }
   ],
   "source": [
    "attributions_guan = - torch.cat([guan_explainer(\n",
    "    model,\n",
    "    inputs_dict,\n",
    "    hidden_state_idx=i,\n",
    "    steps=50,\n",
    "    lr=1e-1,\n",
    "    la=1e-1,\n",
    ") for i in range(14)], 0).T"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 12,
   "metadata": {},
   "outputs": [
    {
     "data": {
      "image/png": "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\n",
      "text/plain": [
       "<Figure size 648x264 with 1 Axes>"
      ]
     },
     "metadata": {},
     "output_type": "display_data"
    }
   ],
   "source": [
    "plot_sst_attributions(\n",
    "    attributions_guan[:inputs_dict[\"mask\"].sum(-1).item()].cpu(),\n",
    "    tokens,\n",
    ")"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "# Attention"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 13,
   "metadata": {},
   "outputs": [],
   "source": [
    "model.net.bert.encoder.output_attentions = True\n",
    "for l in model.net.bert.encoder.layer:\n",
    "    l.attention.self.output_attentions = True\n",
    "    \n",
    "_, attention = model(**inputs_dict)\n",
    "\n",
    "model.net.bert.encoder.output_attentions = False\n",
    "for l in model.net.bert.encoder.layer:\n",
    "    l.attention.self.output_attentions = False\n",
    "    \n",
    "attributions_att = torch.stack(\n",
    "    [e.max(1).values.mean(-2) for e in attention], -1\n",
    ")\n",
    "attributions_att = torch.cat(\n",
    "    (torch.full_like(attributions_att[...,0:1], float(\"nan\")),\n",
    "     attributions_att, \n",
    "     torch.full_like(attributions_att[...,0:1], float(\"nan\"))), -1\n",
    ")[0]"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 14,
   "metadata": {},
   "outputs": [
    {
     "data": {
      "image/png": "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\n",
      "text/plain": [
       "<Figure size 648x264 with 1 Axes>"
      ]
     },
     "metadata": {},
     "output_type": "display_data"
    }
   ],
   "source": [
    "plot_sst_attributions(\n",
    "    attributions_att[:inputs_dict[\"mask\"].sum(-1).item()].cpu(),\n",
    "    tokens,\n",
    ")"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": []
  }
 ],
 "metadata": {
  "kernelspec": {
   "display_name": "Python 3",
   "language": "python",
   "name": "python3"
  },
  "language_info": {
   "codemirror_mode": {
    "name": "ipython",
    "version": 3
   },
   "file_extension": ".py",
   "mimetype": "text/x-python",
   "name": "python",
   "nbconvert_exporter": "python",
   "pygments_lexer": "ipython3",
   "version": "3.6.10"
  }
 },
 "nbformat": 4,
 "nbformat_minor": 4
}