coar/edit/model_errors.ipynb

{
 "cells": [
  {
   "cell_type": "markdown",
   "id": "e526cb83-bae4-4639-b336-4e056debb932",
   "metadata": {},
   "source": [
    "## Changing individual predictions\n",
    "\n",
    "In this notebook, we show that Coar-Edit can change individual predictions of an ImageNet ResNet50 classifier without impacting its overall performance. This experiment reproduces our findings in Section 5.1 of our paper."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 2,
   "id": "d7657066-6a8a-4141-accb-f0052cd73338",
   "metadata": {
    "tags": []
   },
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "The autoreload extension is already loaded. To reload it, use:\n",
      "  %reload_ext autoreload\n"
     ]
    }
   ],
   "source": [
    "import os \n",
    "import sys \n",
    "sys.path.append('..')\n",
    "\n",
    "import functools \n",
    "import torch\n",
    "import torchvision\n",
    "import numpy as np\n",
    "import pandas as pd \n",
    "from torch import nn\n",
    "import seaborn as sns\n",
    "import tqdm\n",
    "from collections import defaultdict\n",
    "import matplotlib.pyplot as plt\n",
    "\n",
    "from src.models import cifar_resnet\n",
    "from src.utils import ndm_utils\n",
    "from src.utils import data_utils\n",
    "from src.utils import eval_utils\n",
    "from src.utils import plot_utils\n",
    "from src.data import ffcv_pipelines\n",
    "import coar_edit\n",
    "\n",
    "%matplotlib inline\n",
    "%load_ext autoreload\n",
    "%autoreload 2"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "d7bc2f04-98bc-4dac-b706-3224087fd7dd",
   "metadata": {
    "tags": []
   },
   "source": [
    "### Download attributions"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 14,
   "id": "0d475686-707d-48fa-8858-241df7f0e03d",
   "metadata": {
    "tags": []
   },
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "Downloading dataloader (FFCV format)\n",
      "File ‘out/imagenet_val.beton’ already there; not retrieving.\n",
      "Downloading attributions\n",
      "--2024-02-14 15:38:14--  https://www.dropbox.com/scl/fi/vd38pyjjyz1g7zu72r741/imagenet_resnet.pt?rlkey=45ideswhv9x3tyu7ctjcd84df&dl=0\n",
      "Resolving www.dropbox.com (www.dropbox.com)... 162.125.4.18, 2620:100:6019:18::a27d:412\n",
      "Connecting to www.dropbox.com (www.dropbox.com)|162.125.4.18|:443... connected.\n",
      "HTTP request sent, awaiting response... 302 Found\n",
      "Location: https://ucdbbe7c0329e385d57a26837175.dl.dropboxusercontent.com/cd/0/inline/CNT9bEzb7AfU5pq7M30q0OFyyRd3_4brDP0lm31aZ2WVZHfAa6U1UwHd6FQbIpYunw3ljyyhLA-F7P476fylTt7lgKvP1DxK06bqegeK_xGQXwxFHUMoYZIvUJvJ4xKLjBcHFcCIxtcneBLemZU9RKrG/file# [following]\n",
      "--2024-02-14 15:38:14--  https://ucdbbe7c0329e385d57a26837175.dl.dropboxusercontent.com/cd/0/inline/CNT9bEzb7AfU5pq7M30q0OFyyRd3_4brDP0lm31aZ2WVZHfAa6U1UwHd6FQbIpYunw3ljyyhLA-F7P476fylTt7lgKvP1DxK06bqegeK_xGQXwxFHUMoYZIvUJvJ4xKLjBcHFcCIxtcneBLemZU9RKrG/file\n",
      "Resolving ucdbbe7c0329e385d57a26837175.dl.dropboxusercontent.com (ucdbbe7c0329e385d57a26837175.dl.dropboxusercontent.com)... 162.125.4.15, 2620:100:6019:15::a27d:40f\n",
      "Connecting to ucdbbe7c0329e385d57a26837175.dl.dropboxusercontent.com (ucdbbe7c0329e385d57a26837175.dl.dropboxusercontent.com)|162.125.4.15|:443... connected.\n",
      "HTTP request sent, awaiting response... 302 Found\n",
      "Location: /cd/0/inline2/CNQldbSfTsvCPMdiq1Uj58Hci7xRXUSeEhjFj0pYLSMNTXJ1ca86kb41N_reViF7RJKjSGzL0GDwYw-bq1cXkYXs5DjzTqG16LOFYPRKM4FXOFzRAmmBKUs1sSDWOgkroiwhcOt-EQjZfgCQKIwjYe9kCf2hnOD_2pQYHA_LtVQaSLGzWh86dWN64DlCDnx6ZrCooLyTaArQV4pAmSQJBTKLHVm9DehwlnESiAa8BWhCnaLe1hKeg2_oeOXIPJRQIacB8ENHWx1wyuv1hzKYdL96MugCPLFD8KWbnFVvv3Q7gY0_UooRa_EGNqbuiMhjo9RikciwaZvCZqIASzt4ra7YYbsIziRfF7cZdkBBbXvlJIH1xs4-cfWpVCAQkeJgdSs/file [following]\n",
      "--2024-02-14 15:38:15--  https://ucdbbe7c0329e385d57a26837175.dl.dropboxusercontent.com/cd/0/inline2/CNQldbSfTsvCPMdiq1Uj58Hci7xRXUSeEhjFj0pYLSMNTXJ1ca86kb41N_reViF7RJKjSGzL0GDwYw-bq1cXkYXs5DjzTqG16LOFYPRKM4FXOFzRAmmBKUs1sSDWOgkroiwhcOt-EQjZfgCQKIwjYe9kCf2hnOD_2pQYHA_LtVQaSLGzWh86dWN64DlCDnx6ZrCooLyTaArQV4pAmSQJBTKLHVm9DehwlnESiAa8BWhCnaLe1hKeg2_oeOXIPJRQIacB8ENHWx1wyuv1hzKYdL96MugCPLFD8KWbnFVvv3Q7gY0_UooRa_EGNqbuiMhjo9RikciwaZvCZqIASzt4ra7YYbsIziRfF7cZdkBBbXvlJIH1xs4-cfWpVCAQkeJgdSs/file\n",
      "Reusing existing connection to ucdbbe7c0329e385d57a26837175.dl.dropboxusercontent.com:443.\n",
      "HTTP request sent, awaiting response... 200 OK\n",
      "Length: 4544401394 (4.2G) [application/octet-stream]\n",
      "Saving to: ‘out/imagenet_attb.pt’\n",
      "\n",
      "out/imagenet_attb.p 100%[===================>]   4.23G  28.1MB/s    in 1m 52s  \n",
      "\n",
      "2024-02-14 15:40:08 (38.8 MB/s) - ‘out/imagenet_attb.pt’ saved [4544401394/4544401394]\n",
      "\n"
     ]
    }
   ],
   "source": [
    "# note: ~20G\n",
    "print ('Downloading dataloader (FFCV format)') \n",
    "!wget -nc 'https://www.dropbox.com/scl/fi/ovztah82lk6qav1bkhpn3/imagenet_val.beton?rlkey=eroq5kcusbwdfecxk0xscnu3n&dl=0' -O out/imagenet_val.beton\n",
    "\n",
    "print ('Downloading attributions')\n",
    "!wget 'https://www.dropbox.com/scl/fi/vd38pyjjyz1g7zu72r741/imagenet_resnet.pt?rlkey=45ideswhv9x3tyu7ctjcd84df&dl=0' -O out/imagenet_attb.pt"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "c39fc6bc-4a08-46ca-8100-bdc4f3af90d1",
   "metadata": {},
   "source": [
    "### Setup"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 21,
   "id": "c3b0715b-7386-40b0-971e-4a4f5bb2821f",
   "metadata": {
    "tags": []
   },
   "outputs": [],
   "source": [
    "DEVICE = torch.device(0) \n",
    "\n",
    "def get_model():\n",
    "    weights = torchvision.models.ResNet50_Weights.IMAGENET1K_V1\n",
    "    model = torchvision.models.resnet50(weights=weights).eval()\n",
    "    return model\n",
    "\n",
    "def get_attributions():\n",
    "    pth = 'out/imagenet_attb.pt'\n",
    "    return torch.load(pth)['weight'].T.float().nan_to_num()\n",
    "\n",
    "def get_loaders(batch_size=512, num_workers=2):\n",
    "    pipeline = ffcv_pipelines.get_pipelines('imagenet', 'val', DEVICE)\n",
    "    loaders = {'val': 'out/imagenet_val.beton'}\n",
    "    return {k: data_utils.get_ffcv_loader(v, batch_size, num_workers, pipeline, is_train=False) for k, v in loaders.items()}\n",
    "\n",
    "def get_model_components(model):\n",
    "    def is_component(comp_name):\n",
    "        valid = 'conv' in comp_name\n",
    "        return valid\n",
    "\n",
    "    def get_comp_dim(comp):\n",
    "        assert type(comp) is nn.Conv2d\n",
    "        return comp.out_channels\n",
    "\n",
    "    comp_dims = {}\n",
    "    for comp_name, comp in model.named_modules():\n",
    "        if not is_component(comp_name): continue\n",
    "        key = comp_name+'.weight'\n",
    "        comp_dims[key] = get_comp_dim(comp)\n",
    "\n",
    "    return comp_dims"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 22,
   "id": "5e0c7322-0145-4cd9-9172-1a53c7352cf0",
   "metadata": {
    "tags": []
   },
   "outputs": [
    {
     "data": {
      "text/plain": [
       "22720"
      ]
     },
     "execution_count": 22,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "# get model, data, attributions \n",
    "model = get_model()\n",
    "comps = get_model_components(model)\n",
    "num_components = sum(comps.values())\n",
    "\n",
    "loader_map = get_loaders() \n",
    "cnames = data_utils.SIMPLE_IMAGENET_CLASS_NAMES\n",
    "\n",
    "attb = get_attributions()\n",
    "editor = coar_edit.CoarEdit()\n",
    "\n",
    "num_components"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 23,
   "id": "6f2a4017-b1c0-4560-b2c7-6b415748e226",
   "metadata": {
    "tags": []
   },
   "outputs": [],
   "source": [
    "# get metadata\n",
    "model = model.to(DEVICE)\n",
    "preds, labels = eval_utils.get_predictions(model, loader_map['val'], DEVICE)\n",
    "margins = eval_utils.get_margins(model, loader_map['val'], DEVICE)"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "5651c48a-0c02-4fb5-bdf9-367050b415ae",
   "metadata": {},
   "source": [
    "### Apply Coar-Edit to induce a misclassification"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 24,
   "id": "be1f58a3-a034-45ab-8664-c69de2e93565",
   "metadata": {
    "tags": []
   },
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "#6671 (bittern)\n",
      "Margin: 7.086\n",
      "target examples torch.Size([1, 22720])\n",
      "reference examples torch.Size([50, 22720])\n"
     ]
    }
   ],
   "source": [
    "# prepare inputs \n",
    "valid_inds = (margins>0).nonzero()[0]\n",
    "\n",
    "mi = np.random.choice(valid_inds)\n",
    "print ('#{} ({})'.format(mi, cnames[labels[mi]], cnames[preds[mi]]))\n",
    "print ('Margin: {:.3f}'.format(margins[mi]))\n",
    "\n",
    "num_ref = 50\n",
    "ref_inds = np.random.choice(range(attb.shape[0]), size=num_ref, replace=False)\n",
    "ref_attb = attb[ref_inds]\n",
    "target_attb = attb[mi][None,:]\n",
    "\n",
    "print ('target examples', target_attb.shape)\n",
    "print ('reference examples', ref_attb.shape)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 25,
   "id": "dbe54377-9e4a-4fff-bc39-e0c8887424a8",
   "metadata": {
    "tags": []
   },
   "outputs": [
    {
     "data": {
      "text/plain": [
       "[('layer4.1.conv2.weight', 386),\n",
       " ('layer3.4.conv3.weight', 1),\n",
       " ('layer4.0.conv2.weight', 168),\n",
       " ('layer4.2.conv1.weight', 236),\n",
       " ('layer3.4.conv2.weight', 35),\n",
       " ('layer4.2.conv2.weight', 124),\n",
       " ('layer4.0.conv1.weight', 369),\n",
       " ('layer4.0.conv2.weight', 466),\n",
       " ('layer4.0.conv3.weight', 257),\n",
       " ('layer4.1.conv3.weight', 257)]"
      ]
     },
     "execution_count": 25,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "# use Coar-Edit to identify a subset of components to ablate \n",
    "subset_size = 200 # max number of components to ablate \n",
    "ablate_order = 'topk' # get components with most negative scores (to increase margin)\n",
    "\n",
    "scores = editor.score_components(target_attb, ref_attb)\n",
    "subset = editor.get_subset(scores, subset_size, ablate_order)\n",
    "subset_inds = subset.indices.numpy()\n",
    "ablate_set = [ndm_utils.get_component_from_index(ind, comps) for ind in subset_inds]\n",
    "ablate_set[:10]"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 26,
   "id": "2771ff9c-7585-4d5f-b6e6-fb346f34845c",
   "metadata": {
    "tags": []
   },
   "outputs": [
    {
     "name": "stderr",
     "output_type": "stream",
     "text": [
      "(avg:3.00, example:-0.45):  10%|████                                     | 20/200 [00:42<06:22,  2.13s/it]\n"
     ]
    },
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "Accuracy: 0.754 -> 0.754\n"
     ]
    },
    {
     "data": {
      "image/png": "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\n",
      "text/plain": [
       "<Figure size 746.75x500 with 1 Axes>"
      ]
     },
     "metadata": {},
     "output_type": "display_data"
    }
   ],
   "source": [
    "# iteratively ablate identified components and evaluate model performance \n",
    "eval_every = 5\n",
    "masked_model = get_model().to(DEVICE)\n",
    "rows = [] \n",
    "\n",
    "for k, (comp_name, comp_idx) in (pbar:=tqdm.tqdm(list(enumerate(ablate_set)))):\n",
    "    if k % eval_every == 0:\n",
    "        # eval + log\n",
    "        masked_margins = eval_utils.get_margins(masked_model, loader_map['val'], DEVICE)\n",
    "        margin_avg = np.mean(masked_margins) \n",
    "        example_margin = masked_margins[mi] \n",
    "\n",
    "        row = {\n",
    "            'Num. components ablated': k,\n",
    "            'Average margin over val set': margin_avg, \n",
    "            'Example margin': example_margin,\n",
    "            'Accuracy': (masked_margins>0).mean()\n",
    "        }\n",
    "\n",
    "        rows.append(row)\n",
    "        pbar.set_description('(avg:{:.2f}, example:{:.2f})'.format(margin_avg, example_margin))\n",
    "        \n",
    "        if example_margin < 0:\n",
    "            break\n",
    "    \n",
    "    # ablate\n",
    "    masked_model = editor.ablate_model_given_component(masked_model, comp_name, comp_idx)\n",
    "    masked_model = masked_model.to(DEVICE).eval()\n",
    "    \n",
    "# plot effect of edit on model performance \n",
    "df = pd.DataFrame(rows)\n",
    "pdf = df.melt(id_vars=['Num. components ablated'], \n",
    "              value_vars=['Average margin over val set', 'Example margin'],\n",
    "              var_name='Split', \n",
    "              value_name='Margin')\n",
    "\n",
    "sns.set_style('darkgrid')\n",
    "fg = sns.relplot(data=pdf, \n",
    "                 x='Num. components ablated', \n",
    "                 y='Margin', \n",
    "                 hue='Split',\n",
    "                 kind='line', \n",
    "                 lw=2, marker='o')\n",
    "\n",
    "fg.axes[0,0].axhline(0, color='red', ls='--', lw=2)\n",
    "fg.fig.suptitle\n",
    "\n",
    "acc_before_edit = df.iloc[0]['Accuracy']\n",
    "acc_after_edit = df[(df['Example margin']>0)].sort_values('Num. components ablated').iloc[0]['Accuracy']\n",
    "print ('Accuracy: {:.3f} -> {:.3f}'.format(acc_before_edit, acc_after_edit))"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "508141fa-b15a-4f69-8a1f-6432f9acb40d",
   "metadata": {
    "tags": []
   },
   "source": [
    "### Apply Coar-Edit to correct a misclassification"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 31,
   "id": "c4207e81-2daf-475d-b0ce-b69486c08beb",
   "metadata": {
    "tags": []
   },
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "#20347 (altar misclassified as throne)\n",
      "Margin: -2.391\n",
      "target examples torch.Size([1, 22720])\n",
      "reference examples torch.Size([50, 22720])\n"
     ]
    }
   ],
   "source": [
    "# prepare inputs \n",
    "valid_inds = (margins>0).nonzero()[0]\n",
    "\n",
    "misclassified_inds = (margins<0).nonzero()[0]\n",
    "mi = misclassified_ind = np.random.choice(misclassified_inds)\n",
    "print ('#{} ({} misclassified as {})'.format(mi, cnames[labels[mi]], cnames[preds[mi]]))\n",
    "print ('Margin: {:.3f}'.format(margins[mi]))\n",
    "\n",
    "num_ref = 50\n",
    "ref_inds = np.random.choice(range(attb.shape[0]), size=num_ref, replace=False)\n",
    "ref_attb = attb[ref_inds]\n",
    "target_attb = attb[mi][None,:]\n",
    "\n",
    "print ('target examples', target_attb.shape)\n",
    "print ('reference examples', ref_attb.shape)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 32,
   "id": "bf714041-586d-41b9-a5b7-a6aadd583bad",
   "metadata": {
    "tags": []
   },
   "outputs": [
    {
     "data": {
      "text/plain": [
       "[('layer4.1.conv2.weight', 504),\n",
       " ('layer4.1.conv3.weight', 246),\n",
       " ('layer4.0.conv3.weight', 246),\n",
       " ('layer4.2.conv3.weight', 246),\n",
       " ('layer4.1.conv3.weight', 893),\n",
       " ('layer4.0.conv2.weight', 81),\n",
       " ('layer4.0.conv2.weight', 406),\n",
       " ('layer4.1.conv3.weight', 1693),\n",
       " ('layer4.2.conv1.weight', 1),\n",
       " ('layer4.0.conv3.weight', 511)]"
      ]
     },
     "execution_count": 32,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "# use Coar-Edit to identify a subset of components to ablate \n",
    "subset_size = 200 # max number of components to ablate \n",
    "ablate_order = 'bottomk' # get components with most negative scores (to increase margin)\n",
    "\n",
    "scores = editor.score_components(target_attb, ref_attb)\n",
    "subset = editor.get_subset(scores, subset_size, ablate_order)\n",
    "subset_inds = subset.indices.numpy()\n",
    "ablate_set = [ndm_utils.get_component_from_index(ind, comps) for ind in subset_inds]\n",
    "ablate_set[:10]"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 33,
   "id": "c6fcba8e-6ca1-4b52-8ae1-95807f073f69",
   "metadata": {
    "tags": []
   },
   "outputs": [
    {
     "name": "stderr",
     "output_type": "stream",
     "text": [
      "(avg:3.03, example:0.31):   3%|█▎                                         | 6/200 [00:34<18:24,  5.69s/it]\n"
     ]
    },
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "Accuracy: 0.754 -> 0.753\n"
     ]
    },
    {
     "data": {
      "image/png": "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\n",
      "text/plain": [
       "<Figure size 746.75x500 with 1 Axes>"
      ]
     },
     "metadata": {},
     "output_type": "display_data"
    }
   ],
   "source": [
    "# iteratively ablate identified components and evaluate model performance \n",
    "eval_every = 2\n",
    "masked_model = get_model().to(DEVICE)\n",
    "rows = [] \n",
    "\n",
    "for k, (comp_name, comp_idx) in (pbar:=tqdm.tqdm(list(enumerate(ablate_set)))):\n",
    "    if k % eval_every == 0:\n",
    "        # eval + log\n",
    "        masked_margins = eval_utils.get_margins(masked_model, loader_map['val'], DEVICE)\n",
    "        margin_avg = np.mean(masked_margins) \n",
    "        example_margin = masked_margins[mi] \n",
    "\n",
    "        row = {\n",
    "            'Num. components ablated': k,\n",
    "            'Average margin over val set': margin_avg, \n",
    "            'Example margin': example_margin,\n",
    "            'Accuracy': (masked_margins>0).mean()\n",
    "        }\n",
    "\n",
    "        rows.append(row)\n",
    "        pbar.set_description('(avg:{:.2f}, example:{:.2f})'.format(margin_avg, example_margin))\n",
    "        \n",
    "        if example_margin > 0:\n",
    "            break\n",
    "    \n",
    "    # ablate\n",
    "    masked_model = editor.ablate_model_given_component(masked_model, comp_name, comp_idx)\n",
    "    masked_model = masked_model.to(DEVICE).eval()\n",
    "    \n",
    "# plot effect of edit on model performance \n",
    "df = pd.DataFrame(rows)\n",
    "pdf = df.melt(id_vars=['Num. components ablated'], \n",
    "              value_vars=['Average margin over val set', 'Example margin'],\n",
    "              var_name='Split', \n",
    "              value_name='Margin')\n",
    "\n",
    "sns.set_style('darkgrid')\n",
    "fg = sns.relplot(data=pdf, \n",
    "                 x='Num. components ablated', \n",
    "                 y='Margin', \n",
    "                 hue='Split',\n",
    "                 kind='line', \n",
    "                 lw=2, marker='o')\n",
    "\n",
    "fg.axes[0,0].axhline(0, color='red', ls='--', lw=2)\n",
    "fg.fig.suptitle\n",
    "\n",
    "acc_before_edit = df.iloc[0]['Accuracy']\n",
    "\n",
    "if len(df[(df['Example margin']>0)]):\n",
    "    acc_after_edit = df[(df['Example margin']>0)].sort_values('Num. components ablated').iloc[0]['Accuracy']\n",
    "else: \n",
    "    acc_after_edit = df.sort_values('Example margin').iloc[0]['Accuracy']\n",
    "    \n",
    "print ('Accuracy: {:.3f} -> {:.3f}'.format(acc_before_edit, acc_after_edit))"
   ]
  }
 ],
 "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": "3.10.9"
  }
 },
 "nbformat": 4,
 "nbformat_minor": 5
}