siim_isic_integrated_model.ipynb

{
 "cells": [
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "%reload_ext autoreload\n",
    "%autoreload 2\n",
    "%matplotlib inline"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "from fastai.vision import *\n",
    "from fastai.tabular import *\n",
    "from image_tabular.core import *\n",
    "from image_tabular.dataset import *\n",
    "from image_tabular.model import *\n",
    "from image_tabular.metric import *\n",
    "\n",
    "# use gpu by default if available\n",
    "device = torch.device('cuda:0' if torch.cuda.is_available() else 'cpu')"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "import warnings\n",
    "warnings.filterwarnings(\"ignore\", category=UserWarning, module=\"torch.nn.functional\")"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "data_path = Path(\"./data/siim-isic-melanoma-classification/\")"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "33126 10982\n"
     ]
    }
   ],
   "source": [
    "train_df = pd.read_csv(data_path/\"train.csv\")\n",
    "test_df = pd.read_csv(data_path/\"test.csv\")\n",
    "\n",
    "print(len(train_df), len(test_df))"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "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>image_name</th>\n",
       "      <th>patient_id</th>\n",
       "      <th>sex</th>\n",
       "      <th>age_approx</th>\n",
       "      <th>anatom_site_general_challenge</th>\n",
       "      <th>diagnosis</th>\n",
       "      <th>benign_malignant</th>\n",
       "      <th>target</th>\n",
       "    </tr>\n",
       "  </thead>\n",
       "  <tbody>\n",
       "    <tr>\n",
       "      <th>0</th>\n",
       "      <td>ISIC_2637011</td>\n",
       "      <td>IP_7279968</td>\n",
       "      <td>male</td>\n",
       "      <td>45.0</td>\n",
       "      <td>head/neck</td>\n",
       "      <td>unknown</td>\n",
       "      <td>benign</td>\n",
       "      <td>0</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>1</th>\n",
       "      <td>ISIC_0015719</td>\n",
       "      <td>IP_3075186</td>\n",
       "      <td>female</td>\n",
       "      <td>45.0</td>\n",
       "      <td>upper extremity</td>\n",
       "      <td>unknown</td>\n",
       "      <td>benign</td>\n",
       "      <td>0</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>2</th>\n",
       "      <td>ISIC_0052212</td>\n",
       "      <td>IP_2842074</td>\n",
       "      <td>female</td>\n",
       "      <td>50.0</td>\n",
       "      <td>lower extremity</td>\n",
       "      <td>nevus</td>\n",
       "      <td>benign</td>\n",
       "      <td>0</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>3</th>\n",
       "      <td>ISIC_0068279</td>\n",
       "      <td>IP_6890425</td>\n",
       "      <td>female</td>\n",
       "      <td>45.0</td>\n",
       "      <td>head/neck</td>\n",
       "      <td>unknown</td>\n",
       "      <td>benign</td>\n",
       "      <td>0</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>4</th>\n",
       "      <td>ISIC_0074268</td>\n",
       "      <td>IP_8723313</td>\n",
       "      <td>female</td>\n",
       "      <td>55.0</td>\n",
       "      <td>upper extremity</td>\n",
       "      <td>unknown</td>\n",
       "      <td>benign</td>\n",
       "      <td>0</td>\n",
       "    </tr>\n",
       "  </tbody>\n",
       "</table>\n",
       "</div>"
      ],
      "text/plain": [
       "     image_name  patient_id     sex  age_approx anatom_site_general_challenge  \\\n",
       "0  ISIC_2637011  IP_7279968    male        45.0                     head/neck   \n",
       "1  ISIC_0015719  IP_3075186  female        45.0               upper extremity   \n",
       "2  ISIC_0052212  IP_2842074  female        50.0               lower extremity   \n",
       "3  ISIC_0068279  IP_6890425  female        45.0                     head/neck   \n",
       "4  ISIC_0074268  IP_8723313  female        55.0               upper extremity   \n",
       "\n",
       "  diagnosis benign_malignant  target  \n",
       "0   unknown           benign       0  \n",
       "1   unknown           benign       0  \n",
       "2     nevus           benign       0  \n",
       "3   unknown           benign       0  \n",
       "4   unknown           benign       0  "
      ]
     },
     "execution_count": null,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "train_df.head()"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [
    {
     "data": {
      "text/plain": [
       "0    0.98237\n",
       "1    0.01763\n",
       "Name: target, dtype: float64"
      ]
     },
     "execution_count": null,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "# extremely unbalanced dataset, most of the images are benign\n",
    "train_df[\"target\"].value_counts(normalize=True)"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "## Image data"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "tfms = get_transforms(flip_vert=True)\n",
    "size = 128"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [
    {
     "data": {
      "text/plain": [
       "6625"
      ]
     },
     "execution_count": null,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "# idx for validation, shared by image and tabular data\n",
    "val_idx = get_valid_index(train_df)\n",
    "len(val_idx)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [
    {
     "data": {
      "text/plain": [
       "LabelLists;\n",
       "\n",
       "Train: LabelList (26501 items)\n",
       "x: ImageList\n",
       "Image (3, 128, 128),Image (3, 128, 128),Image (3, 128, 128),Image (3, 128, 128),Image (3, 128, 128)\n",
       "y: CategoryList\n",
       "0,0,0,0,0\n",
       "Path: data/siim-isic-melanoma-classification;\n",
       "\n",
       "Valid: LabelList (6625 items)\n",
       "x: ImageList\n",
       "Image (3, 128, 128),Image (3, 128, 128),Image (3, 128, 128),Image (3, 128, 128),Image (3, 128, 128)\n",
       "y: CategoryList\n",
       "0,0,0,0,0\n",
       "Path: data/siim-isic-melanoma-classification;\n",
       "\n",
       "Test: LabelList (10982 items)\n",
       "x: ImageList\n",
       "Image (3, 128, 128),Image (3, 128, 128),Image (3, 128, 128),Image (3, 128, 128),Image (3, 128, 128)\n",
       "y: EmptyLabelList\n",
       ",,,,\n",
       "Path: data/siim-isic-melanoma-classification"
      ]
     },
     "execution_count": null,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "# load image data using train_df and prepare fastai LabelLists\n",
    "image_data = (ImageList.from_df(train_df, path=data_path, cols=\"image_name\",\n",
    "                               folder=\"train_128\", suffix=\".jpg\")\n",
    "              .split_by_idx(val_idx)\n",
    "              .label_from_df(cols=\"target\")\n",
    "              .transform(tfms, size=size))\n",
    "\n",
    "# add test data so that we can make predictions\n",
    "test_image_data = ImageList.from_df(test_df, path=data_path, cols=\"image_name\",\n",
    "                                    folder=\"test_128\", suffix=\".jpg\")\n",
    "image_data.add_test(test_image_data)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "0\n"
     ]
    },
    {
     "data": {
      "image/jpeg": "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\n",
      "image/png": "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\n",
      "text/plain": [
       "Image (3, 128, 128)"
      ]
     },
     "execution_count": null,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "# show one example image\n",
    "print(image_data.train[0][1])\n",
    "image_data.train[0][0]"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "## Tabular data"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "dep_var = 'target'\n",
    "cat_names = ['sex', 'anatom_site_general_challenge']\n",
    "cont_names = ['age_approx']\n",
    "procs = [FillMissing, Categorify, Normalize]"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [
    {
     "data": {
      "text/plain": [
       "LabelLists;\n",
       "\n",
       "Train: LabelList (26501 items)\n",
       "x: TabularList\n",
       "sex male; anatom_site_general_challenge head/neck; age_approx_na False; age_approx -0.2713; ,sex female; anatom_site_general_challenge upper extremity; age_approx_na False; age_approx -0.2713; ,sex female; anatom_site_general_challenge lower extremity; age_approx_na False; age_approx 0.0768; ,sex female; anatom_site_general_challenge head/neck; age_approx_na False; age_approx -0.2713; ,sex female; anatom_site_general_challenge lower extremity; age_approx_na False; age_approx -0.6195; \n",
       "y: CategoryList\n",
       "0,0,0,0,0\n",
       "Path: data/siim-isic-melanoma-classification;\n",
       "\n",
       "Valid: LabelList (6625 items)\n",
       "x: TabularList\n",
       "sex female; anatom_site_general_challenge upper extremity; age_approx_na False; age_approx -0.2713; ,sex female; anatom_site_general_challenge upper extremity; age_approx_na False; age_approx -0.2713; ,sex male; anatom_site_general_challenge upper extremity; age_approx_na False; age_approx -0.2713; ,sex male; anatom_site_general_challenge upper extremity; age_approx_na False; age_approx 1.1214; ,sex male; anatom_site_general_challenge torso; age_approx_na False; age_approx -1.3159; \n",
       "y: CategoryList\n",
       "0,0,0,0,0\n",
       "Path: data/siim-isic-melanoma-classification;\n",
       "\n",
       "Test: LabelList (10982 items)\n",
       "x: TabularList\n",
       "sex male; anatom_site_general_challenge #na#; age_approx_na False; age_approx 1.4696; ,sex male; anatom_site_general_challenge lower extremity; age_approx_na False; age_approx -0.6195; ,sex female; anatom_site_general_challenge torso; age_approx_na False; age_approx 0.4250; ,sex female; anatom_site_general_challenge torso; age_approx_na False; age_approx 0.0768; ,sex female; anatom_site_general_challenge lower extremity; age_approx_na False; age_approx -0.2713; \n",
       "y: EmptyLabelList\n",
       ",,,,\n",
       "Path: data/siim-isic-melanoma-classification"
      ]
     },
     "execution_count": null,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "tab_data = (TabularList.from_df(train_df, path=data_path, cat_names=cat_names, cont_names=cont_names, procs=procs)\n",
    "                           .split_by_idx(val_idx)\n",
    "                           .label_from_df(cols=dep_var))\n",
    "\n",
    "# add test\n",
    "tab_data.add_test(TabularList.from_df(test_df, cat_names=cat_names, cont_names=cont_names,\n",
    "                                      processor = tab_data.train.x.processor))"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [
    {
     "data": {
      "text/plain": [
       "(TabularLine [tensor([2, 1, 1]), tensor([-0.2713])], Category 0)"
      ]
     },
     "execution_count": null,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "# one example\n",
    "tab_data.train[0]"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "## Integrate image and tabular data"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "integrate_train, integrate_valid, integrate_test = get_imagetabdatasets(image_data, tab_data)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [
    {
     "data": {
      "text/plain": [
       "DataBunch;\n",
       "\n",
       "Train: <image_tabular.dataset.ImageTabDataset object at 0x7f68646769e8>;\n",
       "\n",
       "Valid: <image_tabular.dataset.ImageTabDataset object at 0x7f6864676b00>;\n",
       "\n",
       "Test: <image_tabular.dataset.ImageTabDataset object at 0x7f6864676b38>"
      ]
     },
     "execution_count": null,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "# package train, valid, and test datasets into a fastai databunch\n",
    "bs = 64\n",
    "\n",
    "db = DataBunch.create(integrate_train, integrate_valid, integrate_test,\n",
    "                      path=data_path, bs=bs)\n",
    "db"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "# image normalization with imagenet_stats\n",
    "db.norm, db.denorm = normalize_funcs_image_tab(*imagenet_stats)\n",
    "db.add_tfm(db.norm)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [
    {
     "data": {
      "text/plain": [
       "2"
      ]
     },
     "execution_count": null,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "# check the shape of one batch\n",
    "x, y = next(iter(db.train_dl))\n",
    "len(x)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [
    {
     "data": {
      "text/plain": [
       "torch.Size([64, 3, 128, 128])"
      ]
     },
     "execution_count": null,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "# images\n",
    "x[0].shape"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [
    {
     "data": {
      "text/plain": [
       "(torch.Size([64, 3]), torch.Size([64, 1]))"
      ]
     },
     "execution_count": null,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "# categorical and continuous tabular data \n",
    "x[1][0].shape, x[1][1].shape"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [
    {
     "data": {
      "text/plain": [
       "torch.Size([64])"
      ]
     },
     "execution_count": null,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "# targets\n",
    "y.shape"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "## Model that trains on image and tabular data simultaneously"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "# cnn model for images, use Resnet50 as an example\n",
    "cnn_arch = models.resnet50\n",
    "\n",
    "# cnn_out_sz is the output size of the cnn model that will be concatenated with tabular model output\n",
    "cnn_out_sz = 256\n",
    "\n",
    "# use fastai functions to get a cnn model\n",
    "image_data_db = image_data.databunch()\n",
    "image_data_db.c = cnn_out_sz\n",
    "cnn_learn = cnn_learner(image_data_db, cnn_arch, ps=0.2)\n",
    "cnn_model = cnn_learn.model"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [
    {
     "data": {
      "text/plain": [
       "TabularModel(\n",
       "  (embeds): ModuleList(\n",
       "    (0): Embedding(3, 3)\n",
       "    (1): Embedding(7, 5)\n",
       "    (2): Embedding(3, 3)\n",
       "  )\n",
       "  (emb_drop): Dropout(p=0.0, inplace=False)\n",
       "  (bn_cont): BatchNorm1d(1, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)\n",
       "  (layers): Sequential(\n",
       "    (0): Linear(in_features=12, out_features=8, bias=True)\n",
       "    (1): ReLU(inplace=True)\n",
       "    (2): BatchNorm1d(8, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)\n",
       "    (3): Dropout(p=0.2, inplace=False)\n",
       "    (4): Linear(in_features=8, out_features=8, bias=True)\n",
       "  )\n",
       ")"
      ]
     },
     "execution_count": null,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "# get embedding sizes of categorical data\n",
    "emb_szs = tab_data.train.get_emb_szs()\n",
    "\n",
    "# output size of the tabular model that will be concatenated with cnn model output\n",
    "tab_out_sz = 8\n",
    "\n",
    "# use fastai functions to get a tabular model\n",
    "tabular_model = TabularModel(emb_szs, len(cont_names), out_sz=tab_out_sz, layers=[8], ps=0.2)\n",
    "tabular_model"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "# get an integrated model that combines the two components and concatenate their outputs\n",
    "# which will pass through additional fully connected layers\n",
    "integrate_model = CNNTabularModel(cnn_model,\n",
    "                                  tabular_model,\n",
    "                                  layers = [cnn_out_sz + tab_out_sz, 32],\n",
    "                                  ps=0.2,\n",
    "                                  out_sz=2).to(device)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [
    {
     "data": {
      "text/plain": [
       "torch.Size([64, 2])"
      ]
     },
     "execution_count": null,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "# check model output dimension, should be (bs, 2)\n",
    "integrate_model(*x).shape"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "# adjust loss function weight because the dataset is extremely unbalanced\n",
    "weights = [1/(1-train_df[\"target\"].mean()), 1/train_df[\"target\"].mean()]\n",
    "loss_func = CrossEntropyFlat(weight=torch.FloatTensor(weights).to(device))"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "# package everything in a fastai learner, add auc roc score as a metric\n",
    "learn = Learner(db, integrate_model, metrics=[accuracy, ROCAUC()], loss_func=loss_func)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "# organize layer groups in order to use differential learning rates provided by fastai\n",
    "# the first two layer groups are earlier layers of resnet\n",
    "# the last layer group consists of the fully connected layers of cnn model, tabular model,\n",
    "# and final fully connected layers for the concatenated data\n",
    "learn.layer_groups = [nn.Sequential(*flatten_model(cnn_learn.layer_groups[0])),\n",
    "                      nn.Sequential(*flatten_model(cnn_learn.layer_groups[1])),\n",
    "                      nn.Sequential(*(flatten_model(cnn_learn.layer_groups[2]) +\n",
    "                                      flatten_model(integrate_model.tabular_model) +\n",
    "                                      flatten_model(integrate_model.layers)))]"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "## Training"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [
    {
     "data": {
      "text/html": [
       "\n",
       "    <div>\n",
       "        <style>\n",
       "            /* Turns off some styling */\n",
       "            progress {\n",
       "                /* gets rid of default border in Firefox and Opera. */\n",
       "                border: none;\n",
       "                /* Needs to be in here for Safari polyfill so background images work as expected. */\n",
       "                background-size: auto;\n",
       "            }\n",
       "            .progress-bar-interrupted, .progress-bar-interrupted::-webkit-progress-bar {\n",
       "                background: #F44336;\n",
       "            }\n",
       "        </style>\n",
       "      <progress value='0' class='' max='1', style='width:300px; height:20px; vertical-align: middle;'></progress>\n",
       "      0.00% [0/1 00:00<00:00]\n",
       "    </div>\n",
       "    \n",
       "<table border=\"1\" class=\"dataframe\">\n",
       "  <thead>\n",
       "    <tr style=\"text-align: left;\">\n",
       "      <th>epoch</th>\n",
       "      <th>train_loss</th>\n",
       "      <th>valid_loss</th>\n",
       "      <th>accuracy</th>\n",
       "      <th>rocauc</th>\n",
       "      <th>time</th>\n",
       "    </tr>\n",
       "  </thead>\n",
       "  <tbody>\n",
       "  </tbody>\n",
       "</table><p>\n",
       "\n",
       "    <div>\n",
       "        <style>\n",
       "            /* Turns off some styling */\n",
       "            progress {\n",
       "                /* gets rid of default border in Firefox and Opera. */\n",
       "                border: none;\n",
       "                /* Needs to be in here for Safari polyfill so background images work as expected. */\n",
       "                background-size: auto;\n",
       "            }\n",
       "            .progress-bar-interrupted, .progress-bar-interrupted::-webkit-progress-bar {\n",
       "                background: #F44336;\n",
       "            }\n",
       "        </style>\n",
       "      <progress value='93' class='' max='414', style='width:300px; height:20px; vertical-align: middle;'></progress>\n",
       "      22.46% [93/414 00:08<00:28 2.3661]\n",
       "    </div>\n",
       "    "
      ],
      "text/plain": [
       "<IPython.core.display.HTML object>"
      ]
     },
     "metadata": {},
     "output_type": "display_data"
    },
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "LR Finder is complete, type {learner_name}.recorder.plot() to see the graph.\n"
     ]
    },
    {
     "data": {
      "image/png": "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\n",
      "text/plain": [
       "<Figure size 432x288 with 1 Axes>"
      ]
     },
     "metadata": {
      "needs_background": "light"
     },
     "output_type": "display_data"
    }
   ],
   "source": [
    "# find learning rate to train the last layer group first \n",
    "learn.freeze()\n",
    "learn.lr_find()\n",
    "learn.recorder.plot()"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [
    {
     "data": {
      "text/html": [
       "<table border=\"1\" class=\"dataframe\">\n",
       "  <thead>\n",
       "    <tr style=\"text-align: left;\">\n",
       "      <th>epoch</th>\n",
       "      <th>train_loss</th>\n",
       "      <th>valid_loss</th>\n",
       "      <th>accuracy</th>\n",
       "      <th>rocauc</th>\n",
       "      <th>time</th>\n",
       "    </tr>\n",
       "  </thead>\n",
       "  <tbody>\n",
       "    <tr>\n",
       "      <td>0</td>\n",
       "      <td>0.705341</td>\n",
       "      <td>0.606712</td>\n",
       "      <td>0.590943</td>\n",
       "      <td>0.789852</td>\n",
       "      <td>00:34</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <td>1</td>\n",
       "      <td>0.624063</td>\n",
       "      <td>0.564126</td>\n",
       "      <td>0.663849</td>\n",
       "      <td>0.838528</td>\n",
       "      <td>00:35</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <td>2</td>\n",
       "      <td>0.566842</td>\n",
       "      <td>0.559899</td>\n",
       "      <td>0.656151</td>\n",
       "      <td>0.844186</td>\n",
       "      <td>00:35</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <td>3</td>\n",
       "      <td>0.564589</td>\n",
       "      <td>0.549224</td>\n",
       "      <td>0.685132</td>\n",
       "      <td>0.853641</td>\n",
       "      <td>00:35</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <td>4</td>\n",
       "      <td>0.554486</td>\n",
       "      <td>0.496429</td>\n",
       "      <td>0.754717</td>\n",
       "      <td>0.871573</td>\n",
       "      <td>00:35</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <td>5</td>\n",
       "      <td>0.513213</td>\n",
       "      <td>0.498709</td>\n",
       "      <td>0.757887</td>\n",
       "      <td>0.870145</td>\n",
       "      <td>00:35</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <td>6</td>\n",
       "      <td>0.523281</td>\n",
       "      <td>0.499963</td>\n",
       "      <td>0.747774</td>\n",
       "      <td>0.870568</td>\n",
       "      <td>00:35</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <td>7</td>\n",
       "      <td>0.506248</td>\n",
       "      <td>0.488594</td>\n",
       "      <td>0.768604</td>\n",
       "      <td>0.875076</td>\n",
       "      <td>00:35</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <td>8</td>\n",
       "      <td>0.478553</td>\n",
       "      <td>0.478242</td>\n",
       "      <td>0.781434</td>\n",
       "      <td>0.877582</td>\n",
       "      <td>00:35</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <td>9</td>\n",
       "      <td>0.484625</td>\n",
       "      <td>0.480606</td>\n",
       "      <td>0.768302</td>\n",
       "      <td>0.877502</td>\n",
       "      <td>00:35</td>\n",
       "    </tr>\n",
       "  </tbody>\n",
       "</table>"
      ],
      "text/plain": [
       "<IPython.core.display.HTML object>"
      ]
     },
     "metadata": {},
     "output_type": "display_data"
    }
   ],
   "source": [
    "# train\n",
    "learn.fit_one_cycle(10, 1e-4)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [
    {
     "data": {
      "text/html": [
       "<table border=\"1\" class=\"dataframe\">\n",
       "  <thead>\n",
       "    <tr style=\"text-align: left;\">\n",
       "      <th>epoch</th>\n",
       "      <th>train_loss</th>\n",
       "      <th>valid_loss</th>\n",
       "      <th>accuracy</th>\n",
       "      <th>rocauc</th>\n",
       "      <th>time</th>\n",
       "    </tr>\n",
       "  </thead>\n",
       "  <tbody>\n",
       "    <tr>\n",
       "      <td>0</td>\n",
       "      <td>0.472651</td>\n",
       "      <td>0.488090</td>\n",
       "      <td>0.792906</td>\n",
       "      <td>0.857081</td>\n",
       "      <td>00:49</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <td>1</td>\n",
       "      <td>0.491386</td>\n",
       "      <td>0.446133</td>\n",
       "      <td>0.838943</td>\n",
       "      <td>0.882625</td>\n",
       "      <td>00:48</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <td>2</td>\n",
       "      <td>0.458774</td>\n",
       "      <td>0.463548</td>\n",
       "      <td>0.821585</td>\n",
       "      <td>0.866160</td>\n",
       "      <td>00:48</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <td>3</td>\n",
       "      <td>0.411819</td>\n",
       "      <td>0.449613</td>\n",
       "      <td>0.821283</td>\n",
       "      <td>0.877032</td>\n",
       "      <td>00:48</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <td>4</td>\n",
       "      <td>0.450356</td>\n",
       "      <td>0.448963</td>\n",
       "      <td>0.836528</td>\n",
       "      <td>0.874391</td>\n",
       "      <td>00:48</td>\n",
       "    </tr>\n",
       "  </tbody>\n",
       "</table>"
      ],
      "text/plain": [
       "<IPython.core.display.HTML object>"
      ]
     },
     "metadata": {},
     "output_type": "display_data"
    }
   ],
   "source": [
    "# unfreeze all layer groups to train the entire model using differential learning rates\n",
    "learn.unfreeze()\n",
    "learn.fit_one_cycle(5, slice(1e-6, 1e-4))"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "The model achieved an ROC AUC score of about 0.87 on the validation set."
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "## Prediction"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [
    {
     "data": {
      "text/html": [],
      "text/plain": [
       "<IPython.core.display.HTML object>"
      ]
     },
     "metadata": {},
     "output_type": "display_data"
    }
   ],
   "source": [
    "# make predictions for the test set\n",
    "preds, y = learn.get_preds(DatasetType.Test)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "# submit predictions to kaggle\n",
    "submit = pd.read_csv(data_path/\"sample_submission.csv\")\n",
    "submit[\"target\"] = preds[:, 1]\n",
    "submit.to_csv(data_path/\"image_tab.csv\", index=False)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": []
  }
 ],
 "metadata": {
  "kernelspec": {
   "display_name": "Python 3",
   "language": "python",
   "name": "python3"
  }
 },
 "nbformat": 4,
 "nbformat_minor": 2
}