Ensemble.ipynb

{
 "cells": [
  {
   "cell_type": "code",
   "execution_count": 1,
   "id": "072fb6f1-2e13-47c1-82c2-2be865392890",
   "metadata": {},
   "outputs": [],
   "source": [
    "import os\n",
    "import pandas as pd\n",
    "import numpy as np\n",
    "from PIL import Image\n",
    "import cv2\n",
    "import albumentations as A\n",
    "from albumentations.pytorch.transforms import ToTensorV2\n",
    "import ttach as tta\n",
    "\n",
    "import torch\n",
    "import torch.nn as nn\n",
    "import torch.nn.functional as F\n",
    "from torch.utils.data import Dataset, DataLoader, WeightedRandomSampler\n",
    "\n",
    "import torchvision\n",
    "from torchvision.models import efficientnet_b0\n",
    "from torchvision import transforms\n",
    "from torchvision.transforms import Resize, ToTensor, Normalize\n",
    "\n",
    "import sklearn\n",
    "from sklearn.model_selection import train_test_split\n",
    "from sklearn.metrics import confusion_matrix\n",
    "from torchsampler import ImbalancedDatasetSampler\n",
    "\n",
    "import matplotlib as mpl\n",
    "import matplotlib.pyplot as plt\n",
    "import seaborn as sns\n",
    "import itertools\n",
    "\n",
    "import time\n",
    "from tqdm import tqdm, trange\n",
    "\n",
    "import wandb\n",
    "\n",
    "import coatnet\n",
    "from efficientnet_pytorch import EfficientNet\n",
    "\n",
    "import inspect"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 3,
   "id": "bdb37e2b-c961-4e47-9c0d-56cf7a9c9f49",
   "metadata": {},
   "outputs": [],
   "source": [
    "device = torch.device('cuda:0' if torch.cuda.is_available() else 'cpu')\n",
    "\n",
    "# 훈련 데이터셋 경로\n",
    "train_dir = '/opt/ml/input/data/train'\n",
    "image_dir = '/opt/ml/input/data/train/images'"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 4,
   "id": "bd9e424f-6524-45cb-87b4-751babb2cd31",
   "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>Unnamed: 0</th>\n",
       "      <th>id</th>\n",
       "      <th>path</th>\n",
       "      <th>class</th>\n",
       "    </tr>\n",
       "  </thead>\n",
       "  <tbody>\n",
       "    <tr>\n",
       "      <th>16976</th>\n",
       "      <td>16976</td>\n",
       "      <td>6392</td>\n",
       "      <td>/opt/ml/input/data/train/images/006392_male_As...</td>\n",
       "      <td>12</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>17649</th>\n",
       "      <td>17649</td>\n",
       "      <td>6511</td>\n",
       "      <td>/opt/ml/input/data/train/images/006511_male_As...</td>\n",
       "      <td>0</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>12452</th>\n",
       "      <td>12452</td>\n",
       "      <td>4228</td>\n",
       "      <td>/opt/ml/input/data/train/images/004228_female_...</td>\n",
       "      <td>10</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>5621</th>\n",
       "      <td>5621</td>\n",
       "      <td>1540</td>\n",
       "      <td>/opt/ml/input/data/train/images/001540_male_As...</td>\n",
       "      <td>2</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>12769</th>\n",
       "      <td>12769</td>\n",
       "      <td>4288</td>\n",
       "      <td>/opt/ml/input/data/train/images/004288_female_...</td>\n",
       "      <td>17</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>11685</th>\n",
       "      <td>11685</td>\n",
       "      <td>3848</td>\n",
       "      <td>/opt/ml/input/data/train/images/003848_male_As...</td>\n",
       "      <td>1</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>13116</th>\n",
       "      <td>13116</td>\n",
       "      <td>4373</td>\n",
       "      <td>/opt/ml/input/data/train/images/004373_male_As...</td>\n",
       "      <td>2</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>57</th>\n",
       "      <td>57</td>\n",
       "      <td>10</td>\n",
       "      <td>/opt/ml/input/data/train/images/000010_female_...</td>\n",
       "      <td>16</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>1076</th>\n",
       "      <td>1076</td>\n",
       "      <td>529</td>\n",
       "      <td>/opt/ml/input/data/train/images/000529_female_...</td>\n",
       "      <td>4</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>7372</th>\n",
       "      <td>7372</td>\n",
       "      <td>1920</td>\n",
       "      <td>/opt/ml/input/data/train/images/001920_female_...</td>\n",
       "      <td>17</td>\n",
       "    </tr>\n",
       "  </tbody>\n",
       "</table>\n",
       "</div>"
      ],
      "text/plain": [
       "       Unnamed: 0    id                                               path  \\\n",
       "16976       16976  6392  /opt/ml/input/data/train/images/006392_male_As...   \n",
       "17649       17649  6511  /opt/ml/input/data/train/images/006511_male_As...   \n",
       "12452       12452  4228  /opt/ml/input/data/train/images/004228_female_...   \n",
       "5621         5621  1540  /opt/ml/input/data/train/images/001540_male_As...   \n",
       "12769       12769  4288  /opt/ml/input/data/train/images/004288_female_...   \n",
       "11685       11685  3848  /opt/ml/input/data/train/images/003848_male_As...   \n",
       "13116       13116  4373  /opt/ml/input/data/train/images/004373_male_As...   \n",
       "57             57    10  /opt/ml/input/data/train/images/000010_female_...   \n",
       "1076         1076   529  /opt/ml/input/data/train/images/000529_female_...   \n",
       "7372         7372  1920  /opt/ml/input/data/train/images/001920_female_...   \n",
       "\n",
       "       class  \n",
       "16976     12  \n",
       "17649      0  \n",
       "12452     10  \n",
       "5621       2  \n",
       "12769     17  \n",
       "11685      1  \n",
       "13116      2  \n",
       "57        16  \n",
       "1076       4  \n",
       "7372      17  "
      ]
     },
     "execution_count": 4,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "# meta 데이터와 이미지 경로를 불러옵니다.\n",
    "train_df = pd.read_csv(os.path.join(train_dir, 'train.csv'))\n",
    "train_class_df = pd.read_csv(os.path.join(train_dir, 'train_class_df.csv'))\n",
    "train_df.sample(10)\n",
    "train_class_df.sample(10)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 5,
   "id": "119e632a-7cd8-4c14-9a86-dbc8f74f36ab",
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "       id  gender   race  age                    path  age_range\n",
      "0  000001  female  Asian   45  000001_female_Asian_45          1\n",
      "1  000002  female  Asian   52  000002_female_Asian_52          1\n",
      "2  000004    male  Asian   54    000004_male_Asian_54          1\n",
      "3  000005  female  Asian   58  000005_female_Asian_58          1\n",
      "4  000006  female  Asian   59  000006_female_Asian_59          1\n"
     ]
    }
   ],
   "source": [
    "# age label 추가\n",
    "ages = []\n",
    "for row in train_df['age']:\n",
    "    if row < 30:\n",
    "        ages.append(0)\n",
    "    elif row < 60:\n",
    "        ages.append(1)\n",
    "    else:\n",
    "        ages.append(2)\n",
    "train_df['age_range'] = ages\n",
    "\n",
    "print(train_df.head(5))"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 6,
   "id": "e4dc4383-008d-47cd-b60a-bb2e90244487",
   "metadata": {},
   "outputs": [],
   "source": [
    "# Dataset 생성\n",
    "class MaskDataset(Dataset):\n",
    "    def __init__(self, df, transform=None):\n",
    "        self.path = df['path']\n",
    "        self.label = df['class']\n",
    "        self.id = df['id']\n",
    "        self.transform = transform\n",
    "    \n",
    "    def __len__(self):\n",
    "        return len(self.path)\n",
    "\n",
    "    def __getitem__(self, idx):\n",
    "        image = Image.open(self.path.iloc[idx])\n",
    "        label = np.array(self.label.iloc[idx])\n",
    "\n",
    "        if self.transform:\n",
    "            image = self.transform(image)\n",
    "            #image = transforms.Resize((256, 256))(image)\n",
    "            #image = cv2.imread(self.path.iloc[idx])\n",
    "            #image = cv2.cvtColor(image, cv2.COLOR_BGR2RGB)\n",
    "            #augmented = self.transform(image=image)\n",
    "            #image = augmented['image']\n",
    "        return image, torch.tensor(label)\n",
    "    \n",
    "    def get_labels(self):\n",
    "        return self.label"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 7,
   "id": "4d6cdddb-1cde-4dd4-9a1e-2a5bf5376477",
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "       Unnamed: 0    id                                               path  \\\n",
      "17339       17339  6451  /opt/ml/input/data/train/images/006451_female_...   \n",
      "1894         1894   678  /opt/ml/input/data/train/images/000678_female_...   \n",
      "10459       10459  3567  /opt/ml/input/data/train/images/003567_female_...   \n",
      "14302       14302  5129  /opt/ml/input/data/train/images/005129_female_...   \n",
      "4598         4598  1340  /opt/ml/input/data/train/images/001340_male_As...   \n",
      "...           ...   ...                                                ...   \n",
      "937           937   378  /opt/ml/input/data/train/images/000378_female_...   \n",
      "5502         5502  1522  /opt/ml/input/data/train/images/001522_female_...   \n",
      "14513       14513  5234  /opt/ml/input/data/train/images/005234_male_As...   \n",
      "18621       18621  6737  /opt/ml/input/data/train/images/006737_female_...   \n",
      "5664         5664  1546  /opt/ml/input/data/train/images/001546_male_As...   \n",
      "\n",
      "       class  \n",
      "17339      3  \n",
      "1894       4  \n",
      "10459     15  \n",
      "14302     16  \n",
      "4598       6  \n",
      "...      ...  \n",
      "937       10  \n",
      "5502       3  \n",
      "14513      0  \n",
      "18621     15  \n",
      "5664      12  \n",
      "\n",
      "[15120 rows x 4 columns]\n"
     ]
    }
   ],
   "source": [
    "# train, valid 나누기 -> 이 코드에서는 valid만 필요\n",
    "train, valid = train_test_split(train_class_df,\n",
    "                                test_size = 0.2,\n",
    "                                shuffle = True,\n",
    "                                stratify = train_class_df['class'],\n",
    "                                random_state = 11)\n",
    "train.shape, valid.shape\n",
    "print(train)"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "e34a9158-5a05-4dda-ad05-97ec6346450a",
   "metadata": {},
   "source": [
    "## Augmentation"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 8,
   "id": "d2011b0a-d3e9-47ff-ab28-e438fad3a4ce",
   "metadata": {},
   "outputs": [],
   "source": [
    "image_paths = pd.read_csv(os.path.join(train_dir, \"train_class_df.csv\"))['path']\n",
    "\n",
    "transformation1 = transforms.Compose([\n",
    "    transforms.RandomRotation(15),\n",
    "    transforms.Resize((224, 224)),\n",
    "    transforms.ToTensor(),\n",
    "    transforms.Normalize([0.485, 0.456, 0.406], [0.229, 0.224, 0.225])\n",
    "])\n",
    "\n",
    "transformation2 = A.Compose([      \n",
    "    # 수평 뒤집기\n",
    "    A.HorizontalFlip(p=0.5),\n",
    "    A.Rotate(15),\n",
    "    # Contrast Limited Adaptive Histogram Equalization 적용\n",
    "    A.CLAHE(p=0.5),\n",
    "    # 무작위로 channel을 섞기\n",
    "    A.ChannelShuffle(p= 0.05),\n",
    "    A.ChannelDropout(),\n",
    "    # normalize\n",
    "    A.GaussNoise(),\n",
    "    A.CenterCrop(224, 224),\n",
    "    ToTensorV2(),\n",
    "])\n",
    "\n",
    "transformation3 = transforms.Compose([\n",
    "    transforms.RandomRotation(30),\n",
    "    transforms.Resize((224, 224)),\n",
    "    transforms.ToTensor(),\n",
    "    transforms.Normalize((0.5,0.5,0.5),(0.5,0.5,0.5))\n",
    "])\n",
    "\n",
    "transformation4 = transforms.Compose([\n",
    "    transforms.CenterCrop((300, 200)),\n",
    "    transforms.Resize((224, 224)),\n",
    "    transforms.ToTensor(),\n",
    "    transforms.Normalize([0.485, 0.456, 0.406], [0.229, 0.224, 0.225])\n",
    "])"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 9,
   "id": "5ccf7d25-d5ad-4025-a65f-35a5fcd8f1e3",
   "metadata": {},
   "outputs": [],
   "source": [
    "valid_dataset = MaskDataset(\n",
    "                valid,\n",
    "                transform = transformation4\n",
    ")"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 10,
   "id": "5e4040e3-947a-4d4a-ba6d-ee9b7edca39e",
   "metadata": {},
   "outputs": [],
   "source": [
    "batch_size = 32\n",
    "num_workers = 4\n",
    "\n",
    "valid_dataloader = DataLoader(\n",
    "    valid_dataset,\n",
    "    batch_size = batch_size,\n",
    "    num_workers = num_workers,\n",
    "    #sampler=ImbalancedDatasetSampler(train_dataset),\n",
    "    drop_last=True\n",
    ")"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "53c2d058-0167-4dab-b974-dd8c719c16ee",
   "metadata": {},
   "source": [
    "## Models"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 14,
   "id": "b158f4a5-61ab-4f6c-a3a3-9f38299cb836",
   "metadata": {},
   "outputs": [],
   "source": [
    "# EfficientNet Model from KKT\n",
    "class KKT_Effi(nn.Module):\n",
    "    def __init__(self, num_classes=18):\n",
    "        super().__init__()\n",
    "        self.models = torchvision.models.efficientnet_b0(pretrained=True)\n",
    "        self.models.classifier[1] = nn.Linear(1280, num_classes)\n",
    "\n",
    "\n",
    "\n",
    "    def forward(self, x):\n",
    "\n",
    "        return self.models(x)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 16,
   "id": "3e5edddd-a514-4a19-ad86-f335e2dc1160",
   "metadata": {},
   "outputs": [],
   "source": [
    "# ResNet model\n",
    "class ResNet(nn.Module):\n",
    "    def __init__(self, num_classes = 18, in_channels = 3):\n",
    "        super(ResNet, self).__init__()\n",
    "        self.network = torchvision.models.resnet18(pretrained=True)\n",
    "        num_ftrs = self.network.fc.in_features\n",
    "        self.network.fc = nn.Linear(num_ftrs, num_classes)\n",
    "\n",
    "    def forward(self, x):\n",
    "        x = self.network(x)\n",
    "        return x"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 17,
   "id": "929213b2-d370-4df7-98f4-8e21d5d471c0",
   "metadata": {},
   "outputs": [],
   "source": [
    "# CoAtNet model\n",
    "class MyModel(nn.Module):\n",
    "    def __init__(self, num_classes=18):\n",
    "        super().__init__()\n",
    "\n",
    "        self.model = coatnet.coatnet_0()\n",
    "        self.fc1 = nn.Linear(1000, 128)\n",
    "        self.dropout = nn.Dropout(p=0.5)\n",
    "        self.fc2 = nn.Linear(128, num_classes)\n",
    "        \n",
    "        \n",
    "        \n",
    "    def forward(self, x):\n",
    "        x = self.model(x)\n",
    "        x = F.relu(self.fc1(x))\n",
    "        x = self.dropout(x)\n",
    "        x = self.fc2(x)\n",
    "        return x"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "246cd77d-5f92-4da5-a0e8-59bb633aba99",
   "metadata": {},
   "source": [
    "## Inference"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "271114d4-e24f-4fe3-be7d-97c95237296c",
   "metadata": {},
   "source": [
    "### ensemble"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 45,
   "id": "5d8c530a-97c7-4ef4-8091-28927b1ed8e8",
   "metadata": {},
   "outputs": [
    {
     "name": "stderr",
     "output_type": "stream",
     "text": [
      "100%|██████████| 118/118 [00:12<00:00,  9.41it/s]\n"
     ]
    },
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "tensor([[[0.0236, 0.0628, 0.0092,  ..., 0.0071, 0.0178, 0.0101],\n",
      "         [0.1015, 0.1558, 0.1931,  ..., 0.0222, 0.0159, 0.0259],\n",
      "         [0.0071, 0.3402, 0.5796,  ..., 0.0057, 0.0020, 0.0013],\n",
      "         ...,\n",
      "         [0.0057, 0.0344, 0.0071,  ..., 0.0018, 0.0066, 0.0056],\n",
      "         [0.6819, 0.1052, 0.0151,  ..., 0.0062, 0.0047, 0.0067],\n",
      "         [0.0043, 0.0992, 0.8327,  ..., 0.0032, 0.0010, 0.0015]]]) tensor([4., 0., 1.,  ..., 4., 0., 2.])\n"
     ]
    },
    {
     "name": "stderr",
     "output_type": "stream",
     "text": [
      "100%|██████████| 118/118 [00:11<00:00, 10.19it/s]\n"
     ]
    },
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "tensor([[[2.3612e-02, 6.2821e-02, 9.1537e-03,  ..., 7.0571e-03,\n",
      "          1.7813e-02, 1.0057e-02],\n",
      "         [1.0150e-01, 1.5576e-01, 1.9305e-01,  ..., 2.2215e-02,\n",
      "          1.5863e-02, 2.5863e-02],\n",
      "         [7.1256e-03, 3.4020e-01, 5.7956e-01,  ..., 5.7358e-03,\n",
      "          1.9503e-03, 1.2744e-03],\n",
      "         ...,\n",
      "         [5.7352e-03, 3.4354e-02, 7.1103e-03,  ..., 1.8499e-03,\n",
      "          6.5644e-03, 5.5551e-03],\n",
      "         [6.8194e-01, 1.0519e-01, 1.5120e-02,  ..., 6.2375e-03,\n",
      "          4.7257e-03, 6.7118e-03],\n",
      "         [4.3128e-03, 9.9219e-02, 8.3268e-01,  ..., 3.1577e-03,\n",
      "          9.7451e-04, 1.4951e-03]],\n",
      "\n",
      "        [[1.0491e-03, 1.9064e-03, 1.0744e-02,  ..., 1.3052e-03,\n",
      "          2.5086e-03, 3.8112e-04],\n",
      "         [8.7946e-01, 1.7661e-02, 4.4420e-03,  ..., 4.6465e-03,\n",
      "          1.4305e-03, 3.9944e-04],\n",
      "         [1.4280e-04, 9.8222e-01, 6.9996e-05,  ..., 8.1679e-03,\n",
      "          3.6283e-03, 9.9072e-04],\n",
      "         ...,\n",
      "         [1.7967e-03, 2.4337e-02, 7.5004e-03,  ..., 9.4930e-04,\n",
      "          5.2375e-04, 4.0302e-04],\n",
      "         [9.3969e-01, 9.0301e-03, 6.1710e-03,  ..., 3.5786e-04,\n",
      "          5.9097e-04, 5.3556e-04],\n",
      "         [6.1931e-04, 8.2615e-03, 9.7595e-01,  ..., 2.3021e-03,\n",
      "          5.8235e-04, 4.0193e-04]]]) tensor([4., 0., 1.,  ..., 4., 0., 2.])\n",
      "Loaded pretrained weights for efficientnet-b0\n"
     ]
    },
    {
     "name": "stderr",
     "output_type": "stream",
     "text": [
      "100%|██████████| 118/118 [00:12<00:00,  9.45it/s]\n"
     ]
    },
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "tensor([[[2.3612e-02, 6.2821e-02, 9.1537e-03,  ..., 7.0571e-03,\n",
      "          1.7813e-02, 1.0057e-02],\n",
      "         [1.0150e-01, 1.5576e-01, 1.9305e-01,  ..., 2.2215e-02,\n",
      "          1.5863e-02, 2.5863e-02],\n",
      "         [7.1256e-03, 3.4020e-01, 5.7956e-01,  ..., 5.7358e-03,\n",
      "          1.9503e-03, 1.2744e-03],\n",
      "         ...,\n",
      "         [5.7352e-03, 3.4354e-02, 7.1103e-03,  ..., 1.8499e-03,\n",
      "          6.5644e-03, 5.5551e-03],\n",
      "         [6.8194e-01, 1.0519e-01, 1.5120e-02,  ..., 6.2375e-03,\n",
      "          4.7257e-03, 6.7118e-03],\n",
      "         [4.3128e-03, 9.9219e-02, 8.3268e-01,  ..., 3.1577e-03,\n",
      "          9.7451e-04, 1.4951e-03]],\n",
      "\n",
      "        [[1.0491e-03, 1.9064e-03, 1.0744e-02,  ..., 1.3052e-03,\n",
      "          2.5086e-03, 3.8112e-04],\n",
      "         [8.7946e-01, 1.7661e-02, 4.4420e-03,  ..., 4.6465e-03,\n",
      "          1.4305e-03, 3.9944e-04],\n",
      "         [1.4280e-04, 9.8222e-01, 6.9996e-05,  ..., 8.1679e-03,\n",
      "          3.6283e-03, 9.9072e-04],\n",
      "         ...,\n",
      "         [1.7967e-03, 2.4337e-02, 7.5004e-03,  ..., 9.4930e-04,\n",
      "          5.2375e-04, 4.0302e-04],\n",
      "         [9.3969e-01, 9.0301e-03, 6.1710e-03,  ..., 3.5786e-04,\n",
      "          5.9097e-04, 5.3556e-04],\n",
      "         [6.1931e-04, 8.2615e-03, 9.7595e-01,  ..., 2.3021e-03,\n",
      "          5.8235e-04, 4.0193e-04]],\n",
      "\n",
      "        [[5.7639e-03, 3.6574e-03, 7.0647e-03,  ..., 4.3900e-03,\n",
      "          8.3498e-03, 2.6439e-03],\n",
      "         [7.1809e-01, 3.2697e-02, 9.9916e-03,  ..., 5.4382e-03,\n",
      "          8.5135e-03, 4.9327e-03],\n",
      "         [3.2152e-03, 9.1135e-01, 2.5719e-02,  ..., 4.0236e-03,\n",
      "          3.6383e-03, 4.2431e-03],\n",
      "         ...,\n",
      "         [1.5045e-03, 2.2217e-03, 3.0938e-03,  ..., 1.6092e-03,\n",
      "          4.0718e-03, 1.5265e-03],\n",
      "         [9.1519e-01, 1.5085e-02, 4.2438e-03,  ..., 3.0597e-03,\n",
      "          6.4174e-03, 3.0304e-03],\n",
      "         [4.2478e-03, 7.2718e-03, 9.1162e-01,  ..., 4.8148e-03,\n",
      "          7.5715e-03, 4.5044e-03]]]) tensor([4., 0., 1.,  ..., 4., 0., 2.])\n"
     ]
    },
    {
     "name": "stderr",
     "output_type": "stream",
     "text": [
      "100%|██████████| 118/118 [00:11<00:00, 10.23it/s]\n"
     ]
    },
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "tensor([[[2.3612e-02, 6.2821e-02, 9.1537e-03,  ..., 7.0571e-03,\n",
      "          1.7813e-02, 1.0057e-02],\n",
      "         [1.0150e-01, 1.5576e-01, 1.9305e-01,  ..., 2.2215e-02,\n",
      "          1.5863e-02, 2.5863e-02],\n",
      "         [7.1256e-03, 3.4020e-01, 5.7956e-01,  ..., 5.7358e-03,\n",
      "          1.9503e-03, 1.2744e-03],\n",
      "         ...,\n",
      "         [5.7352e-03, 3.4354e-02, 7.1103e-03,  ..., 1.8499e-03,\n",
      "          6.5644e-03, 5.5551e-03],\n",
      "         [6.8194e-01, 1.0519e-01, 1.5120e-02,  ..., 6.2375e-03,\n",
      "          4.7257e-03, 6.7118e-03],\n",
      "         [4.3128e-03, 9.9219e-02, 8.3268e-01,  ..., 3.1577e-03,\n",
      "          9.7451e-04, 1.4951e-03]],\n",
      "\n",
      "        [[1.0491e-03, 1.9064e-03, 1.0744e-02,  ..., 1.3052e-03,\n",
      "          2.5086e-03, 3.8112e-04],\n",
      "         [8.7946e-01, 1.7661e-02, 4.4420e-03,  ..., 4.6465e-03,\n",
      "          1.4305e-03, 3.9944e-04],\n",
      "         [1.4280e-04, 9.8222e-01, 6.9996e-05,  ..., 8.1679e-03,\n",
      "          3.6283e-03, 9.9072e-04],\n",
      "         ...,\n",
      "         [1.7967e-03, 2.4337e-02, 7.5004e-03,  ..., 9.4930e-04,\n",
      "          5.2375e-04, 4.0302e-04],\n",
      "         [9.3969e-01, 9.0301e-03, 6.1710e-03,  ..., 3.5786e-04,\n",
      "          5.9097e-04, 5.3556e-04],\n",
      "         [6.1931e-04, 8.2615e-03, 9.7595e-01,  ..., 2.3021e-03,\n",
      "          5.8235e-04, 4.0193e-04]],\n",
      "\n",
      "        [[5.7639e-03, 3.6574e-03, 7.0647e-03,  ..., 4.3900e-03,\n",
      "          8.3498e-03, 2.6439e-03],\n",
      "         [7.1809e-01, 3.2697e-02, 9.9916e-03,  ..., 5.4382e-03,\n",
      "          8.5135e-03, 4.9327e-03],\n",
      "         [3.2152e-03, 9.1135e-01, 2.5719e-02,  ..., 4.0236e-03,\n",
      "          3.6383e-03, 4.2431e-03],\n",
      "         ...,\n",
      "         [1.5045e-03, 2.2217e-03, 3.0938e-03,  ..., 1.6092e-03,\n",
      "          4.0718e-03, 1.5265e-03],\n",
      "         [9.1519e-01, 1.5085e-02, 4.2438e-03,  ..., 3.0597e-03,\n",
      "          6.4174e-03, 3.0304e-03],\n",
      "         [4.2478e-03, 7.2718e-03, 9.1162e-01,  ..., 4.8148e-03,\n",
      "          7.5715e-03, 4.5044e-03]],\n",
      "\n",
      "        [[1.0408e-02, 4.9634e-03, 7.2577e-03,  ..., 7.4070e-03,\n",
      "          6.4214e-03, 9.3232e-03],\n",
      "         [8.2006e-01, 1.5136e-02, 8.9444e-03,  ..., 9.0015e-03,\n",
      "          7.1973e-03, 9.9827e-03],\n",
      "         [2.0956e-03, 9.3745e-01, 5.4091e-03,  ..., 3.7857e-03,\n",
      "          3.9113e-03, 2.7825e-03],\n",
      "         ...,\n",
      "         [6.2555e-03, 1.5194e-02, 6.6946e-03,  ..., 1.1248e-02,\n",
      "          4.5101e-03, 7.6977e-03],\n",
      "         [9.0131e-01, 6.3610e-03, 6.4459e-03,  ..., 4.0332e-03,\n",
      "          5.0155e-03, 8.1460e-03],\n",
      "         [2.7978e-03, 2.0815e-02, 9.2974e-01,  ..., 3.4942e-03,\n",
      "          3.1873e-03, 3.6920e-03]]]) tensor([4., 0., 1.,  ..., 4., 0., 2.])\n",
      "tensor([4, 0, 1,  ..., 4, 0, 2])\n",
      "F1: 0.8505911185129894, Average accuracy: 0.914989406779661\n"
     ]
    },
    {
     "data": {
      "text/plain": [
       "<matplotlib.axes._subplots.AxesSubplot at 0x7fbe534a5640>"
      ]
     },
     "execution_count": 45,
     "metadata": {},
     "output_type": "execute_result"
    },
    {
     "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": [
    "LOAD_PATH = '/opt/ml/result/Best/Team7/'\n",
    "best_models = os.listdir(LOAD_PATH)\n",
    "total_outputs = torch.tensor([])\n",
    "for m in best_models:\n",
    "    if 'ipynb' in m:\n",
    "        continue\n",
    "    if 'three' in m:\n",
    "        continue\n",
    "\n",
    "    if 'efficient' in m:\n",
    "        model = KKT_Effi(num_classes=18).to(device)\n",
    "    elif 'mask' in m:\n",
    "        # from KKH\n",
    "        model = EfficientNet.from_pretrained('efficientnet-b0', num_classes=18).to(device)\n",
    "    elif 'Resnet' in m:\n",
    "        model = ResNet().to(device)\n",
    "    else:\n",
    "        model = MyModel().to(device)\n",
    "    model.load_state_dict(torch.load(LOAD_PATH + m))\n",
    "\n",
    "    with torch.no_grad():\n",
    "        model.eval()\n",
    "        total_labels = torch.tensor([])\n",
    "        t_outputs = torch.tensor([])\n",
    "        for val_batch in tqdm(valid_dataloader):\n",
    "            inputs, labels = val_batch\n",
    "            outputs = F.softmax(model(inputs.to(device)).cpu(), dim=1)\n",
    "            t_outputs = torch.cat((t_outputs, outputs), dim=0)\n",
    "            total_labels = torch.cat((total_labels, labels), dim = 0)       \n",
    "        total_outputs =  torch.cat((total_outputs, t_outputs.unsqueeze(0)), dim=0)\n",
    "        \n",
    "    print(total_outputs, total_labels)    \n",
    "prediction_array = torch.argmax(total_outputs.mean(axis=0), dim=1)\n",
    "print(prediction_array)\n",
    "\n",
    "insert_array = {'TF': total_labels == prediction_array, 'label':total_labels}\n",
    "\n",
    "validation_eval = pd.DataFrame(None, columns = [\"TF\", \"label\"])\n",
    "validation_eval = validation_eval.append(pd.DataFrame(data=insert_array), ignore_index=True)\n",
    "\n",
    "accuracy = sklearn.metrics.accuracy_score(prediction_array, total_labels)\n",
    "f1 = sklearn.metrics.f1_score(prediction_array, total_labels, average=\"macro\")\n",
    "print('F1: {}, Average accuracy: {}'.format(f1, accuracy))\n",
    "\n",
    "sns.countplot(\n",
    "    data = validation_eval,\n",
    "    x = \"label\",\n",
    "    hue = \"TF\"\n",
    ")"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "f3a7453c-511a-44fa-ae08-6e0fe6ad8f39",
   "metadata": {},
   "source": [
    "### tta"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 26,
   "id": "d52b78b6-5a0e-4326-b7fa-0b3e2a089b91",
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "Loaded pretrained weights for efficientnet-b0\n"
     ]
    },
    {
     "name": "stderr",
     "output_type": "stream",
     "text": [
      "  0%|          | 0/118 [00:00<?, ?it/s]/tmp/ipykernel_68110/4121189927.py:43: UserWarning: Implicit dimension choice for softmax has been deprecated. Change the call to include dim=X as an argument.\n",
      "  outputs = F.softmax(outputs).cpu()\n",
      "100%|██████████| 118/118 [02:09<00:00,  1.09s/it]\n"
     ]
    },
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "tensor([[[0.0059, 0.0033, 0.0068,  ..., 0.0045, 0.0108, 0.0030],\n",
      "         [0.7792, 0.0259, 0.0067,  ..., 0.0048, 0.0067, 0.0043],\n",
      "         [0.0038, 0.9059, 0.0277,  ..., 0.0043, 0.0033, 0.0049],\n",
      "         ...,\n",
      "         [0.0014, 0.0017, 0.0024,  ..., 0.0013, 0.0037, 0.0010],\n",
      "         [0.8633, 0.0446, 0.0055,  ..., 0.0040, 0.0073, 0.0038],\n",
      "         [0.0046, 0.0083, 0.9043,  ..., 0.0051, 0.0082, 0.0051]]]) tensor([4., 0., 1.,  ..., 4., 0., 2.])\n"
     ]
    },
    {
     "name": "stderr",
     "output_type": "stream",
     "text": [
      "100%|██████████| 118/118 [01:01<00:00,  1.91it/s]\n"
     ]
    },
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "tensor([[[0.0059, 0.0033, 0.0068,  ..., 0.0045, 0.0108, 0.0030],\n",
      "         [0.7792, 0.0259, 0.0067,  ..., 0.0048, 0.0067, 0.0043],\n",
      "         [0.0038, 0.9059, 0.0277,  ..., 0.0043, 0.0033, 0.0049],\n",
      "         ...,\n",
      "         [0.0014, 0.0017, 0.0024,  ..., 0.0013, 0.0037, 0.0010],\n",
      "         [0.8633, 0.0446, 0.0055,  ..., 0.0040, 0.0073, 0.0038],\n",
      "         [0.0046, 0.0083, 0.9043,  ..., 0.0051, 0.0082, 0.0051]],\n",
      "\n",
      "        [[0.0117, 0.0052, 0.0074,  ..., 0.0081, 0.0061, 0.0089],\n",
      "         [0.8611, 0.0200, 0.0046,  ..., 0.0069, 0.0048, 0.0057],\n",
      "         [0.0022, 0.9342, 0.0073,  ..., 0.0040, 0.0039, 0.0033],\n",
      "         ...,\n",
      "         [0.0067, 0.0163, 0.0070,  ..., 0.0089, 0.0052, 0.0079],\n",
      "         [0.9188, 0.0054, 0.0055,  ..., 0.0036, 0.0055, 0.0070],\n",
      "         [0.0019, 0.0114, 0.9503,  ..., 0.0027, 0.0025, 0.0028]]]) tensor([4., 0., 1.,  ..., 4., 0., 2.])\n",
      "tensor([4, 0, 1,  ..., 4, 0, 2])\n",
      "F1: 0.8071558304968679, Average accuracy: 0.881885593220339\n"
     ]
    },
    {
     "data": {
      "text/plain": [
       "<matplotlib.axes._subplots.AxesSubplot at 0x7fbee3829520>"
      ]
     },
     "execution_count": 26,
     "metadata": {},
     "output_type": "execute_result"
    },
    {
     "data": {
      "image/png": "iVBORw0KGgoAAAANSUhEUgAAAYYAAAEGCAYAAABhMDI9AAAABHNCSVQICAgIfAhkiAAAAAlwSFlzAAALEgAACxIB0t1+/AAAADh0RVh0U29mdHdhcmUAbWF0cGxvdGxpYiB2ZXJzaW9uMy4yLjEsIGh0dHA6Ly9tYXRwbG90bGliLm9yZy+j8jraAAAdcElEQVR4nO3de5RcZZnv8e+PpGMjARKSJmI6saMiikRjaCQMDkfpAYVRAsr1oHQAzcwZ9OhBHCOuxYkcPOI46oCjuHJASSRgEA1kkNFg5KIol6RJuEUkIiGdyY2AUcBAIs/5Y78da3equqs7vatvv89avWrvd7/11LMrlXpqv3vXW4oIzMzMOuzV3wmYmdnA4sJgZmY5LgxmZpbjwmBmZjkuDGZmljOyvxPYE+PHj4+mpqb+TsPMbFBZsWLFMxHRUGn7oC4MTU1NLF++vL/TMDMbVCSt7Wq7h5LMzCzHhcHMzHJcGMzMLGdQn2Mw686OHTtob29n+/bt/Z1KTdTX19PY2EhdXV1/p2KDmAuDDWnt7e3su+++NDU1Iam/0ylURLB161ba29uZMmVKf6djg5iHkmxI2759O+PGjRvyRQFAEuPGjRs2R0dWHBcGG/KGQ1HoMJz21YrjwmBmZjk+x2C2h7Zu3UpLSwsAGzduZMSIETQ0ZF8qXbVqFW9/+9t39b355pvxt/VtoCusMEg6BFhU0vR64BJgQWpvAp4CTo+I55QdA18BnAi8CMyKiLai8hsqnr50apfbJ1/ycI0yGb7GjRvHypUrAZg7dy6jR4/moosuAmD06NG7tpkNFoUNJUXE4xExLSKmAYeTvdkvBuYAyyLiYGBZWgc4ATg4/c0GrioqNzMzq6xW5xhagN9FxFpgJjA/tc8HTk7LM4EFkbkXGCPpoBrlZ1aIP//5z0ybNo1p06Zxyimn9Hc6ZlWp1TmGM4Eb0vKEiNiQljcCE9LyRGBdyX3aU9uGkjYkzSY7omDy5MlF5WvWJ/bee28PJdmgU/gRg6RRwEnADzpvi4gAoifxImJeRDRHRHPHCT4zM+s7tRhKOgFoi4hNaX1TxxBRut2c2tcDk0ru15jazMyshmpRGM7ir8NIAEuA1rTcCtxS0n6OMjOAbSVDTmZmViOFnmOQtA9wHPAPJc2XAzdKOh9YC5ye2m8ju1R1DdkVTOcWmZtZEebOnZtbf/755/snEbM9UGhhiIgXgHGd2raSXaXUuW8AFxSZj5mZdc9TYpiZWY4Lg5mZ5bgwmJlZjguDmZnluDCYmVmOp922YeXwzyzo03grvnJOt31GjBjB1Kl/nQW3q6m3R48e7Utcrd+5MJgVzPMl2WDjoSSzGnv++edpaWlh+vTpTJ06lVtuuWW3Phs2bOCYY45h2rRpHHbYYfziF78AYOnSpRx11FFMnz6d0047zUcXVggXBrOCdZ56u76+nsWLF9PW1sYdd9zBpz/9abLvd/7V9ddfz3vf+15WrlzJqlWrmDZtGs888wyXXXYZP/vZz2hra6O5uZmvfe1r/bRXNpR5KMmsYJ2Hknbs2MHFF1/M3XffzV577cX69evZtGkTr3nNa3b1OeKIIzjvvPPYsWMHJ598MtOmTeOuu+7iscce4+ijjwbg5Zdf5qijjqr5/tjQ58JgVmMLFy5ky5YtrFixgrq6Opqamti+fXuuzzHHHMPdd9/Nj3/8Y2bNmsWFF17I2LFjOe6447jhhhsqRDbrGx5KMquxbdu2ceCBB1JXV8cdd9zB2rVrd+uzdu1aJkyYwMc+9jE++tGP0tbWxowZM7jnnntYs2YNAC+88AK//e1va52+DQM+YrBhpZrLS4t29tln84EPfICpU6fS3NzMm9/85t363HnnnXzlK1+hrq6O0aNHs2DBAhoaGrj22ms566yzeOmllwC47LLLeNOb3lTrXbAhzoXBrGCdrxwaP348v/71r7vs29raSmtr627bjz32WB544IG+T9KshIeSzMwsx4XBzMxyXBjMzCzHhcHMzHJcGMzMLMdXJdkuT186teK2yZc8XMNMzKw/FVoYJI0BrgYOAwI4D3gcWAQ0AU8Bp0fEc5IEXAGcCLwIzIqItiLzs+Gnq+LXG90VzK1bt9LS0gLAxo0bGTFiBA0NDQDcf//9jBo1qk/zMesLRR8xXAH8JCJOlTQKeDVwMbAsIi6XNAeYA3wWOAE4OP0dCVyVbs0GrXHjxu2aJ2nu3LmMHj2aiy66aNf2nTt3MnKkD9xtYCnsFSlpf+AYYBZARLwMvCxpJvDu1G0+cCdZYZgJLIhsmsl7JY2RdFBEbCgqR7P+MGvWLOrr63nwwQc5+uij2W+//XIF47DDDuPWW2+lqamJ6667jiuvvJKXX36ZI488km9961uMGDGin/fAhroiTz5PAbYA35X0oKSrJe0DTCh5s98ITEjLE4F1JfdvT205kmZLWi5p+ZYtWwpM36w47e3t/OpXv+py2uzVq1ezaNEi7rnnHlauXMmIESNYuHBhDbO04arIY9iRwHTgExFxn6QryIaNdomIkBRl711BRMwD5gE0Nzf36L5mA8Vpp53W7Sf/ZcuWsWLFCo444ggg+12HAw88sBbp2TBXZGFoB9oj4r60fhNZYdjUMUQk6SBgc9q+HphUcv/G1GY25Oyzzz67lkeOHMkrr7yya71jCu6IoLW1lS996Us1z8+Gt8KGkiJiI7BO0iGpqQV4DFgCdMwO1gp0/K7hEuAcZWYA23x+wYaDpqYm2tqyC/Da2tr4/e9/D0BLSws33XQTmzdnn52effbZslN0m/W1oi+H+ASwMF2R9CRwLlkxulHS+cBa4PTU9zayS1XXkF2uem7BudkwNBC/j/GhD32IBQsW8Na3vpUjjzxy1zTahx56KJdddhnHH388r7zyCnV1dXzzm9/kda97XT9nbENdoYUhIlYCzWU2tZTpG8AFReZj1p/mzp1btn3vvfdm6dKlZbedccYZnHHGGQVmZbY7T4lhZmY5LgxmZpbjwmBDXjZKOTwMp3214rgw2JBWX1/P1q1bh8UbZkSwdetW6uvr+zsVG+Q8SYsNaY2NjbS3tzNcviVfX19PY2Njf6dhg5wLgw1pdXV1TJkypb/TMBtUPJRkZmY5LgxmZpYz5IeSuvthloH4TVgzs/7kIwYzM8txYTAzsxwXBjMzy3FhMDOzHBcGMzPLcWEwM7McFwYzM8txYTAzsxwXBjMzy3FhMDOzHBcGMzPLKbQwSHpK0sOSVkpantoOkHS7pCfS7djULklXSloj6SFJ04vMzczMyqvFEcN7ImJaRDSn9TnAsog4GFiW1gFOAA5Of7OBq2qQm5mZddIfQ0kzgflpeT5wckn7gsjcC4yRdFA/5GdmNqwVXRgCWCpphaTZqW1CRGxIyxuBCWl5IrCu5L7tqS1H0mxJyyUtHy4/12hmVktF/x7DuyJivaQDgdsl/aZ0Y0SEpB79SntEzAPmATQ3Nw/9X3g3M6uxQo8YImJ9ut0MLAbeCWzqGCJKt5tT9/XApJK7N6Y2MzOrocIKg6R9JO3bsQwcDzwCLAFaU7dW4Ja0vAQ4J12dNAPYVjLkZGZmNVLkUNIEYLGkjse5PiJ+IukB4EZJ5wNrgdNT/9uAE4E1wIvAuQXmZmZmFRRWGCLiSeDtZdq3Ai1l2gO4oKh8zMysOv7ms5mZ5bgwmJlZjguDmZnluDCYmVmOC4OZmeW4MJiZWY4Lg5mZ5bgwmJlZjguDmZnluDCYmVmOC4OZmeW4MJiZWY4Lg5mZ5bgwmJlZjguDmZnluDCYmVmOC4OZmeW4MJiZWY4Lg5mZ5bgwmJlZTlWFQdKyatoq3HeEpAcl3ZrWp0i6T9IaSYskjUrtr0rra9L2pup3w8zM+kqXhUFSvaQDgPGSxko6IP01AROrfIxPAqtL1r8MfD0i3gg8B5yf2s8HnkvtX0/9zMysxro7YvgHYAXw5nTb8XcL8O/dBZfUCPw9cHVaF3AscFPqMh84OS3PTOuk7S2pv5mZ1dDIrjZGxBXAFZI+ERHf6EX8fwP+Gdg3rY8D/hARO9N6O3898pgIrEuPu1PSttT/mdKAkmYDswEmT57ci5TMzKwrXRaGDhHxDUl/AzSV3iciFlS6j6T3A5sjYoWkd+9hnqW5zAPmATQ3N0dfxTUzs0xVhUHS94A3ACuBv6TmACoWBuBo4CRJJwL1wH7AFcAYSSPTUUMjsD71Xw9MAtoljQT2B7b2bHfMzGxPVVUYgGbg0Iio+hN6RHwO+BxAOmK4KCLOlvQD4FTg+0Ar2fkKgCVp/ddp+8978nhmZtY3qv0ewyPAa/roMT8LXChpDdk5hGtS+zXAuNR+ITCnjx7PzMx6oNojhvHAY5LuB17qaIyIk6q5c0TcCdyZlp8E3lmmz3bgtCrzMTOzglRbGOYWmYSZmQ0c1V6VdFfRiZiZ2cBQ7VVJfyK7CglgFFAHvBAR+xWVmJmZ9Y9qjxg6vqDW8e3lmcCMopIyM7P+0+PZVSNzM/DeAvIxM7N+Vu1Q0gdLVvci+17D9kIyMjOzflXtVUkfKFneCTxFNpxkZmZDTLXnGM4tOhEzMxsYqv2hnkZJiyVtTn8/TFNqm5nZEFPtyefvks1l9Nr09x+pzczMhphqC0NDRHw3Inamv2uBhgLzMjOzflJtYdgq6cPp95tHSPownhLbzGxIqrYwnAecDmwENpBNiz2roJzMzKwfVXu56qVAa0Q8ByDpAOBfyQqGmZkNIdUeMbytoygARMSzwDuKScnMzPpTtYVhL0ljO1bSEUO1RxtmZjaIVPvm/lXg1+lnOSH7QZ0vFpOSmZn1p2q/+bxA0nLg2NT0wYh4rLi0zMysv1Q9HJQKgYuBmdkQ1+Npt83MbGgrrDBIqpd0v6RVkh6V9IXUPkXSfZLWSFokaVRqf1VaX5O2NxWVm5mZVVbklUUvAcdGxPOS6oBfSvpP4ELg6xHxfUnfBs4Hrkq3z0XEGyWdCXwZOKPA/Hrs6UunVtw2+ZKHa5iJmVlxCjtiSL/09nxarUt/QXYC+6bUPh84OS3PTOuk7S3pZ0TNzKyGCj3HkOZVWglsBm4Hfgf8ISJ2pi7twMS0PBFYB5C2bwPGlYk5W9JyScu3bNlSZPpmZsNSoYUhIv4SEdOARuCdwJv7IOa8iGiOiOaGBk/wambW12pyVVJE/AG4AzgKGCOp49xGI7A+La8HJgGk7fvjGVzNzGquyKuSGiSNSct7A8cBq8kKxKmpWytwS1pektZJ238eEVFUfmZmVl6RVyUdBMyXNIKsAN0YEbdKegz4vqTLgAeBa1L/a4DvSVoDPAucWWBuZmZWQWGFISIeoswMrBHxJNn5hs7t28nmYDIzs37kbz6bmVmOC4OZmeW4MJiZWY4Lg5mZ5bgwmJlZjguDmZnluDCYmVmOC4OZmeW4MJiZWY4Lg5mZ5bgwmJlZjguDmZnluDCYmVmOC4OZmeW4MJiZWY4Lg5mZ5bgwmJlZjguDmZnluDCYmVmOC4OZmeUUVhgkTZJ0h6THJD0q6ZOp/QBJt0t6It2OTe2SdKWkNZIekjS9qNzMzKyyIo8YdgKfjohDgRnABZIOBeYAyyLiYGBZWgc4ATg4/c0GriowNzMzq6CwwhARGyKiLS3/CVgNTARmAvNTt/nAyWl5JrAgMvcCYyQdVFR+ZmZW3shaPIikJuAdwH3AhIjYkDZtBCak5YnAupK7tae2DVifOPwzC7rcvnjfGiViZgNa4SefJY0Gfgh8KiL+WLotIgKIHsabLWm5pOVbtmzpw0zNzAwKLgyS6siKwsKI+FFq3tQxRJRuN6f29cCkkrs3praciJgXEc0R0dzQ0FBc8mZmw1SRVyUJuAZYHRFfK9m0BGhNy63ALSXt56Srk2YA20qGnMzMrEaKPMdwNPAR4GFJK1PbxcDlwI2SzgfWAqenbbcBJwJrgBeBcwvMzczMKiisMETELwFV2NxSpn8AFxSVj5mZVcfffDYzsxwXBjMzy3FhMDOzHBcGMzPLcWEwM7McFwYzM8txYTAzsxwXBjMzy3FhMDOzHBcGMzPLcWEwM7OcmvxQj5nZUPP0pVO73D75kodrlEnfc2Ew60ddvbkM5jcWG9w8lGRmZjkuDGZmluOhJDOzfjbQhhR9xGBmZjkuDGZmluOhJLMhYChfOmm15yMGMzPL8RGDDWgD7aSc2XBQWGGQ9B3g/cDmiDgstR0ALAKagKeA0yPiOUkCrgBOBF4EZkVEW1G5mVltuLAPTkUeMVwL/DuwoKRtDrAsIi6XNCetfxY4ATg4/R0JXJVubRDyeLfZ4FbYOYaIuBt4tlPzTGB+Wp4PnFzSviAy9wJjJB1UVG5mZlZZrc8xTIiIDWl5IzAhLU8E1pX0a09tG+hE0mxgNsDkyZP7PMHDP7Og4rbF+/b5w5mZDTj9dlVSRAQQvbjfvIhojojmhoaGAjIzMxveal0YNnUMEaXbzal9PTCppF9jajMzsxqrdWFYArSm5VbglpL2c5SZAWwrGXIyM7MaKvJy1RuAdwPjJbUD/xu4HLhR0vnAWuD01P02sktV15BdrnpuUXmZmVnXCisMEXFWhU0tZfoGcEFRuQxEvr7bzAYqT4lhZmY5LgxmZpbjwmBmZjkuDGZmluPZVW3Y8Al/s+q4MJjZgOeJGWvLQ0lmZpbjwmBmZjkuDGZmluPCYGZmOT75bDZI+LdCrFZ8xGBmZjkuDGZmluOhJDOzGhhMQ4E+YjAzsxwXBjMzy/FQkpkNCINpqGUg64s5wVwYbMjo6o0F/OZSLU82aC4MZj3kCd0GNxe+7rkwDAI+xDazWhpQJ58lvU/S45LWSJrT3/mYmQ1HA6YwSBoBfBM4ATgUOEvSof2blZnZ8DOQhpLeCayJiCcBJH0fmAk81t0dPdRSe37Oq+MT4jYYKSL6OwcAJJ0KvC8iPprWPwIcGREf79RvNjA7rR4CPN5N6PHAM32U5kCM5ZxqH8s51T6Wc+rbWK+LiIZKGwfSEUNVImIeMK/a/pKWR0RzXzz2QIzlnGofyznVPpZzqm2sAXOOAVgPTCpZb0xtZmZWQwOpMDwAHCxpiqRRwJnAkn7Oycxs2BkwQ0kRsVPSx4GfAiOA70TEo30Quuphp0EayznVPpZzqn0s51TDWAPm5LOZmQ0MA2koyczMBgAXBjMzyxkyhaG76TQkvUrSorT9PklNFeJ8R9JmSY9U2C5JV6Y4D0maXqHfJEl3SHpM0qOSPrkHseol3S9pVYr1hd7uX+o7QtKDkm7dwzhPSXpY0kpJy3u7f6nvGEk3SfqNpNWSjuppLEmHpFw6/v4o6VN7kNP/Ss/3I5JukFTfm+dK0idTjEc759NdTuVej5IOkHS7pCfS7dgKj9ua+jyRlsvFOi3l9Yqkipc4qtP/rwqxvpL+/R6StFjSmCpiPVAmzv9JMVZKWirptb3dv5K+n5YUksZXEevuMjnNlbS+5LV1Ym+fp9TvE+m5elTSv/Q2Vnr9deT0lKSV1cQq1ycnIgb9H9nJ6t8BrwdGAauAQzv1+Sfg22n5TGBRhVjHANOBRypsPxH4T0DADOC+Cv0OAqan5X2B35bJqdpYAkan5TrgPmBGb/Yvbb8QuB64tcy2nsR5Chjfxfaq9i/1nQ98NC2PAsb0NlbJa2Ij2Rd5evOcTwR+D+yd1m8EZvX0uQIOAx4BXk12scfPgDdWm1O51yPwL8CctDwH+HKZxz0AeDLdjk3LJ5aJ9RayL4reCTT34P/XR8rEOh4YmZa/XCGvzrHWAB/qFGe/kuX/2fEc92b/Ut9JZBe1rC33ei0T67+A/9Ypp7nARVW85qp5nt6TXgevSusH9jZWp/t8FbikyliHdrUvQ+WIYdd0GhHxMtAxnUapmWRvPgA3AS2S1DlQRNwNPNvFY80EFkTmXmCMpIPKxNkQEW1p+U/AarI3m97Eioh4Pq3Wpb/OVw1UtX+SGoG/B67uYv+6jVOlqvZP0v5kb4DXAETEyxHxh97EKtEC/C4i1u5BnJHA3pJGkr2x/1eZWN09V28he6N/MSJ2AncBH6w2pwqvx9LHnQ+cXCb39wK3R8SzEfEccDuwf+dYEbE6IrqbPaDc/6/GMrGWpn0EuDf16S7WNamtNM4fS1b3YffXetX7l3wd+OcKccrF+g/gbRX6dqWq5wn4H8DlEfESQERs3oNYQHbUCZwO3FBlrM7vjzlDpTBMBNaVrLez+5vwrj7pxbsNGFfQY+WkIYZ3kH3S71UsZcM/K4HNZC/iirG62b9/I/tP8kqFdHvyPAWwVNIKZVOVVIyVVNq/KcAW4LvKhriulrRPL2N1OJPy/0mqihMR64F/BZ4GNgDbImJppVhdPFePAH8raZykV5N9qp3UqU9P921CRGxIyxuBCWX69Ph12oXexDqP7Ciomliv6dxJ0hclrQPOBi7pbU6SZgLrI2JVF7lWlRPw8TTE9Z0Kw3fVPk9vIntN3CfpLklH7EGsDn8LbIqIJ/og1pApDAOWpNHAD4FPdfok1CMR8ZeImEb2qeGdkg7rRS7vBzZHxIre5tHJuyJiOtmMuBdIOqaXcUaSHSJfFRHvAF4gGyLpFWVfkDwJ+MEexBhL9qlqCvBaYB9JH+5pnIhYTTasshT4CbAS+Etv8yoTP6j8SbhfSPo8sBNY2NsYEfH5iJiUYny8u/4V8ng1cDHlC0tPXQW8AZhG9kHhq3sQayTZsNUM4DPAjXtwVN7hLMp/EOqVoVIYqplOY1efNDSwP7C1oMciPU4dWVFYGBE/2pNYHdIQyx3A+yrF6mL/jgZOkvQU2eHksZKu60WcjlzWp9vNwGI6DQn0YP/agfaSo6CbyApFb2JBVqjaImJTmW3Vxvk74PcRsSUidgA/Av6mUqyunquIuCYiDo+IY4DnyM439XbfADZ1DDWl23JDEX05xUxPXvOzgPcDZ6eiVU2sjV089kKycxC9yekNZIV9VXrNNwJtkjofDXSbU0RsSh/OXgH+H7u/1qvNCbLX+4/S0OH9ZEfvnU+K9+Q5H0k2PLmo3PaexOowVApDNdNpLAFa0/KpwM8rvHC7swQ4R5kZZEMMGzp3Sp8ArgFWR8TX9jBWg9IVHpL2Bo4DftPT/YuIz0VEY0Q0kT1HP4+Izp+Cq3qeJO0jad+OZbITj52vBqlq/yJiI7BO0iGpqYXdp1uvKlbS1aenauM8DcyQ9Or0b9lCdp6oc6xqnqsD0+1ksv/A1+/BvnV+3FbgljJ9fgocL2lsOvo5PrX1RlXT1Uh6H9kw5UkR8WIPYv2sU5yDS1ZnsvtrHarYv4h4OCIOjIim9JpvJ7sgpHMhKhfr7k45lZ6HOoXdX+uV9q3ctD43k52ARtKbyE4Id54NtSdTBP0d8JuIaK+wvefTDUUXZ6YH0x/Z2O1vyc6+fz61XUr2IgWoJxtaWAPcD7y+QpwbyA4Vd5C9kM4H/hH4x7RdZD8o9DvgYSpfyfEuskP8h8iGD1amHHsT623AgynWI6QrD3qzfyUx3026KqmXz9Prya5uWAU8WvKc93j/Ut9pwPK0jzeTXR3Sm+dqH7JP7fuXtPU2py+QvSk9AnwPeFUvn6tfkBW6VUBLT3Kq8HocBywDniB7Uz0g9W0Gri6573kptzXAuRVinZKWXwI2AT9N930tcFul/18VYq0hG8vueL1/u4pYq8rE+WF6zh8iOxE8sbf71+nf4SnSVUndxLq3TE7fS/82D5G9qR60B8/TKOC6tI9twLG9jZX6XUt6LZXct8tY3b2fekoMMzPLGSpDSWZm1kdcGMzMLMeFwczMclwYzMwsx4XBzMxyXBjMqiTp+W62N6nCrLxd3OdaSafuWWZmfcuFwczMclwYzHpI0mhJyyS1Kfs9itKZKkdKWqjsNyVuSvP1IOnwNGHaCkk/Vdczw5r1KxcGs57bDpwS2QSC7wG+WjIJ2iHAtyLiLcAfgX9Kc2Z9Azg1Ig4HvgN8sR/yNqvKyP5OwGwQEvB/02yyr5BNYdwx9fW6iLgnLV9H9kMzPyH7wZ7bU/0YQTa1gdmA5MJg1nNnAw3A4RGxI83c2fGzn53nmAmyQvJoRByF2SDgoSSzntuf7Hctdkh6D/C6km2T9dffq/7vwC+Bx4GGjnZJdZLeWtOMzXrAhcGs5xYCzZIeBs4hPy3042Q/WrSabIbYqyL7OcVTgS9LWkU282jn33YwGzA8u6qZmeX4iMHMzHJcGMzMLMeFwczMclwYzMwsx4XBzMxyXBjMzCzHhcHMzHL+P/TlcEbDtQU1AAAAAElFTkSuQmCC\n",
      "text/plain": [
       "<Figure size 432x288 with 1 Axes>"
      ]
     },
     "metadata": {
      "needs_background": "light"
     },
     "output_type": "display_data"
    }
   ],
   "source": [
    "tf = tta.Compose(\n",
    "[\n",
    "    tta.HorizontalFlip(),\n",
    "    tta.Rotate90(angles=[0,15]),\n",
    "    tta.Multiply(factors=[0.8,0.9,1,1.1]),\n",
    "])\n",
    "\n",
    "LOAD_PATH = '/opt/ml/result/Best/Team5/'\n",
    "best_models = os.listdir(LOAD_PATH)\n",
    "total_outputs = torch.tensor([])\n",
    "\n",
    "for m in best_models:\n",
    "    if 'ipynb' in m:\n",
    "        continue\n",
    "    if 'three' in m:\n",
    "        continue   \n",
    "        \n",
    "    elif 'efficient' in m:\n",
    "        model = KKT_Effi(num_classes=18).to(device)\n",
    "    elif 'mask' in m:\n",
    "        model = EfficientNet.from_pretrained('efficientnet-b0', num_classes=18).to(device)\n",
    "    elif 'Resnet' in m:\n",
    "        model = ResNet().to(device)\n",
    "    else:\n",
    "        model = MyModel().to(device)\n",
    "    model.load_state_dict(torch.load(LOAD_PATH + m))\n",
    "    tta_model = tta.ClassificationTTAWrapper(model,tf)\n",
    "\n",
    "    with torch.no_grad():\n",
    "        model.eval()\n",
    "        total_labels = torch.tensor([])\n",
    "        t_outputs = torch.tensor([])\n",
    "        for val_batch in tqdm(valid_dataloader):\n",
    "            inputs, labels = val_batch\n",
    "            outputs = tta_model(inputs.to(device))\n",
    "            outputs = F.softmax(outputs).cpu()\n",
    "            t_outputs = torch.cat((t_outputs, outputs), dim=0)\n",
    "            total_labels = torch.cat((total_labels, labels), dim = 0)       \n",
    "        total_outputs =  torch.cat((total_outputs, t_outputs.unsqueeze(0)), dim=0)\n",
    "        \n",
    "    print(total_outputs, total_labels)    \n",
    "prediction_array = torch.argmax(total_outputs.mean(axis=0), dim=1)\n",
    "print(prediction_array)\n",
    "\n",
    "insert_array = {'TF': total_labels == prediction_array, 'label':total_labels}\n",
    "\n",
    "validation_eval = pd.DataFrame(None, columns = [\"TF\", \"label\"])\n",
    "validation_eval = validation_eval.append(pd.DataFrame(data=insert_array), ignore_index=True)\n",
    "\n",
    "accuarcy = sklearn.metrics.accuracy_score(prediction_array, total_labels)\n",
    "f1 = sklearn.metrics.f1_score(prediction_array, total_labels, average=\"macro\")\n",
    "print('F1: {}, Average accuracy: {}'.format(f1, accuarcy))\n",
    "\n",
    "#model = ExampleModel()\n",
    "#model.load_state_dict(torch.load(LOAD_PATH + os.listdir(LOAD_PATH)[0]))\n",
    "sns.countplot(\n",
    "    data = validation_eval,\n",
    "    x = \"label\",\n",
    "    hue = \"TF\"\n",
    ")"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "4e68780b-4608-491f-a573-f264ffa50bcf",
   "metadata": {},
   "source": [
    "### inference evaluation"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "17e485d8-3229-45c4-bfa9-79622bc5e354",
   "metadata": {},
   "source": [
    "#### confusion matrix"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 24,
   "id": "57864715-1c24-4bcc-bce9-6cc49ffe1ab7",
   "metadata": {},
   "outputs": [],
   "source": [
    "class_name = ['WM<30', 'WM<60', 'WM>60', 'WF<30', 'WF<60', 'WF>60', 'IM<30', 'IM<60', 'IM>60', 'IF<30', 'IF<60', 'IF>60', 'NM<30', 'NM<60', 'NM>60', 'NF<30', 'NF<60', 'NF>60']\n",
    "def plot_confusion_matrix(cm, target_names=class_name, cmap=None, normalize=True, labels=True, title='Confusion matrix'):\n",
    "    accuracy = np.trace(cm) / float(np.sum(cm))\n",
    "    misclass = 1 - accuracy\n",
    "\n",
    "    if cmap is None:\n",
    "        cmap = plt.get_cmap('Blues')\n",
    "\n",
    "    if normalize:\n",
    "        cm = cm.astype('float') / cm.sum(axis=1)[:, np.newaxis]\n",
    "        \n",
    "    plt.figure(figsize=(15, 15))\n",
    "    plt.imshow(cm, interpolation='nearest', cmap=cmap)\n",
    "    plt.title(title)\n",
    "    plt.colorbar()\n",
    "\n",
    "    thresh = cm.max() / 1.5 if normalize else cm.max() / 2\n",
    "    \n",
    "    if target_names is not None:\n",
    "        tick_marks = np.arange(len(target_names))\n",
    "        plt.xticks(tick_marks, target_names)\n",
    "        plt.yticks(tick_marks, target_names)\n",
    "    \n",
    "    if labels:\n",
    "        for i, j in itertools.product(range(cm.shape[0]), range(cm.shape[1])):\n",
    "            if normalize:\n",
    "                plt.text(j, i, \"{:0.4f}\".format(cm[i, j]),\n",
    "                         horizontalalignment=\"center\",\n",
    "                         color=\"white\" if cm[i, j] > thresh else \"black\")\n",
    "            else:\n",
    "                plt.text(j, i, \"{:,}\".format(cm[i, j]),\n",
    "                         horizontalalignment=\"center\",\n",
    "                         color=\"white\" if cm[i, j] > thresh else \"black\")\n",
    "\n",
    "    plt.tight_layout()\n",
    "    plt.ylabel('True label')\n",
    "    plt.xlabel('Predicted label\\naccuracy={:0.4f}; misclass={:0.4f}'.format(accuracy, misclass))\n",
    "    plt.show()"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 25,
   "id": "b82971ac-b50f-4cba-89b0-8599ccd684be",
   "metadata": {},
   "outputs": [
    {
     "data": {
      "image/png": "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\n",
      "text/plain": [
       "<Figure size 1080x1080 with 2 Axes>"
      ]
     },
     "metadata": {
      "needs_background": "light"
     },
     "output_type": "display_data"
    }
   ],
   "source": [
    "confusion_mat = confusion_matrix(total_labels, prediction_array)\n",
    "plot_confusion_matrix(confusion_mat)"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "a0358447-bff1-4f66-8b57-735b4d3d6889",
   "metadata": {},
   "source": [
    "## Submission"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 28,
   "id": "3f515462-0024-4007-979e-efb6d43accc3",
   "metadata": {},
   "outputs": [],
   "source": [
    "class TestDataset(Dataset):\n",
    "    def __init__(self, img_paths, transform=None):\n",
    "        self.img_paths = img_paths\n",
    "        self.transform = transform\n",
    "\n",
    "    def __getitem__(self, index):\n",
    "        image = Image.open(self.img_paths[index])\n",
    "\n",
    "        if self.transform:\n",
    "            image = self.transform(image)\n",
    "            #image = cv2.imread(self.img_paths[index])\n",
    "            #image = cv2.cvtColor(image, cv2.COLOR_BGR2RGB)\n",
    "            #augmented = self.transform(image=image)\n",
    "            #image = augmented['image']\n",
    "        return image\n",
    "\n",
    "    def __len__(self):\n",
    "        return len(self.img_paths)"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "6501073f-e432-45bb-b773-bc4f7b087017",
   "metadata": {},
   "source": [
    "### ensemble"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 51,
   "id": "c2aff66f-0f07-4cc7-97b0-94565cce2d41",
   "metadata": {},
   "outputs": [
    {
     "name": "stderr",
     "output_type": "stream",
     "text": [
      "100%|██████████| 12600/12600 [03:39<00:00, 57.35it/s]\n"
     ]
    },
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "tensor([[[0.0018, 0.0032, 0.0031,  ..., 0.0030, 0.0051, 0.0022],\n",
      "         [0.0019, 0.0021, 0.9514,  ..., 0.0021, 0.0026, 0.0015],\n",
      "         [0.0102, 0.0106, 0.0174,  ..., 0.0049, 0.0085, 0.0833],\n",
      "         ...,\n",
      "         [0.0043, 0.0041, 0.0049,  ..., 0.0038, 0.0025, 0.0038],\n",
      "         [0.1255, 0.2824, 0.0440,  ..., 0.0132, 0.0149, 0.0076],\n",
      "         [0.0023, 0.0013, 0.0013,  ..., 0.0019, 0.0017, 0.0013]]])\n"
     ]
    },
    {
     "name": "stderr",
     "output_type": "stream",
     "text": [
      "100%|██████████| 12600/12600 [04:08<00:00, 50.65it/s]\n"
     ]
    },
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "tensor([[[0.0018, 0.0032, 0.0031,  ..., 0.0030, 0.0051, 0.0022],\n",
      "         [0.0019, 0.0021, 0.9514,  ..., 0.0021, 0.0026, 0.0015],\n",
      "         [0.0102, 0.0106, 0.0174,  ..., 0.0049, 0.0085, 0.0833],\n",
      "         ...,\n",
      "         [0.0043, 0.0041, 0.0049,  ..., 0.0038, 0.0025, 0.0038],\n",
      "         [0.1255, 0.2824, 0.0440,  ..., 0.0132, 0.0149, 0.0076],\n",
      "         [0.0023, 0.0013, 0.0013,  ..., 0.0019, 0.0017, 0.0013]],\n",
      "\n",
      "        [[0.0035, 0.0136, 0.0184,  ..., 0.0175, 0.0152, 0.0312],\n",
      "         [0.0244, 0.4102, 0.3360,  ..., 0.0136, 0.0084, 0.0053],\n",
      "         [0.0145, 0.0227, 0.0460,  ..., 0.0202, 0.0426, 0.1392],\n",
      "         ...,\n",
      "         [0.0174, 0.0188, 0.0352,  ..., 0.1731, 0.0480, 0.0387],\n",
      "         [0.0339, 0.5603, 0.0683,  ..., 0.0140, 0.0136, 0.0062],\n",
      "         [0.0114, 0.0303, 0.0318,  ..., 0.0157, 0.0142, 0.0085]]])\n"
     ]
    },
    {
     "name": "stderr",
     "output_type": "stream",
     "text": [
      "100%|██████████| 12600/12600 [02:29<00:00, 84.56it/s] \n"
     ]
    },
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "tensor([[[1.8466e-03, 3.2066e-03, 3.1246e-03,  ..., 2.9559e-03,\n",
      "          5.1358e-03, 2.1898e-03],\n",
      "         [1.9368e-03, 2.1377e-03, 9.5143e-01,  ..., 2.1427e-03,\n",
      "          2.5556e-03, 1.4822e-03],\n",
      "         [1.0178e-02, 1.0635e-02, 1.7370e-02,  ..., 4.9179e-03,\n",
      "          8.5292e-03, 8.3256e-02],\n",
      "         ...,\n",
      "         [4.2746e-03, 4.0539e-03, 4.9432e-03,  ..., 3.8013e-03,\n",
      "          2.4913e-03, 3.8460e-03],\n",
      "         [1.2545e-01, 2.8241e-01, 4.4017e-02,  ..., 1.3198e-02,\n",
      "          1.4853e-02, 7.5744e-03],\n",
      "         [2.3306e-03, 1.3460e-03, 1.2530e-03,  ..., 1.9334e-03,\n",
      "          1.7035e-03, 1.3387e-03]],\n",
      "\n",
      "        [[3.4531e-03, 1.3579e-02, 1.8435e-02,  ..., 1.7488e-02,\n",
      "          1.5155e-02, 3.1185e-02],\n",
      "         [2.4402e-02, 4.1017e-01, 3.3600e-01,  ..., 1.3555e-02,\n",
      "          8.4229e-03, 5.2816e-03],\n",
      "         [1.4511e-02, 2.2724e-02, 4.5953e-02,  ..., 2.0182e-02,\n",
      "          4.2594e-02, 1.3919e-01],\n",
      "         ...,\n",
      "         [1.7384e-02, 1.8800e-02, 3.5235e-02,  ..., 1.7308e-01,\n",
      "          4.7977e-02, 3.8674e-02],\n",
      "         [3.3917e-02, 5.6035e-01, 6.8294e-02,  ..., 1.4040e-02,\n",
      "          1.3618e-02, 6.1602e-03],\n",
      "         [1.1399e-02, 3.0302e-02, 3.1831e-02,  ..., 1.5701e-02,\n",
      "          1.4164e-02, 8.5059e-03]],\n",
      "\n",
      "        [[2.1133e-04, 1.6567e-03, 1.4845e-03,  ..., 1.8283e-02,\n",
      "          5.2673e-02, 7.5381e-04],\n",
      "         [2.6317e-04, 9.2749e-01, 1.1157e-02,  ..., 4.9185e-03,\n",
      "          2.3264e-03, 1.1106e-03],\n",
      "         [3.7110e-03, 2.1540e-03, 2.0524e-03,  ..., 1.2794e-02,\n",
      "          4.6781e-02, 4.8506e-02],\n",
      "         ...,\n",
      "         [3.2397e-03, 1.0487e-03, 1.0785e-03,  ..., 1.3837e-03,\n",
      "          2.1254e-03, 7.8987e-04],\n",
      "         [1.2624e-02, 8.9391e-01, 3.7157e-02,  ..., 2.1393e-02,\n",
      "          8.6200e-04, 4.0946e-04],\n",
      "         [1.3661e-03, 7.7844e-04, 1.4083e-03,  ..., 6.1253e-02,\n",
      "          1.0504e-02, 1.9007e-03]]])\n",
      "Loaded pretrained weights for efficientnet-b0\n"
     ]
    },
    {
     "name": "stderr",
     "output_type": "stream",
     "text": [
      "100%|██████████| 12600/12600 [05:05<00:00, 41.28it/s]\n"
     ]
    },
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "tensor([[[1.8466e-03, 3.2066e-03, 3.1246e-03,  ..., 2.9559e-03,\n",
      "          5.1358e-03, 2.1898e-03],\n",
      "         [1.9368e-03, 2.1377e-03, 9.5143e-01,  ..., 2.1427e-03,\n",
      "          2.5556e-03, 1.4822e-03],\n",
      "         [1.0178e-02, 1.0635e-02, 1.7370e-02,  ..., 4.9179e-03,\n",
      "          8.5292e-03, 8.3256e-02],\n",
      "         ...,\n",
      "         [4.2746e-03, 4.0539e-03, 4.9432e-03,  ..., 3.8013e-03,\n",
      "          2.4913e-03, 3.8460e-03],\n",
      "         [1.2545e-01, 2.8241e-01, 4.4017e-02,  ..., 1.3198e-02,\n",
      "          1.4853e-02, 7.5744e-03],\n",
      "         [2.3306e-03, 1.3460e-03, 1.2530e-03,  ..., 1.9334e-03,\n",
      "          1.7035e-03, 1.3387e-03]],\n",
      "\n",
      "        [[3.4531e-03, 1.3579e-02, 1.8435e-02,  ..., 1.7488e-02,\n",
      "          1.5155e-02, 3.1185e-02],\n",
      "         [2.4402e-02, 4.1017e-01, 3.3600e-01,  ..., 1.3555e-02,\n",
      "          8.4229e-03, 5.2816e-03],\n",
      "         [1.4511e-02, 2.2724e-02, 4.5953e-02,  ..., 2.0182e-02,\n",
      "          4.2594e-02, 1.3919e-01],\n",
      "         ...,\n",
      "         [1.7384e-02, 1.8800e-02, 3.5235e-02,  ..., 1.7308e-01,\n",
      "          4.7977e-02, 3.8674e-02],\n",
      "         [3.3917e-02, 5.6035e-01, 6.8294e-02,  ..., 1.4040e-02,\n",
      "          1.3618e-02, 6.1602e-03],\n",
      "         [1.1399e-02, 3.0302e-02, 3.1831e-02,  ..., 1.5701e-02,\n",
      "          1.4164e-02, 8.5059e-03]],\n",
      "\n",
      "        [[2.1133e-04, 1.6567e-03, 1.4845e-03,  ..., 1.8283e-02,\n",
      "          5.2673e-02, 7.5381e-04],\n",
      "         [2.6317e-04, 9.2749e-01, 1.1157e-02,  ..., 4.9185e-03,\n",
      "          2.3264e-03, 1.1106e-03],\n",
      "         [3.7110e-03, 2.1540e-03, 2.0524e-03,  ..., 1.2794e-02,\n",
      "          4.6781e-02, 4.8506e-02],\n",
      "         ...,\n",
      "         [3.2397e-03, 1.0487e-03, 1.0785e-03,  ..., 1.3837e-03,\n",
      "          2.1254e-03, 7.8987e-04],\n",
      "         [1.2624e-02, 8.9391e-01, 3.7157e-02,  ..., 2.1393e-02,\n",
      "          8.6200e-04, 4.0946e-04],\n",
      "         [1.3661e-03, 7.7844e-04, 1.4083e-03,  ..., 6.1253e-02,\n",
      "          1.0504e-02, 1.9007e-03]],\n",
      "\n",
      "        [[9.0971e-03, 6.9135e-03, 9.3023e-03,  ..., 8.2400e-03,\n",
      "          1.5840e-02, 1.0077e-02],\n",
      "         [2.4323e-03, 3.9783e-02, 8.9570e-01,  ..., 2.1321e-03,\n",
      "          3.8855e-03, 2.7370e-03],\n",
      "         [3.2256e-02, 2.4843e-02, 2.5653e-02,  ..., 2.0978e-02,\n",
      "          7.1687e-02, 1.1807e-01],\n",
      "         ...,\n",
      "         [1.2222e-02, 6.4141e-03, 5.1741e-03,  ..., 9.0728e-03,\n",
      "          5.9941e-03, 3.3303e-03],\n",
      "         [3.9867e-02, 1.3828e-01, 5.9923e-01,  ..., 1.3676e-02,\n",
      "          1.3094e-02, 1.6890e-02],\n",
      "         [4.0860e-03, 1.3022e-03, 2.8554e-03,  ..., 3.5527e-03,\n",
      "          3.9563e-03, 4.0613e-03]]])\n"
     ]
    },
    {
     "name": "stderr",
     "output_type": "stream",
     "text": [
      "100%|██████████| 12600/12600 [02:05<00:00, 100.29it/s]\n"
     ]
    },
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "tensor([[[1.8466e-03, 3.2066e-03, 3.1246e-03,  ..., 2.9559e-03,\n",
      "          5.1358e-03, 2.1898e-03],\n",
      "         [1.9368e-03, 2.1377e-03, 9.5143e-01,  ..., 2.1427e-03,\n",
      "          2.5556e-03, 1.4822e-03],\n",
      "         [1.0178e-02, 1.0635e-02, 1.7370e-02,  ..., 4.9179e-03,\n",
      "          8.5292e-03, 8.3256e-02],\n",
      "         ...,\n",
      "         [4.2746e-03, 4.0539e-03, 4.9432e-03,  ..., 3.8013e-03,\n",
      "          2.4913e-03, 3.8460e-03],\n",
      "         [1.2545e-01, 2.8241e-01, 4.4017e-02,  ..., 1.3198e-02,\n",
      "          1.4853e-02, 7.5744e-03],\n",
      "         [2.3306e-03, 1.3460e-03, 1.2530e-03,  ..., 1.9334e-03,\n",
      "          1.7035e-03, 1.3387e-03]],\n",
      "\n",
      "        [[3.4531e-03, 1.3579e-02, 1.8435e-02,  ..., 1.7488e-02,\n",
      "          1.5155e-02, 3.1185e-02],\n",
      "         [2.4402e-02, 4.1017e-01, 3.3600e-01,  ..., 1.3555e-02,\n",
      "          8.4229e-03, 5.2816e-03],\n",
      "         [1.4511e-02, 2.2724e-02, 4.5953e-02,  ..., 2.0182e-02,\n",
      "          4.2594e-02, 1.3919e-01],\n",
      "         ...,\n",
      "         [1.7384e-02, 1.8800e-02, 3.5235e-02,  ..., 1.7308e-01,\n",
      "          4.7977e-02, 3.8674e-02],\n",
      "         [3.3917e-02, 5.6035e-01, 6.8294e-02,  ..., 1.4040e-02,\n",
      "          1.3618e-02, 6.1602e-03],\n",
      "         [1.1399e-02, 3.0302e-02, 3.1831e-02,  ..., 1.5701e-02,\n",
      "          1.4164e-02, 8.5059e-03]],\n",
      "\n",
      "        [[2.1133e-04, 1.6567e-03, 1.4845e-03,  ..., 1.8283e-02,\n",
      "          5.2673e-02, 7.5381e-04],\n",
      "         [2.6317e-04, 9.2749e-01, 1.1157e-02,  ..., 4.9185e-03,\n",
      "          2.3264e-03, 1.1106e-03],\n",
      "         [3.7110e-03, 2.1540e-03, 2.0524e-03,  ..., 1.2794e-02,\n",
      "          4.6781e-02, 4.8506e-02],\n",
      "         ...,\n",
      "         [3.2397e-03, 1.0487e-03, 1.0785e-03,  ..., 1.3837e-03,\n",
      "          2.1254e-03, 7.8987e-04],\n",
      "         [1.2624e-02, 8.9391e-01, 3.7157e-02,  ..., 2.1393e-02,\n",
      "          8.6200e-04, 4.0946e-04],\n",
      "         [1.3661e-03, 7.7844e-04, 1.4083e-03,  ..., 6.1253e-02,\n",
      "          1.0504e-02, 1.9007e-03]],\n",
      "\n",
      "        [[9.0971e-03, 6.9135e-03, 9.3023e-03,  ..., 8.2400e-03,\n",
      "          1.5840e-02, 1.0077e-02],\n",
      "         [2.4323e-03, 3.9783e-02, 8.9570e-01,  ..., 2.1321e-03,\n",
      "          3.8855e-03, 2.7370e-03],\n",
      "         [3.2256e-02, 2.4843e-02, 2.5653e-02,  ..., 2.0978e-02,\n",
      "          7.1687e-02, 1.1807e-01],\n",
      "         ...,\n",
      "         [1.2222e-02, 6.4141e-03, 5.1741e-03,  ..., 9.0728e-03,\n",
      "          5.9941e-03, 3.3303e-03],\n",
      "         [3.9867e-02, 1.3828e-01, 5.9923e-01,  ..., 1.3676e-02,\n",
      "          1.3094e-02, 1.6890e-02],\n",
      "         [4.0860e-03, 1.3022e-03, 2.8554e-03,  ..., 3.5527e-03,\n",
      "          3.9563e-03, 4.0613e-03]],\n",
      "\n",
      "        [[6.1860e-03, 4.4764e-03, 1.1065e-02,  ..., 6.9330e-03,\n",
      "          8.4697e-03, 8.5640e-03],\n",
      "         [2.4759e-03, 1.1678e-02, 9.2810e-01,  ..., 4.7705e-03,\n",
      "          2.2906e-03, 2.2043e-03],\n",
      "         [1.0939e-02, 1.2168e-02, 3.9478e-02,  ..., 8.1954e-03,\n",
      "          2.4680e-02, 3.4929e-01],\n",
      "         ...,\n",
      "         [2.2624e-03, 3.5185e-03, 2.9562e-03,  ..., 2.6096e-03,\n",
      "          3.0521e-03, 3.2313e-03],\n",
      "         [1.2221e-02, 7.4902e-01, 1.4788e-01,  ..., 4.8917e-03,\n",
      "          5.9298e-03, 3.8767e-03],\n",
      "         [2.7937e-03, 2.0661e-03, 2.5083e-03,  ..., 4.2662e-03,\n",
      "          4.4849e-03, 2.3606e-03]]])\n",
      "tensor([14,  2, 14,  ...,  9,  1,  8])\n",
      "Submission file generated..!\n"
     ]
    }
   ],
   "source": [
    "LOAD_PATH = '/opt/ml/result/Best/Team7/'\n",
    "test_dir = '/opt/ml/input/data/eval'\n",
    "\n",
    "submission = pd.read_csv(os.path.join(test_dir, 'info.csv'))\n",
    "image_dir = os.path.join(test_dir, 'images')\n",
    "image_paths = [os.path.join(image_dir, img_id) for img_id in submission.ImageID]\n",
    "\n",
    "test_dataset = TestDataset(image_paths, transform = transformation4)\n",
    "loader = DataLoader(\n",
    "    test_dataset,\n",
    "    shuffle=False,\n",
    "    num_workers=num_workers\n",
    ")\n",
    "\n",
    "best_models = os.listdir(LOAD_PATH)\n",
    "total_outputs = torch.tensor([])\n",
    "for m in best_models:\n",
    "    if 'ipynb' in m:\n",
    "        continue\n",
    "    if 'three' in m:\n",
    "        continue\n",
    "        \n",
    "    elif 'efficient' in m:\n",
    "        model = KKT_Effi(num_classes=18).to(device)\n",
    "    elif 'mask' in m:\n",
    "        model = EfficientNet.from_pretrained('efficientnet-b0', num_classes=18).to(device)\n",
    "    elif 'Resnet' in m:\n",
    "        model = ResNet().to(device)\n",
    "    elif 'three' in m:\n",
    "        continue\n",
    "    else:\n",
    "        model = MyModel().to(device)\n",
    "    model.load_state_dict(torch.load(LOAD_PATH + m))\n",
    "\n",
    "    with torch.no_grad():\n",
    "        model.eval()\n",
    "        t_outputs = torch.tensor([])\n",
    "        for val_batch in tqdm(loader):            \n",
    "            outputs = F.softmax(model(val_batch.to(device, dtype=torch.float32)).cpu(), dim=1)\n",
    "            t_outputs = torch.cat((t_outputs, outputs), dim=0)   \n",
    "        total_outputs =  torch.cat((total_outputs, t_outputs.unsqueeze(0)), dim=0)\n",
    "        \n",
    "    print(total_outputs)    \n",
    "prediction_array = torch.argmax(total_outputs.mean(axis=0), dim=1)\n",
    "print(prediction_array)\n",
    "submission['ans'] = prediction_array\n",
    "submission.to_csv(os.path.join(test_dir, 'submission.csv'), index=False)\n",
    "print(\"Submission file generated..!\")"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "1bec7969-9c06-4a23-b341-27b7ea7ce677",
   "metadata": {},
   "source": [
    "### tta"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 57,
   "id": "375256f4-33b1-4137-a0a7-2043dade40fe",
   "metadata": {},
   "outputs": [
    {
     "name": "stderr",
     "output_type": "stream",
     "text": [
      "100%|██████████| 12600/12600 [37:04<00:00,  5.66it/s]\n"
     ]
    },
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "tensor([[[0.0026, 0.0131, 0.0087,  ..., 0.0121, 0.0166, 0.0138],\n",
      "         [0.0343, 0.3637, 0.2850,  ..., 0.0167, 0.0118, 0.0084],\n",
      "         [0.0174, 0.0282, 0.0400,  ..., 0.0204, 0.0686, 0.1531],\n",
      "         ...,\n",
      "         [0.0203, 0.0245, 0.0416,  ..., 0.1611, 0.0503, 0.0392],\n",
      "         [0.0224, 0.6892, 0.0561,  ..., 0.0097, 0.0094, 0.0034],\n",
      "         [0.0113, 0.0233, 0.0403,  ..., 0.0149, 0.0121, 0.0109]]]) tensor([4., 0., 1.,  ..., 4., 0., 2.])\n"
     ]
    },
    {
     "name": "stderr",
     "output_type": "stream",
     "text": [
      "100%|██████████| 12600/12600 [08:54<00:00, 23.58it/s]\n",
      "  0%|          | 0/12600 [00:00<?, ?it/s]"
     ]
    },
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "tensor([[[2.6327e-03, 1.3065e-02, 8.6945e-03,  ..., 1.2138e-02,\n",
      "          1.6590e-02, 1.3815e-02],\n",
      "         [3.4264e-02, 3.6370e-01, 2.8497e-01,  ..., 1.6733e-02,\n",
      "          1.1781e-02, 8.3753e-03],\n",
      "         [1.7439e-02, 2.8225e-02, 3.9994e-02,  ..., 2.0390e-02,\n",
      "          6.8561e-02, 1.5314e-01],\n",
      "         ...,\n",
      "         [2.0288e-02, 2.4516e-02, 4.1560e-02,  ..., 1.6108e-01,\n",
      "          5.0337e-02, 3.9171e-02],\n",
      "         [2.2355e-02, 6.8916e-01, 5.6104e-02,  ..., 9.7398e-03,\n",
      "          9.4342e-03, 3.4021e-03],\n",
      "         [1.1315e-02, 2.3257e-02, 4.0328e-02,  ..., 1.4941e-02,\n",
      "          1.2130e-02, 1.0936e-02]],\n",
      "\n",
      "        [[2.4611e-04, 2.2259e-03, 4.0374e-03,  ..., 1.1243e-02,\n",
      "          3.1737e-02, 4.0405e-04],\n",
      "         [2.8814e-04, 8.6655e-01, 1.5966e-02,  ..., 3.0559e-03,\n",
      "          1.5199e-03, 8.6711e-04],\n",
      "         [2.0236e-03, 4.4097e-03, 7.0924e-03,  ..., 9.5996e-03,\n",
      "          8.3819e-02, 4.1251e-02],\n",
      "         ...,\n",
      "         [4.8723e-03, 3.0930e-03, 1.9718e-03,  ..., 1.6114e-03,\n",
      "          2.5227e-03, 8.1363e-04],\n",
      "         [1.1767e-02, 9.2533e-01, 1.6673e-02,  ..., 1.8170e-02,\n",
      "          8.8647e-04, 4.5337e-04],\n",
      "         [2.0743e-03, 1.0807e-03, 2.0288e-03,  ..., 4.7625e-02,\n",
      "          1.0029e-02, 2.0529e-03]]]) tensor([4., 0., 1.,  ..., 4., 0., 2.])\n",
      "Loaded pretrained weights for efficientnet-b0\n"
     ]
    },
    {
     "name": "stderr",
     "output_type": "stream",
     "text": [
      "100%|██████████| 12600/12600 [42:59<00:00,  4.88it/s] \n"
     ]
    },
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "tensor([[[2.6327e-03, 1.3065e-02, 8.6945e-03,  ..., 1.2138e-02,\n",
      "          1.6590e-02, 1.3815e-02],\n",
      "         [3.4264e-02, 3.6370e-01, 2.8497e-01,  ..., 1.6733e-02,\n",
      "          1.1781e-02, 8.3753e-03],\n",
      "         [1.7439e-02, 2.8225e-02, 3.9994e-02,  ..., 2.0390e-02,\n",
      "          6.8561e-02, 1.5314e-01],\n",
      "         ...,\n",
      "         [2.0288e-02, 2.4516e-02, 4.1560e-02,  ..., 1.6108e-01,\n",
      "          5.0337e-02, 3.9171e-02],\n",
      "         [2.2355e-02, 6.8916e-01, 5.6104e-02,  ..., 9.7398e-03,\n",
      "          9.4342e-03, 3.4021e-03],\n",
      "         [1.1315e-02, 2.3257e-02, 4.0328e-02,  ..., 1.4941e-02,\n",
      "          1.2130e-02, 1.0936e-02]],\n",
      "\n",
      "        [[2.4611e-04, 2.2259e-03, 4.0374e-03,  ..., 1.1243e-02,\n",
      "          3.1737e-02, 4.0405e-04],\n",
      "         [2.8814e-04, 8.6655e-01, 1.5966e-02,  ..., 3.0559e-03,\n",
      "          1.5199e-03, 8.6711e-04],\n",
      "         [2.0236e-03, 4.4097e-03, 7.0924e-03,  ..., 9.5996e-03,\n",
      "          8.3819e-02, 4.1251e-02],\n",
      "         ...,\n",
      "         [4.8723e-03, 3.0930e-03, 1.9718e-03,  ..., 1.6114e-03,\n",
      "          2.5227e-03, 8.1363e-04],\n",
      "         [1.1767e-02, 9.2533e-01, 1.6673e-02,  ..., 1.8170e-02,\n",
      "          8.8647e-04, 4.5337e-04],\n",
      "         [2.0743e-03, 1.0807e-03, 2.0288e-03,  ..., 4.7625e-02,\n",
      "          1.0029e-02, 2.0529e-03]],\n",
      "\n",
      "        [[1.0419e-02, 7.6227e-03, 1.1257e-02,  ..., 9.0554e-03,\n",
      "          1.4149e-02, 8.5933e-03],\n",
      "         [2.6537e-03, 2.4066e-02, 9.0642e-01,  ..., 2.3535e-03,\n",
      "          4.6376e-03, 3.4770e-03],\n",
      "         [2.7631e-02, 1.9826e-02, 2.2040e-02,  ..., 1.7891e-02,\n",
      "          4.5152e-02, 8.5196e-02],\n",
      "         ...,\n",
      "         [9.3616e-03, 5.8718e-03, 5.0998e-03,  ..., 9.2792e-03,\n",
      "          6.9238e-03, 3.1931e-03],\n",
      "         [4.4618e-02, 3.4825e-01, 3.1490e-01,  ..., 1.7446e-02,\n",
      "          1.5244e-02, 1.6570e-02],\n",
      "         [3.8381e-03, 1.3768e-03, 3.2446e-03,  ..., 3.5282e-03,\n",
      "          3.6323e-03, 3.6008e-03]]]) tensor([4., 0., 1.,  ..., 4., 0., 2.])\n"
     ]
    },
    {
     "name": "stderr",
     "output_type": "stream",
     "text": [
      "100%|██████████| 12600/12600 [08:48<00:00, 23.85it/s]"
     ]
    },
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "tensor([[[2.6327e-03, 1.3065e-02, 8.6945e-03,  ..., 1.2138e-02,\n",
      "          1.6590e-02, 1.3815e-02],\n",
      "         [3.4264e-02, 3.6370e-01, 2.8497e-01,  ..., 1.6733e-02,\n",
      "          1.1781e-02, 8.3753e-03],\n",
      "         [1.7439e-02, 2.8225e-02, 3.9994e-02,  ..., 2.0390e-02,\n",
      "          6.8561e-02, 1.5314e-01],\n",
      "         ...,\n",
      "         [2.0288e-02, 2.4516e-02, 4.1560e-02,  ..., 1.6108e-01,\n",
      "          5.0337e-02, 3.9171e-02],\n",
      "         [2.2355e-02, 6.8916e-01, 5.6104e-02,  ..., 9.7398e-03,\n",
      "          9.4342e-03, 3.4021e-03],\n",
      "         [1.1315e-02, 2.3257e-02, 4.0328e-02,  ..., 1.4941e-02,\n",
      "          1.2130e-02, 1.0936e-02]],\n",
      "\n",
      "        [[2.4611e-04, 2.2259e-03, 4.0374e-03,  ..., 1.1243e-02,\n",
      "          3.1737e-02, 4.0405e-04],\n",
      "         [2.8814e-04, 8.6655e-01, 1.5966e-02,  ..., 3.0559e-03,\n",
      "          1.5199e-03, 8.6711e-04],\n",
      "         [2.0236e-03, 4.4097e-03, 7.0924e-03,  ..., 9.5996e-03,\n",
      "          8.3819e-02, 4.1251e-02],\n",
      "         ...,\n",
      "         [4.8723e-03, 3.0930e-03, 1.9718e-03,  ..., 1.6114e-03,\n",
      "          2.5227e-03, 8.1363e-04],\n",
      "         [1.1767e-02, 9.2533e-01, 1.6673e-02,  ..., 1.8170e-02,\n",
      "          8.8647e-04, 4.5337e-04],\n",
      "         [2.0743e-03, 1.0807e-03, 2.0288e-03,  ..., 4.7625e-02,\n",
      "          1.0029e-02, 2.0529e-03]],\n",
      "\n",
      "        [[1.0419e-02, 7.6227e-03, 1.1257e-02,  ..., 9.0554e-03,\n",
      "          1.4149e-02, 8.5933e-03],\n",
      "         [2.6537e-03, 2.4066e-02, 9.0642e-01,  ..., 2.3535e-03,\n",
      "          4.6376e-03, 3.4770e-03],\n",
      "         [2.7631e-02, 1.9826e-02, 2.2040e-02,  ..., 1.7891e-02,\n",
      "          4.5152e-02, 8.5196e-02],\n",
      "         ...,\n",
      "         [9.3616e-03, 5.8718e-03, 5.0998e-03,  ..., 9.2792e-03,\n",
      "          6.9238e-03, 3.1931e-03],\n",
      "         [4.4618e-02, 3.4825e-01, 3.1490e-01,  ..., 1.7446e-02,\n",
      "          1.5244e-02, 1.6570e-02],\n",
      "         [3.8381e-03, 1.3768e-03, 3.2446e-03,  ..., 3.5282e-03,\n",
      "          3.6323e-03, 3.6008e-03]],\n",
      "\n",
      "        [[6.7723e-03, 5.3564e-03, 8.1634e-03,  ..., 7.6208e-03,\n",
      "          9.0789e-03, 1.0624e-02],\n",
      "         [2.7176e-03, 2.9816e-02, 8.9494e-01,  ..., 5.0245e-03,\n",
      "          3.1152e-03, 2.5834e-03],\n",
      "         [1.1059e-02, 1.2395e-02, 3.6879e-02,  ..., 8.6800e-03,\n",
      "          2.9366e-02, 1.4202e-01],\n",
      "         ...,\n",
      "         [2.9307e-03, 4.1914e-03, 3.6842e-03,  ..., 4.2771e-03,\n",
      "          3.9506e-03, 3.5709e-03],\n",
      "         [2.6725e-02, 6.8980e-01, 1.2556e-01,  ..., 7.7784e-03,\n",
      "          8.9901e-03, 4.1905e-03],\n",
      "         [3.0273e-03, 2.1232e-03, 2.8807e-03,  ..., 4.7880e-03,\n",
      "          4.2869e-03, 2.3787e-03]]]) tensor([4., 0., 1.,  ..., 4., 0., 2.])\n",
      "tensor([14,  2, 14,  ...,  9,  1,  8])\n",
      "Submission file generated..!\n"
     ]
    },
    {
     "name": "stderr",
     "output_type": "stream",
     "text": [
      "\n"
     ]
    }
   ],
   "source": [
    "import ttach as tta\n",
    "\n",
    "tf = tta.Compose(\n",
    "[\n",
    "    tta.HorizontalFlip(),\n",
    "    tta.Rotate90(angles=[0,15]),\n",
    "    tta.Multiply(factors=[0.8,0.9,1,1.1]),\n",
    "])\n",
    "\n",
    "LOAD_PATH = '/opt/ml/result/Best/Team7/'\n",
    "test_dir = '/opt/ml/input/data/eval'\n",
    "\n",
    "submission = pd.read_csv(os.path.join(test_dir, 'info.csv'))\n",
    "image_dir = os.path.join(test_dir, 'images')\n",
    "image_paths = [os.path.join(image_dir, img_id) for img_id in submission.ImageID]\n",
    "\n",
    "test_dataset = TestDataset(image_paths, transform = transformation4)\n",
    "loader = DataLoader(\n",
    "    test_dataset,\n",
    "    shuffle=False,\n",
    "    num_workers=num_workers\n",
    ")\n",
    "\n",
    "best_models = os.listdir(LOAD_PATH)\n",
    "total_outputs = torch.tensor([])\n",
    "for m in best_models:\n",
    "    if 'ipynb' in m:\n",
    "        continue\n",
    "    if 'three' in m:\n",
    "        continue\n",
    "        \n",
    "    elif 'efficient' in m:\n",
    "        model = KKT_Effi(num_classes=18).to(device)\n",
    "    elif 'mask' in m:\n",
    "        model = EfficientNet.from_pretrained('efficientnet-b0', num_classes=18).to(device)\n",
    "    elif 'Resnet' in m:\n",
    "        model = ResNet().to(device)\n",
    "    elif 'three' in m:\n",
    "        continue\n",
    "    else:\n",
    "        model = MyModel().to(device)\n",
    "    model.load_state_dict(torch.load(LOAD_PATH + m))\n",
    "    tta_model = tta.ClassificationTTAWrapper(model,tf)\n",
    "\n",
    "    with torch.no_grad():\n",
    "        model.eval()\n",
    "        t_outputs = torch.tensor([])\n",
    "        for val_batch in tqdm(loader):\n",
    "            outputs = F.softmax(tta_model(val_batch.to(device, dtype=torch.float32)).cpu(), dim=1)\n",
    "            t_outputs = torch.cat((t_outputs, outputs), dim=0)   \n",
    "        total_outputs =  torch.cat((total_outputs, t_outputs.unsqueeze(0)), dim=0)\n",
    "        \n",
    "    print(total_outputs, total_labels)    \n",
    "prediction_array = torch.argmax(total_outputs.mean(axis=0), dim=1)\n",
    "print(prediction_array)\n",
    "submission['ans'] = prediction_array\n",
    "submission.to_csv(os.path.join(test_dir, 'submission.csv'), index=False)\n",
    "print(\"Submission file generated..!\")"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "37ebb54a-f1c3-47db-ab57-d8b2f087b399",
   "metadata": {},
   "source": [
    "### submission 개형"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 50,
   "id": "7f76ae3b-94f6-43cc-aeb8-827b5820ebed",
   "metadata": {},
   "outputs": [
    {
     "data": {
      "text/plain": [
       "<matplotlib.axes._subplots.AxesSubplot at 0x7fbe258aae80>"
      ]
     },
     "execution_count": 50,
     "metadata": {},
     "output_type": "execute_result"
    },
    {
     "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": [
    "submission = pd.read_csv(os.path.join(test_dir, \"submission_team7.csv\"))\n",
    "sns.countplot(\n",
    "    data = submission,\n",
    "    x = \"ans\",\n",
    ")"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "2880e17c-5db8-4491-b91b-dafcb26dcd25",
   "metadata": {},
   "source": [
    "### 지금까지의 best submission과 비교"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "id": "0dc816bc-ff6e-45a3-892f-941120a4b91b",
   "metadata": {
    "pycharm": {
     "name": "#%%\n"
    }
   },
   "outputs": [],
   "source": [
    "# submission: KKT+KKH, TTA\n",
    "\n",
    "test_dir = '/opt/ml/input/data/eval'\n",
    "\n",
    "submission_best = pd.read_csv(os.path.join(test_dir, \"submission_team7.csv\"))\n",
    "submission_compare = pd.read_csv(os.path.join(test_dir, \"submission_team8.csv\"))\n",
    "\n",
    "submission_best_ans = submission_best['ans'].tolist()\n",
    "submission_compare_ans = submission_compare['ans'].tolist()\n",
    "\n",
    "f1 = sklearn.metrics.f1_score(submission_compare_ans, submission_best_ans, average=\"macro\")\n",
    "print('F1: {}'.format(f1))\n",
    "\n",
    "compare_matrix = confusion_matrix(submission_best_ans, submission_compare_ans)\n",
    "plot_confusion_matrix(compare_matrix)\n",
    "#plt.ylabel('best submission')\n",
    "#plt.xlabel('compare submission\\naccuracy={:0.4f}; misclass={:0.4f}'.format(accuracy, misclass))"
   ]
  }
 ],
 "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": "3.8.5"
  }
 },
 "nbformat": 4,
 "nbformat_minor": 5
}