├── Accident Detection from CCTV
    └── spinal-tl-for-accident-detection-from-cctv-images.ipynb
├── Hyperspectral Image Classification
    ├── HSI.ipynb
    ├── HSI_Confusion Matrix.png
    ├── HSI_KSC.py
    ├── HSI_cmap.png
    └── HSI_gt.png
└── README.md


/Accident Detection from CCTV/spinal-tl-for-accident-detection-from-cctv-images.ipynb:
--------------------------------------------------------------------------------
   1 | {
   2 |  "cells": [
   3 |   {
   4 |    "cell_type": "markdown",
   5 |    "id": "municipal-constitutional",
   6 |    "metadata": {
   7 |     "papermill": {
   8 |      "duration": 0.010268,
   9 |      "end_time": "2021-06-17T06:12:08.694665",
  10 |      "exception": false,
  11 |      "start_time": "2021-06-17T06:12:08.684397",
  12 |      "status": "completed"
  13 |     },
  14 |     "tags": []
  15 |    },
  16 |    "source": [
  17 |     "# Transfer learning with SpinalNet fully connected layer for accident detection from CCTV footage.\n",
  18 |     "\n",
  19 |     "### We write this code with the help of PyTorch demo:\n",
  20 |     "###    https://pytorch.org/tutorials/beginner/transfer_learning_tutorial.html\n",
  21 |     "\n",
  22 |     "## More scripts are available in earlier versions of the notebook.\n",
  23 |     "\n",
  24 |     "### For more code of SpinalNet, please visit: https://github.com/dipuk0506/SpinalNet"
  25 |    ]
  26 |   },
  27 |   {
  28 |    "cell_type": "markdown",
  29 |    "id": "architectural-browse",
  30 |    "metadata": {
  31 |     "papermill": {
  32 |      "duration": 0.008843,
  33 |      "end_time": "2021-06-17T06:12:08.712718",
  34 |      "exception": false,
  35 |      "start_time": "2021-06-17T06:12:08.703875",
  36 |      "status": "completed"
  37 |     },
  38 |     "tags": []
  39 |    },
  40 |    "source": [
  41 |     "# Dataloader and Visualization"
  42 |    ]
  43 |   },
  44 |   {
  45 |    "cell_type": "code",
  46 |    "execution_count": 1,
  47 |    "id": "thousand-drawing",
  48 |    "metadata": {
  49 |     "execution": {
  50 |      "iopub.execute_input": "2021-06-17T06:12:08.747498Z",
  51 |      "iopub.status.busy": "2021-06-17T06:12:08.744374Z",
  52 |      "iopub.status.idle": "2021-06-17T06:12:12.004408Z",
  53 |      "shell.execute_reply": "2021-06-17T06:12:12.003464Z",
  54 |      "shell.execute_reply.started": "2021-06-17T06:04:24.072558Z"
  55 |     },
  56 |     "papermill": {
  57 |      "duration": 3.282871,
  58 |      "end_time": "2021-06-17T06:12:12.004559",
  59 |      "exception": false,
  60 |      "start_time": "2021-06-17T06:12:08.721688",
  61 |      "status": "completed"
  62 |     },
  63 |     "tags": []
  64 |    },
  65 |    "outputs": [
  66 |     {
  67 |      "data": {
  68 |       "image/png": "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\n",
  69 |       "text/plain": [
  70 |        "<Figure size 432x288 with 1 Axes>"
  71 |       ]
  72 |      },
  73 |      "metadata": {
  74 |       "needs_background": "light"
  75 |      },
  76 |      "output_type": "display_data"
  77 |     }
  78 |    ],
  79 |    "source": [
  80 |     "\n",
  81 |     "from __future__ import print_function, division\n",
  82 |     "\n",
  83 |     "import torch\n",
  84 |     "import torch.nn as nn\n",
  85 |     "import torch.optim as optim\n",
  86 |     "from torch.optim import lr_scheduler\n",
  87 |     "import numpy as np\n",
  88 |     "import torchvision\n",
  89 |     "from torchvision import datasets, models, transforms\n",
  90 |     "import matplotlib.pyplot as plt\n",
  91 |     "import time\n",
  92 |     "import os\n",
  93 |     "import copy\n",
  94 |     "\n",
  95 |     "\n",
  96 |     "\n",
  97 |     "\n",
  98 |     "plt.ion()   # interactive mode\n",
  99 |     "\n",
 100 |     "# Data augmentation and normalization for training\n",
 101 |     "# Just normalization for validation\n",
 102 |     "data_transforms = {\n",
 103 |     "    'train': transforms.Compose([\n",
 104 |     "        transforms.Resize((512,512)),\n",
 105 |     "        transforms.RandomRotation(15,),\n",
 106 |     "        transforms.RandomHorizontalFlip(),\n",
 107 |     "        transforms.ToTensor(),\n",
 108 |     "        transforms.Normalize(mean=[0.507, 0.487, 0.441], std=[0.267, 0.256, 0.276])\n",
 109 |     "    ]),\n",
 110 |     "    'val': transforms.Compose([\n",
 111 |     "        transforms.Resize((512,512)),\n",
 112 |     "        transforms.ToTensor(),\n",
 113 |     "        transforms.Normalize(mean=[0.507, 0.487, 0.441], std=[0.267, 0.256, 0.276])\n",
 114 |     "    ]),\n",
 115 |     "    'test': transforms.Compose([\n",
 116 |     "        transforms.Resize((512,512)),\n",
 117 |     "        transforms.ToTensor(),\n",
 118 |     "        transforms.Normalize(mean=[0.507, 0.487, 0.441], std=[0.267, 0.256, 0.276])\n",
 119 |     "    ]),\n",
 120 |     "}\n",
 121 |     "\n",
 122 |     "data_dir = '../input/accident-detection-from-cctv-footage/data'\n",
 123 |     "image_datasets = {x: datasets.ImageFolder(os.path.join(data_dir, x),\n",
 124 |     "                                          data_transforms[x])\n",
 125 |     "                  for x in ['train', 'val', 'test']}\n",
 126 |     "dataloaders = {x: torch.utils.data.DataLoader(image_datasets[x], batch_size=8,\n",
 127 |     "                                             shuffle=True, num_workers=0)\n",
 128 |     "              for x in ['train', 'val', 'test']}\n",
 129 |     "dataset_sizes = {x: len(image_datasets[x]) for x in ['train', 'val', 'test']}\n",
 130 |     "class_names = image_datasets['train'].classes\n",
 131 |     "\n",
 132 |     "device = torch.device(\"cuda:0\" if torch.cuda.is_available() else \"cpu\")\n",
 133 |     "\n",
 134 |     "def imshow(inp, title=None):\n",
 135 |     "    \"\"\"Imshow for Tensor.\"\"\"\n",
 136 |     "    inp = inp.numpy().transpose((1, 2, 0))\n",
 137 |     "    mean = np.array([0.485, 0.456, 0.406])\n",
 138 |     "    std = np.array([0.229, 0.224, 0.225])\n",
 139 |     "    inp = std * inp + mean\n",
 140 |     "    inp = np.clip(inp, 0, 1)\n",
 141 |     "    plt.imshow(inp)\n",
 142 |     "    if title is not None:\n",
 143 |     "        plt.title(title)\n",
 144 |     "    plt.pause(0.001)  # pause a bit so that plots are updated\n",
 145 |     "\n",
 146 |     "\n",
 147 |     "# Get a batch of training data\n",
 148 |     "inputs, classes = next(iter(dataloaders['train']))\n",
 149 |     "\n",
 150 |     "# Make a grid from batch\n",
 151 |     "out = torchvision.utils.make_grid(inputs)\n",
 152 |     "\n",
 153 |     "imshow(out)#, title=[class_names[x] for x in classes])"
 154 |    ]
 155 |   },
 156 |   {
 157 |    "cell_type": "markdown",
 158 |    "id": "acting-immune",
 159 |    "metadata": {
 160 |     "papermill": {
 161 |      "duration": 0.010411,
 162 |      "end_time": "2021-06-17T06:12:12.026314",
 163 |      "exception": false,
 164 |      "start_time": "2021-06-17T06:12:12.015903",
 165 |      "status": "completed"
 166 |     },
 167 |     "tags": []
 168 |    },
 169 |    "source": [
 170 |     "# Downloading Model and Changing Fully-connected/ Classifier Layer"
 171 |    ]
 172 |   },
 173 |   {
 174 |    "cell_type": "code",
 175 |    "execution_count": 2,
 176 |    "id": "driving-newfoundland",
 177 |    "metadata": {
 178 |     "execution": {
 179 |      "iopub.execute_input": "2021-06-17T06:12:12.072560Z",
 180 |      "iopub.status.busy": "2021-06-17T06:12:12.071720Z",
 181 |      "iopub.status.idle": "2021-06-17T06:12:38.510818Z",
 182 |      "shell.execute_reply": "2021-06-17T06:12:38.510343Z",
 183 |      "shell.execute_reply.started": "2021-06-17T06:04:26.954454Z"
 184 |     },
 185 |     "papermill": {
 186 |      "duration": 26.474321,
 187 |      "end_time": "2021-06-17T06:12:38.510950",
 188 |      "exception": false,
 189 |      "start_time": "2021-06-17T06:12:12.036629",
 190 |      "status": "completed"
 191 |     },
 192 |     "tags": []
 193 |    },
 194 |    "outputs": [
 195 |     {
 196 |      "name": "stderr",
 197 |      "output_type": "stream",
 198 |      "text": [
 199 |       "Downloading: \"https://download.pytorch.org/models/vgg19_bn-c79401a0.pth\" to /root/.cache/torch/hub/checkpoints/vgg19_bn-c79401a0.pth\n"
 200 |      ]
 201 |     },
 202 |     {
 203 |      "data": {
 204 |       "application/vnd.jupyter.widget-view+json": {
 205 |        "model_id": "1534c8af727a4f6bbfee39b99b598433",
 206 |        "version_major": 2,
 207 |        "version_minor": 0
 208 |       },
 209 |       "text/plain": [
 210 |        "  0%|          | 0.00/548M [00:00<?, ?B/s]"
 211 |       ]
 212 |      },
 213 |      "metadata": {},
 214 |      "output_type": "display_data"
 215 |     }
 216 |    ],
 217 |    "source": [
 218 |     "#%%\n",
 219 |     "\n",
 220 |     "model_ft = models.vgg19_bn(pretrained=True)\n",
 221 |     "num_ftrs = model_ft.classifier[0].in_features\n",
 222 |     "\n",
 223 |     "#model_ft = models.resnet18(pretrained=True)\n",
 224 |     "#num_ftrs = model_ft.fc.in_features\n",
 225 |     "\n",
 226 |     "half_in_size = round(num_ftrs/2)\n",
 227 |     "layer_width = 1024\n",
 228 |     "Num_class=2\n",
 229 |     "\n",
 230 |     "class SpinalNet_ResNet(nn.Module):\n",
 231 |     "    def __init__(self):\n",
 232 |     "        super(SpinalNet_ResNet, self).__init__()\n",
 233 |     "        \n",
 234 |     "        self.fc_spinal_layer1 = nn.Sequential(\n",
 235 |     "            #nn.Dropout(p = 0.5), \n",
 236 |     "            nn.Linear(half_in_size, layer_width),\n",
 237 |     "            #nn.BatchNorm1d(layer_width), \n",
 238 |     "            nn.ReLU(inplace=True),)\n",
 239 |     "        self.fc_spinal_layer2 = nn.Sequential(\n",
 240 |     "            #nn.Dropout(p = 0.5), \n",
 241 |     "            nn.Linear(half_in_size+layer_width, layer_width),\n",
 242 |     "            #nn.BatchNorm1d(layer_width), \n",
 243 |     "            nn.ReLU(inplace=True),)\n",
 244 |     "        self.fc_spinal_layer3 = nn.Sequential(\n",
 245 |     "            #nn.Dropout(p = 0.5), \n",
 246 |     "            nn.Linear(half_in_size+layer_width, layer_width),\n",
 247 |     "            #nn.BatchNorm1d(layer_width), \n",
 248 |     "            nn.ReLU(inplace=True),)\n",
 249 |     "        self.fc_spinal_layer4 = nn.Sequential(\n",
 250 |     "            #nn.Dropout(p = 0.5), \n",
 251 |     "            nn.Linear(half_in_size+layer_width, layer_width),\n",
 252 |     "            #nn.BatchNorm1d(layer_width), \n",
 253 |     "            nn.ReLU(inplace=True),)\n",
 254 |     "        self.fc_out = nn.Sequential(\n",
 255 |     "            #nn.Dropout(p = 0.5), \n",
 256 |     "            nn.Linear(layer_width*4, Num_class),)\n",
 257 |     "        \n",
 258 |     "    def forward(self, x):\n",
 259 |     "        x1 = self.fc_spinal_layer1(x[:, 0:half_in_size])\n",
 260 |     "        x2 = self.fc_spinal_layer2(torch.cat([ x[:,half_in_size:2*half_in_size], x1], dim=1))\n",
 261 |     "        x3 = self.fc_spinal_layer3(torch.cat([ x[:,0:half_in_size], x2], dim=1))\n",
 262 |     "        x4 = self.fc_spinal_layer4(torch.cat([ x[:,half_in_size:2*half_in_size], x3], dim=1))\n",
 263 |     "        \n",
 264 |     "        \n",
 265 |     "        x = torch.cat([x1, x2], dim=1)\n",
 266 |     "        x = torch.cat([x, x3], dim=1)\n",
 267 |     "        x = torch.cat([x, x4], dim=1)\n",
 268 |     "\n",
 269 |     "        \n",
 270 |     "        x = self.fc_out(x)\n",
 271 |     "        return x\n",
 272 |     "    \n",
 273 |     "class SpinalNet_VGG(nn.Module):\n",
 274 |     "    def __init__(self):\n",
 275 |     "        super(SpinalNet_VGG, self).__init__()\n",
 276 |     "        \n",
 277 |     "        self.fc_spinal_layer1 = nn.Sequential(\n",
 278 |     "            nn.Dropout(p = 0.5), nn.Linear(half_in_size, layer_width),\n",
 279 |     "            nn.BatchNorm1d(layer_width), nn.ReLU(inplace=True),)\n",
 280 |     "        self.fc_spinal_layer2 = nn.Sequential(\n",
 281 |     "            nn.Dropout(p = 0.5), \n",
 282 |     "            nn.Linear(half_in_size+layer_width, layer_width),\n",
 283 |     "            nn.BatchNorm1d(layer_width), \n",
 284 |     "            nn.ReLU(inplace=True),)\n",
 285 |     "        self.fc_spinal_layer3 = nn.Sequential(\n",
 286 |     "            nn.Dropout(p = 0.5), \n",
 287 |     "            nn.Linear(half_in_size+layer_width, layer_width),\n",
 288 |     "            nn.BatchNorm1d(layer_width), \n",
 289 |     "            nn.ReLU(inplace=True),)\n",
 290 |     "        self.fc_spinal_layer4 = nn.Sequential(\n",
 291 |     "            nn.Dropout(p = 0.5), \n",
 292 |     "            nn.Linear(half_in_size+layer_width, layer_width),\n",
 293 |     "            nn.BatchNorm1d(layer_width), \n",
 294 |     "            nn.ReLU(inplace=True),)\n",
 295 |     "        self.fc_out = nn.Sequential(\n",
 296 |     "            nn.Dropout(p = 0.5), \n",
 297 |     "            nn.Linear(layer_width*4, Num_class),)        \n",
 298 |     "\n",
 299 |     "    def forward(self, x):\n",
 300 |     "        x1 = self.fc_spinal_layer1(x[:, 0:half_in_size])\n",
 301 |     "        x2 = self.fc_spinal_layer2(torch.cat([ x[:,half_in_size:2*half_in_size], x1], dim=1))\n",
 302 |     "        x3 = self.fc_spinal_layer3(torch.cat([ x[:,0:half_in_size], x2], dim=1))\n",
 303 |     "        x4 = self.fc_spinal_layer4(torch.cat([ x[:,half_in_size:2*half_in_size], x3], dim=1))\n",
 304 |     "        \n",
 305 |     "        \n",
 306 |     "        x = torch.cat([x1, x2], dim=1)\n",
 307 |     "        x = torch.cat([x, x3], dim=1)\n",
 308 |     "        x = torch.cat([x, x4], dim=1)\n",
 309 |     "\n",
 310 |     "        \n",
 311 |     "        x = self.fc_out(x)\n",
 312 |     "        return x\n",
 313 |     "\n",
 314 |     "\n",
 315 |     "\n",
 316 |     "net_classifier = nn.Sequential(\n",
 317 |     "            nn.Linear(num_ftrs, 4096),\n",
 318 |     "            nn.ReLU(inplace=True),\n",
 319 |     "            nn.Dropout(),\n",
 320 |     "            nn.Linear(4096, 4096),\n",
 321 |     "            nn.ReLU(inplace=True),\n",
 322 |     "            nn.Dropout(),\n",
 323 |     "            nn.Linear(4096, Num_class)\n",
 324 |     "        )\n",
 325 |     "\n",
 326 |     "\n",
 327 |     "'''\n",
 328 |     "Changing the fully connected layer to SpinalNet or ResNet or VGG\n",
 329 |     "'''\n",
 330 |     "\n",
 331 |     "#model_ft.fc = nn.Linear(num_ftrs, Num_class)\n",
 332 |     "#model_ft.fc = SpinalNet_ResNet()\n",
 333 |     "\n",
 334 |     "model_ft.classifier = SpinalNet_ResNet()\n",
 335 |     "#model_ft.classifier = net_classifier"
 336 |    ]
 337 |   },
 338 |   {
 339 |    "cell_type": "markdown",
 340 |    "id": "immune-spell",
 341 |    "metadata": {
 342 |     "papermill": {
 343 |      "duration": 0.010862,
 344 |      "end_time": "2021-06-17T06:12:38.533127",
 345 |      "exception": false,
 346 |      "start_time": "2021-06-17T06:12:38.522265",
 347 |      "status": "completed"
 348 |     },
 349 |     "tags": []
 350 |    },
 351 |    "source": [
 352 |     "# Training Function"
 353 |    ]
 354 |   },
 355 |   {
 356 |    "cell_type": "code",
 357 |    "execution_count": 3,
 358 |    "id": "enabling-glenn",
 359 |    "metadata": {
 360 |     "execution": {
 361 |      "iopub.execute_input": "2021-06-17T06:12:38.568282Z",
 362 |      "iopub.status.busy": "2021-06-17T06:12:38.567547Z",
 363 |      "iopub.status.idle": "2021-06-17T06:12:38.570328Z",
 364 |      "shell.execute_reply": "2021-06-17T06:12:38.569906Z",
 365 |      "shell.execute_reply.started": "2021-06-17T06:04:54.024367Z"
 366 |     },
 367 |     "papermill": {
 368 |      "duration": 0.024852,
 369 |      "end_time": "2021-06-17T06:12:38.570430",
 370 |      "exception": false,
 371 |      "start_time": "2021-06-17T06:12:38.545578",
 372 |      "status": "completed"
 373 |     },
 374 |     "tags": []
 375 |    },
 376 |    "outputs": [],
 377 |    "source": [
 378 |     "def train_model(model, criterion, optimizer, scheduler, num_epochs=25):\n",
 379 |     "    since = time.time()\n",
 380 |     "\n",
 381 |     "    best_model_wts = copy.deepcopy(model.state_dict())\n",
 382 |     "    best_acc = 0.0\n",
 383 |     "    test_token=0\n",
 384 |     "\n",
 385 |     "    for epoch in range(num_epochs):\n",
 386 |     "        print('Epoch {}/{}'.format(epoch, num_epochs - 1))\n",
 387 |     "        print('-' * 10)\n",
 388 |     "\n",
 389 |     "        # Each epoch has a training and validation phase\n",
 390 |     "        for phase in ['train', 'val', 'test']:\n",
 391 |     "             \n",
 392 |     "            \n",
 393 |     "            '''\n",
 394 |     "            Test when a better validation result is found\n",
 395 |     "            '''\n",
 396 |     "            if test_token ==0 and phase == 'test':\n",
 397 |     "                continue\n",
 398 |     "            test_token =0\n",
 399 |     "            \n",
 400 |     "            \n",
 401 |     "            if phase == 'train':\n",
 402 |     "                model.train()  # Set model to training mode\n",
 403 |     "            else:\n",
 404 |     "                model.eval()   # Set model to evaluate mode\n",
 405 |     "\n",
 406 |     "            running_loss = 0.0\n",
 407 |     "            running_corrects = 0\n",
 408 |     "\n",
 409 |     "            # Iterate over data.\n",
 410 |     "            for inputs, labels in dataloaders[phase]:\n",
 411 |     "                inputs = inputs.to(device)\n",
 412 |     "                labels = labels.to(device)\n",
 413 |     "\n",
 414 |     "                # zero the parameter gradients\n",
 415 |     "                optimizer.zero_grad()\n",
 416 |     "\n",
 417 |     "                # forward\n",
 418 |     "                # track history if only in train\n",
 419 |     "                with torch.set_grad_enabled(phase == 'train'):\n",
 420 |     "                    outputs = model(inputs)\n",
 421 |     "                    _, preds = torch.max(outputs, 1)\n",
 422 |     "                    loss = criterion(outputs, labels)\n",
 423 |     "\n",
 424 |     "                    # backward + optimize only if in training phase\n",
 425 |     "                    if phase == 'train':\n",
 426 |     "                        loss.backward()\n",
 427 |     "                        optimizer.step()\n",
 428 |     "\n",
 429 |     "                # statistics\n",
 430 |     "                running_loss += loss.item() * inputs.size(0)\n",
 431 |     "                running_corrects += torch.sum(preds == labels.data)\n",
 432 |     "            if phase == 'train':\n",
 433 |     "                scheduler.step()\n",
 434 |     "\n",
 435 |     "            epoch_loss = running_loss / dataset_sizes[phase]\n",
 436 |     "            epoch_acc = running_corrects.double() / dataset_sizes[phase]\n",
 437 |     "\n",
 438 |     "            print('{} Loss: {:.4f} Acc: {:.4f}'.format(\n",
 439 |     "                phase, epoch_loss, epoch_acc))\n",
 440 |     "\n",
 441 |     "            # deep copy the model\n",
 442 |     "            if phase == 'val' and epoch_acc > best_acc:\n",
 443 |     "                best_acc = epoch_acc\n",
 444 |     "                best_model_wts = copy.deepcopy(model.state_dict())\n",
 445 |     "                test_token =1\n",
 446 |     "        print()\n",
 447 |     "\n",
 448 |     "    time_elapsed = time.time() - since\n",
 449 |     "    print('Training complete in {:.0f}m {:.0f}s'.format(\n",
 450 |     "        time_elapsed // 60, time_elapsed % 60))\n",
 451 |     "    print('Best val Acc: {:4f}'.format(best_acc))\n",
 452 |     "\n",
 453 |     "    # load best model weights\n",
 454 |     "    model.load_state_dict(best_model_wts)\n",
 455 |     "    return model\n",
 456 |     "\n"
 457 |    ]
 458 |   },
 459 |   {
 460 |    "cell_type": "markdown",
 461 |    "id": "random-triple",
 462 |    "metadata": {
 463 |     "papermill": {
 464 |      "duration": 0.010677,
 465 |      "end_time": "2021-06-17T06:12:38.591815",
 466 |      "exception": false,
 467 |      "start_time": "2021-06-17T06:12:38.581138",
 468 |      "status": "completed"
 469 |     },
 470 |     "tags": []
 471 |    },
 472 |    "source": [
 473 |     "# Training with lr=0.01"
 474 |    ]
 475 |   },
 476 |   {
 477 |    "cell_type": "code",
 478 |    "execution_count": 4,
 479 |    "id": "suited-japan",
 480 |    "metadata": {
 481 |     "execution": {
 482 |      "iopub.execute_input": "2021-06-17T06:12:38.621711Z",
 483 |      "iopub.status.busy": "2021-06-17T06:12:38.621108Z",
 484 |      "iopub.status.idle": "2021-06-17T06:19:05.241260Z",
 485 |      "shell.execute_reply": "2021-06-17T06:19:05.240637Z",
 486 |      "shell.execute_reply.started": "2021-06-17T06:04:54.039059Z"
 487 |     },
 488 |     "papermill": {
 489 |      "duration": 386.638834,
 490 |      "end_time": "2021-06-17T06:19:05.241402",
 491 |      "exception": false,
 492 |      "start_time": "2021-06-17T06:12:38.602568",
 493 |      "status": "completed"
 494 |     },
 495 |     "tags": []
 496 |    },
 497 |    "outputs": [
 498 |     {
 499 |      "name": "stdout",
 500 |      "output_type": "stream",
 501 |      "text": [
 502 |       "Epoch 0/4\n",
 503 |       "----------\n",
 504 |       "train Loss: 0.6149 Acc: 0.6726\n",
 505 |       "val Loss: 0.4118 Acc: 0.8265\n",
 506 |       "test Loss: 0.4686 Acc: 0.7600\n",
 507 |       "\n",
 508 |       "Epoch 1/4\n",
 509 |       "----------\n",
 510 |       "train Loss: 0.4793 Acc: 0.7737\n",
 511 |       "val Loss: 0.3923 Acc: 0.8265\n",
 512 |       "\n",
 513 |       "Epoch 2/4\n",
 514 |       "----------\n",
 515 |       "train Loss: 0.3616 Acc: 0.8420\n",
 516 |       "val Loss: 0.4368 Acc: 0.7755\n",
 517 |       "\n",
 518 |       "Epoch 3/4\n",
 519 |       "----------\n",
 520 |       "train Loss: 0.3238 Acc: 0.8710\n",
 521 |       "val Loss: 0.3438 Acc: 0.8878\n",
 522 |       "test Loss: 0.3077 Acc: 0.8800\n",
 523 |       "\n",
 524 |       "Epoch 4/4\n",
 525 |       "----------\n",
 526 |       "train Loss: 0.2772 Acc: 0.9014\n",
 527 |       "val Loss: 0.2953 Acc: 0.8980\n",
 528 |       "test Loss: 0.3144 Acc: 0.8600\n",
 529 |       "\n",
 530 |       "Training complete in 6m 23s\n",
 531 |       "Best val Acc: 0.897959\n"
 532 |      ]
 533 |     }
 534 |    ],
 535 |    "source": [
 536 |     "model_ft = model_ft.to(device)\n",
 537 |     "\n",
 538 |     "criterion = nn.CrossEntropyLoss()\n",
 539 |     "\n",
 540 |     "# Observe that all parameters are being optimized\n",
 541 |     "optimizer_ft = optim.SGD(model_ft.parameters(), lr=0.01, momentum=0.9)\n",
 542 |     "\n",
 543 |     "# Decay LR by a factor of 0.1 every 7 epochs\n",
 544 |     "exp_lr_scheduler = lr_scheduler.StepLR(optimizer_ft, step_size=7, gamma=0.1)\n",
 545 |     "\n",
 546 |     "model_ft = train_model(model_ft, criterion, optimizer_ft, exp_lr_scheduler,\n",
 547 |     "                       num_epochs=5)"
 548 |    ]
 549 |   },
 550 |   {
 551 |    "cell_type": "markdown",
 552 |    "id": "brown-suite",
 553 |    "metadata": {
 554 |     "papermill": {
 555 |      "duration": 0.014777,
 556 |      "end_time": "2021-06-17T06:19:05.270913",
 557 |      "exception": false,
 558 |      "start_time": "2021-06-17T06:19:05.256136",
 559 |      "status": "completed"
 560 |     },
 561 |     "tags": []
 562 |    },
 563 |    "source": [
 564 |     "# Training with lr=0.001"
 565 |    ]
 566 |   },
 567 |   {
 568 |    "cell_type": "code",
 569 |    "execution_count": 5,
 570 |    "id": "union-grammar",
 571 |    "metadata": {
 572 |     "execution": {
 573 |      "iopub.execute_input": "2021-06-17T06:19:05.306597Z",
 574 |      "iopub.status.busy": "2021-06-17T06:19:05.305860Z",
 575 |      "iopub.status.idle": "2021-06-17T06:25:00.723039Z",
 576 |      "shell.execute_reply": "2021-06-17T06:25:00.723448Z",
 577 |      "shell.execute_reply.started": "2021-06-17T06:04:58.031104Z"
 578 |     },
 579 |     "papermill": {
 580 |      "duration": 355.437759,
 581 |      "end_time": "2021-06-17T06:25:00.723626",
 582 |      "exception": false,
 583 |      "start_time": "2021-06-17T06:19:05.285867",
 584 |      "status": "completed"
 585 |     },
 586 |     "tags": []
 587 |    },
 588 |    "outputs": [
 589 |     {
 590 |      "name": "stdout",
 591 |      "output_type": "stream",
 592 |      "text": [
 593 |       "Epoch 0/4\n",
 594 |       "----------\n",
 595 |       "train Loss: 0.1904 Acc: 0.9191\n",
 596 |       "val Loss: 0.1863 Acc: 0.9490\n",
 597 |       "test Loss: 0.1356 Acc: 0.9500\n",
 598 |       "\n",
 599 |       "Epoch 1/4\n",
 600 |       "----------\n",
 601 |       "train Loss: 0.1159 Acc: 0.9558\n",
 602 |       "val Loss: 0.1612 Acc: 0.9490\n",
 603 |       "\n",
 604 |       "Epoch 2/4\n",
 605 |       "----------\n",
 606 |       "train Loss: 0.1196 Acc: 0.9570\n",
 607 |       "val Loss: 0.1808 Acc: 0.9490\n",
 608 |       "\n",
 609 |       "Epoch 3/4\n",
 610 |       "----------\n",
 611 |       "train Loss: 0.0949 Acc: 0.9583\n",
 612 |       "val Loss: 0.1228 Acc: 0.9490\n",
 613 |       "\n",
 614 |       "Epoch 4/4\n",
 615 |       "----------\n",
 616 |       "train Loss: 0.0810 Acc: 0.9697\n",
 617 |       "val Loss: 0.1396 Acc: 0.9490\n",
 618 |       "\n",
 619 |       "Training complete in 5m 55s\n",
 620 |       "Best val Acc: 0.948980\n"
 621 |      ]
 622 |     }
 623 |    ],
 624 |    "source": [
 625 |     "# Observe that all parameters are being optimized\n",
 626 |     "optimizer_ft = optim.SGD(model_ft.parameters(), lr=0.001, momentum=0.9)\n",
 627 |     "\n",
 628 |     "# Decay LR by a factor of 0.1 every 7 epochs\n",
 629 |     "exp_lr_scheduler = lr_scheduler.StepLR(optimizer_ft, step_size=7, gamma=0.1)\n",
 630 |     "\n",
 631 |     "model_ft = train_model(model_ft, criterion, optimizer_ft, exp_lr_scheduler,\n",
 632 |     "                       num_epochs=5)"
 633 |    ]
 634 |   },
 635 |   {
 636 |    "cell_type": "markdown",
 637 |    "id": "finite-chester",
 638 |    "metadata": {
 639 |     "papermill": {
 640 |      "duration": 0.017422,
 641 |      "end_time": "2021-06-17T06:25:00.758656",
 642 |      "exception": false,
 643 |      "start_time": "2021-06-17T06:25:00.741234",
 644 |      "status": "completed"
 645 |     },
 646 |     "tags": []
 647 |    },
 648 |    "source": [
 649 |     "## Confusion Matrix\n",
 650 |     "\n",
 651 |     "### Validation Data"
 652 |    ]
 653 |   },
 654 |   {
 655 |    "cell_type": "code",
 656 |    "execution_count": 6,
 657 |    "id": "above-fellow",
 658 |    "metadata": {
 659 |     "execution": {
 660 |      "iopub.execute_input": "2021-06-17T06:25:00.801299Z",
 661 |      "iopub.status.busy": "2021-06-17T06:25:00.800697Z",
 662 |      "iopub.status.idle": "2021-06-17T06:25:06.244961Z",
 663 |      "shell.execute_reply": "2021-06-17T06:25:06.245333Z",
 664 |      "shell.execute_reply.started": "2021-06-17T06:11:18.983248Z"
 665 |     },
 666 |     "papermill": {
 667 |      "duration": 5.469343,
 668 |      "end_time": "2021-06-17T06:25:06.245480",
 669 |      "exception": false,
 670 |      "start_time": "2021-06-17T06:25:00.776137",
 671 |      "status": "completed"
 672 |     },
 673 |     "tags": []
 674 |    },
 675 |    "outputs": [
 676 |     {
 677 |      "data": {
 678 |       "text/plain": [
 679 |        "<AxesSubplot:>"
 680 |       ]
 681 |      },
 682 |      "execution_count": 6,
 683 |      "metadata": {},
 684 |      "output_type": "execute_result"
 685 |     },
 686 |     {
 687 |      "data": {
 688 |       "image/png": "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\n",
 689 |       "text/plain": [
 690 |        "<Figure size 600x200 with 2 Axes>"
 691 |       ]
 692 |      },
 693 |      "metadata": {
 694 |       "needs_background": "light"
 695 |      },
 696 |      "output_type": "display_data"
 697 |     }
 698 |    ],
 699 |    "source": [
 700 |     "from sklearn.metrics import confusion_matrix\n",
 701 |     "import seaborn as sn\n",
 702 |     "import pandas as pd\n",
 703 |     "\n",
 704 |     "y_pred = []\n",
 705 |     "y_true = []\n",
 706 |     "\n",
 707 |     "# iterate over test data\n",
 708 |     "for inputs, labels in dataloaders['val']:\n",
 709 |     "        inputs = inputs.to(device)\n",
 710 |     "        labels = labels.to(device)\n",
 711 |     "        \n",
 712 |     "        output = model_ft(inputs) # Feed Network\n",
 713 |     "\n",
 714 |     "        output = (torch.max(torch.exp(output), 1)[1]).data.cpu().numpy()\n",
 715 |     "        y_pred.extend(output) # Save Prediction\n",
 716 |     "        \n",
 717 |     "        labels = labels.data.cpu().numpy()\n",
 718 |     "        y_true.extend(labels) # Save Truth\n",
 719 |     "\n",
 720 |     "# constant for classes\n",
 721 |     "classes = ('Accident', 'Normal')\n",
 722 |     "\n",
 723 |     "# Build confusion matrix\n",
 724 |     "cf_matrix = confusion_matrix(y_true, y_pred)\n",
 725 |     "df_cm = pd.DataFrame(cf_matrix, index = [i for i in classes],\n",
 726 |     "                     columns = [i for i in classes])\n",
 727 |     "\n",
 728 |     "\n",
 729 |     "\n",
 730 |     "plt.figure(figsize = (6,2),dpi=100)\n",
 731 |     "plt.rcParams['font.size'] = '16'\n",
 732 |     "sn.heatmap(df_cm, annot=True, fmt=\".0f\")"
 733 |    ]
 734 |   },
 735 |   {
 736 |    "cell_type": "markdown",
 737 |    "id": "constitutional-equity",
 738 |    "metadata": {
 739 |     "papermill": {
 740 |      "duration": 0.01864,
 741 |      "end_time": "2021-06-17T06:25:06.283530",
 742 |      "exception": false,
 743 |      "start_time": "2021-06-17T06:25:06.264890",
 744 |      "status": "completed"
 745 |     },
 746 |     "tags": []
 747 |    },
 748 |    "source": [
 749 |     "### Test Data"
 750 |    ]
 751 |   },
 752 |   {
 753 |    "cell_type": "code",
 754 |    "execution_count": 7,
 755 |    "id": "shaped-journalist",
 756 |    "metadata": {
 757 |     "execution": {
 758 |      "iopub.execute_input": "2021-06-17T06:25:06.328395Z",
 759 |      "iopub.status.busy": "2021-06-17T06:25:06.327868Z",
 760 |      "iopub.status.idle": "2021-06-17T06:25:10.858779Z",
 761 |      "shell.execute_reply": "2021-06-17T06:25:10.859207Z",
 762 |      "shell.execute_reply.started": "2021-06-17T06:11:07.062855Z"
 763 |     },
 764 |     "papermill": {
 765 |      "duration": 4.556761,
 766 |      "end_time": "2021-06-17T06:25:10.859362",
 767 |      "exception": false,
 768 |      "start_time": "2021-06-17T06:25:06.302601",
 769 |      "status": "completed"
 770 |     },
 771 |     "tags": []
 772 |    },
 773 |    "outputs": [
 774 |     {
 775 |      "data": {
 776 |       "text/plain": [
 777 |        "<AxesSubplot:>"
 778 |       ]
 779 |      },
 780 |      "execution_count": 7,
 781 |      "metadata": {},
 782 |      "output_type": "execute_result"
 783 |     },
 784 |     {
 785 |      "data": {
 786 |       "image/png": "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\n",
 787 |       "text/plain": [
 788 |        "<Figure size 600x200 with 2 Axes>"
 789 |       ]
 790 |      },
 791 |      "metadata": {
 792 |       "needs_background": "light"
 793 |      },
 794 |      "output_type": "display_data"
 795 |     }
 796 |    ],
 797 |    "source": [
 798 |     "y_pred = []\n",
 799 |     "y_true = []\n",
 800 |     "\n",
 801 |     "# iterate over test data\n",
 802 |     "for inputs, labels in dataloaders['test']:\n",
 803 |     "        inputs = inputs.to(device)\n",
 804 |     "        labels = labels.to(device)\n",
 805 |     "        \n",
 806 |     "        output = model_ft(inputs) # Feed Network\n",
 807 |     "\n",
 808 |     "        output = (torch.max(torch.exp(output), 1)[1]).data.cpu().numpy()\n",
 809 |     "        y_pred.extend(output) # Save Prediction\n",
 810 |     "        \n",
 811 |     "        labels = labels.data.cpu().numpy()\n",
 812 |     "        y_true.extend(labels) # Save Truth\n",
 813 |     "\n",
 814 |     "\n",
 815 |     "# Build confusion matrix\n",
 816 |     "cf_matrix = confusion_matrix(y_true, y_pred)\n",
 817 |     "df_cm = pd.DataFrame(cf_matrix, index = [i for i in classes],\n",
 818 |     "                     columns = [i for i in classes])\n",
 819 |     "\n",
 820 |     "\n",
 821 |     "\n",
 822 |     "plt.figure(figsize = (6,2),dpi=100)\n",
 823 |     "plt.rcParams['font.size'] = '16'\n",
 824 |     "sn.heatmap(df_cm, annot=True, fmt=\".0f\")"
 825 |    ]
 826 |   }
 827 |  ],
 828 |  "metadata": {
 829 |   "kernelspec": {
 830 |    "display_name": "Python 3",
 831 |    "language": "python",
 832 |    "name": "python3"
 833 |   },
 834 |   "language_info": {
 835 |    "codemirror_mode": {
 836 |     "name": "ipython",
 837 |     "version": 3
 838 |    },
 839 |    "file_extension": ".py",
 840 |    "mimetype": "text/x-python",
 841 |    "name": "python",
 842 |    "nbconvert_exporter": "python",
 843 |    "pygments_lexer": "ipython3",
 844 |    "version": "3.7.9"
 845 |   },
 846 |   "papermill": {
 847 |    "default_parameters": {},
 848 |    "duration": 788.176224,
 849 |    "end_time": "2021-06-17T06:25:11.789186",
 850 |    "environment_variables": {},
 851 |    "exception": null,
 852 |    "input_path": "__notebook__.ipynb",
 853 |    "output_path": "__notebook__.ipynb",
 854 |    "parameters": {},
 855 |    "start_time": "2021-06-17T06:12:03.612962",
 856 |    "version": "2.3.2"
 857 |   },
 858 |   "widgets": {
 859 |    "application/vnd.jupyter.widget-state+json": {
 860 |     "state": {
 861 |      "1534c8af727a4f6bbfee39b99b598433": {
 862 |       "model_module": "@jupyter-widgets/controls",
 863 |       "model_module_version": "1.5.0",
 864 |       "model_name": "HBoxModel",
 865 |       "state": {
 866 |        "_dom_classes": [],
 867 |        "_model_module": "@jupyter-widgets/controls",
 868 |        "_model_module_version": "1.5.0",
 869 |        "_model_name": "HBoxModel",
 870 |        "_view_count": null,
 871 |        "_view_module": "@jupyter-widgets/controls",
 872 |        "_view_module_version": "1.5.0",
 873 |        "_view_name": "HBoxView",
 874 |        "box_style": "",
 875 |        "children": [
 876 |         "IPY_MODEL_a7c6dcfddd614bb68dd82e716d8377a0",
 877 |         "IPY_MODEL_9a4dca1f75454ba5837c55beec9286f3",
 878 |         "IPY_MODEL_93dc77cdf0b74bf1bc5ca2b837f8019d"
 879 |        ],
 880 |        "layout": "IPY_MODEL_f23b1483f4124b3b8abb31c800ff57e0"
 881 |       }
 882 |      },
 883 |      "93dc77cdf0b74bf1bc5ca2b837f8019d": {
 884 |       "model_module": "@jupyter-widgets/controls",
 885 |       "model_module_version": "1.5.0",
 886 |       "model_name": "HTMLModel",
 887 |       "state": {
 888 |        "_dom_classes": [],
 889 |        "_model_module": "@jupyter-widgets/controls",
 890 |        "_model_module_version": "1.5.0",
 891 |        "_model_name": "HTMLModel",
 892 |        "_view_count": null,
 893 |        "_view_module": "@jupyter-widgets/controls",
 894 |        "_view_module_version": "1.5.0",
 895 |        "_view_name": "HTMLView",
 896 |        "description": "",
 897 |        "description_tooltip": null,
 898 |        "layout": "IPY_MODEL_f9f319b591d04a5582ba154669f8fcae",
 899 |        "placeholder": "​",
 900 |        "style": "IPY_MODEL_dc73a049f79145518f4b7616c1bf787b",
 901 |        "value": " 548M/548M [00:23&lt;00:00, 25.6MB/s]"
 902 |       }
 903 |      },
 904 |      "9a4dca1f75454ba5837c55beec9286f3": {
 905 |       "model_module": "@jupyter-widgets/controls",
 906 |       "model_module_version": "1.5.0",
 907 |       "model_name": "FloatProgressModel",
 908 |       "state": {
 909 |        "_dom_classes": [],
 910 |        "_model_module": "@jupyter-widgets/controls",
 911 |        "_model_module_version": "1.5.0",
 912 |        "_model_name": "FloatProgressModel",
 913 |        "_view_count": null,
 914 |        "_view_module": "@jupyter-widgets/controls",
 915 |        "_view_module_version": "1.5.0",
 916 |        "_view_name": "ProgressView",
 917 |        "bar_style": "success",
 918 |        "description": "",
 919 |        "description_tooltip": null,
 920 |        "layout": "IPY_MODEL_bb7f62c39f984afeb9e825e138110ef4",
 921 |        "max": 574769405.0,
 922 |        "min": 0.0,
 923 |        "orientation": "horizontal",
 924 |        "style": "IPY_MODEL_b29fe8088a5f478a896b84d4dd19975c",
 925 |        "value": 574769405.0
 926 |       }
 927 |      },
 928 |      "a7c6dcfddd614bb68dd82e716d8377a0": {
 929 |       "model_module": "@jupyter-widgets/controls",
 930 |       "model_module_version": "1.5.0",
 931 |       "model_name": "HTMLModel",
 932 |       "state": {
 933 |        "_dom_classes": [],
 934 |        "_model_module": "@jupyter-widgets/controls",
 935 |        "_model_module_version": "1.5.0",
 936 |        "_model_name": "HTMLModel",
 937 |        "_view_count": null,
 938 |        "_view_module": "@jupyter-widgets/controls",
 939 |        "_view_module_version": "1.5.0",
 940 |        "_view_name": "HTMLView",
 941 |        "description": "",
 942 |        "description_tooltip": null,
 943 |        "layout": "IPY_MODEL_f0ede7f0058949e6962971a3aafea582",
 944 |        "placeholder": "​",
 945 |        "style": "IPY_MODEL_cf8a8d029ee94ba39d73afc14fbd7dcf",
 946 |        "value": "100%"
 947 |       }
 948 |      },
 949 |      "b29fe8088a5f478a896b84d4dd19975c": {
 950 |       "model_module": "@jupyter-widgets/controls",
 951 |       "model_module_version": "1.5.0",
 952 |       "model_name": "ProgressStyleModel",
 953 |       "state": {
 954 |        "_model_module": "@jupyter-widgets/controls",
 955 |        "_model_module_version": "1.5.0",
 956 |        "_model_name": "ProgressStyleModel",
 957 |        "_view_count": null,
 958 |        "_view_module": "@jupyter-widgets/base",
 959 |        "_view_module_version": "1.2.0",
 960 |        "_view_name": "StyleView",
 961 |        "bar_color": null,
 962 |        "description_width": ""
 963 |       }
 964 |      },
 965 |      "bb7f62c39f984afeb9e825e138110ef4": {
 966 |       "model_module": "@jupyter-widgets/base",
 967 |       "model_module_version": "1.2.0",
 968 |       "model_name": "LayoutModel",
 969 |       "state": {
 970 |        "_model_module": "@jupyter-widgets/base",
 971 |        "_model_module_version": "1.2.0",
 972 |        "_model_name": "LayoutModel",
 973 |        "_view_count": null,
 974 |        "_view_module": "@jupyter-widgets/base",
 975 |        "_view_module_version": "1.2.0",
 976 |        "_view_name": "LayoutView",
 977 |        "align_content": null,
 978 |        "align_items": null,
 979 |        "align_self": null,
 980 |        "border": null,
 981 |        "bottom": null,
 982 |        "display": null,
 983 |        "flex": null,
 984 |        "flex_flow": null,
 985 |        "grid_area": null,
 986 |        "grid_auto_columns": null,
 987 |        "grid_auto_flow": null,
 988 |        "grid_auto_rows": null,
 989 |        "grid_column": null,
 990 |        "grid_gap": null,
 991 |        "grid_row": null,
 992 |        "grid_template_areas": null,
 993 |        "grid_template_columns": null,
 994 |        "grid_template_rows": null,
 995 |        "height": null,
 996 |        "justify_content": null,
 997 |        "justify_items": null,
 998 |        "left": null,
 999 |        "margin": null,
1000 |        "max_height": null,
1001 |        "max_width": null,
1002 |        "min_height": null,
1003 |        "min_width": null,
1004 |        "object_fit": null,
1005 |        "object_position": null,
1006 |        "order": null,
1007 |        "overflow": null,
1008 |        "overflow_x": null,
1009 |        "overflow_y": null,
1010 |        "padding": null,
1011 |        "right": null,
1012 |        "top": null,
1013 |        "visibility": null,
1014 |        "width": null
1015 |       }
1016 |      },
1017 |      "cf8a8d029ee94ba39d73afc14fbd7dcf": {
1018 |       "model_module": "@jupyter-widgets/controls",
1019 |       "model_module_version": "1.5.0",
1020 |       "model_name": "DescriptionStyleModel",
1021 |       "state": {
1022 |        "_model_module": "@jupyter-widgets/controls",
1023 |        "_model_module_version": "1.5.0",
1024 |        "_model_name": "DescriptionStyleModel",
1025 |        "_view_count": null,
1026 |        "_view_module": "@jupyter-widgets/base",
1027 |        "_view_module_version": "1.2.0",
1028 |        "_view_name": "StyleView",
1029 |        "description_width": ""
1030 |       }
1031 |      },
1032 |      "dc73a049f79145518f4b7616c1bf787b": {
1033 |       "model_module": "@jupyter-widgets/controls",
1034 |       "model_module_version": "1.5.0",
1035 |       "model_name": "DescriptionStyleModel",
1036 |       "state": {
1037 |        "_model_module": "@jupyter-widgets/controls",
1038 |        "_model_module_version": "1.5.0",
1039 |        "_model_name": "DescriptionStyleModel",
1040 |        "_view_count": null,
1041 |        "_view_module": "@jupyter-widgets/base",
1042 |        "_view_module_version": "1.2.0",
1043 |        "_view_name": "StyleView",
1044 |        "description_width": ""
1045 |       }
1046 |      },
1047 |      "f0ede7f0058949e6962971a3aafea582": {
1048 |       "model_module": "@jupyter-widgets/base",
1049 |       "model_module_version": "1.2.0",
1050 |       "model_name": "LayoutModel",
1051 |       "state": {
1052 |        "_model_module": "@jupyter-widgets/base",
1053 |        "_model_module_version": "1.2.0",
1054 |        "_model_name": "LayoutModel",
1055 |        "_view_count": null,
1056 |        "_view_module": "@jupyter-widgets/base",
1057 |        "_view_module_version": "1.2.0",
1058 |        "_view_name": "LayoutView",
1059 |        "align_content": null,
1060 |        "align_items": null,
1061 |        "align_self": null,
1062 |        "border": null,
1063 |        "bottom": null,
1064 |        "display": null,
1065 |        "flex": null,
1066 |        "flex_flow": null,
1067 |        "grid_area": null,
1068 |        "grid_auto_columns": null,
1069 |        "grid_auto_flow": null,
1070 |        "grid_auto_rows": null,
1071 |        "grid_column": null,
1072 |        "grid_gap": null,
1073 |        "grid_row": null,
1074 |        "grid_template_areas": null,
1075 |        "grid_template_columns": null,
1076 |        "grid_template_rows": null,
1077 |        "height": null,
1078 |        "justify_content": null,
1079 |        "justify_items": null,
1080 |        "left": null,
1081 |        "margin": null,
1082 |        "max_height": null,
1083 |        "max_width": null,
1084 |        "min_height": null,
1085 |        "min_width": null,
1086 |        "object_fit": null,
1087 |        "object_position": null,
1088 |        "order": null,
1089 |        "overflow": null,
1090 |        "overflow_x": null,
1091 |        "overflow_y": null,
1092 |        "padding": null,
1093 |        "right": null,
1094 |        "top": null,
1095 |        "visibility": null,
1096 |        "width": null
1097 |       }
1098 |      },
1099 |      "f23b1483f4124b3b8abb31c800ff57e0": {
1100 |       "model_module": "@jupyter-widgets/base",
1101 |       "model_module_version": "1.2.0",
1102 |       "model_name": "LayoutModel",
1103 |       "state": {
1104 |        "_model_module": "@jupyter-widgets/base",
1105 |        "_model_module_version": "1.2.0",
1106 |        "_model_name": "LayoutModel",
1107 |        "_view_count": null,
1108 |        "_view_module": "@jupyter-widgets/base",
1109 |        "_view_module_version": "1.2.0",
1110 |        "_view_name": "LayoutView",
1111 |        "align_content": null,
1112 |        "align_items": null,
1113 |        "align_self": null,
1114 |        "border": null,
1115 |        "bottom": null,
1116 |        "display": null,
1117 |        "flex": null,
1118 |        "flex_flow": null,
1119 |        "grid_area": null,
1120 |        "grid_auto_columns": null,
1121 |        "grid_auto_flow": null,
1122 |        "grid_auto_rows": null,
1123 |        "grid_column": null,
1124 |        "grid_gap": null,
1125 |        "grid_row": null,
1126 |        "grid_template_areas": null,
1127 |        "grid_template_columns": null,
1128 |        "grid_template_rows": null,
1129 |        "height": null,
1130 |        "justify_content": null,
1131 |        "justify_items": null,
1132 |        "left": null,
1133 |        "margin": null,
1134 |        "max_height": null,
1135 |        "max_width": null,
1136 |        "min_height": null,
1137 |        "min_width": null,
1138 |        "object_fit": null,
1139 |        "object_position": null,
1140 |        "order": null,
1141 |        "overflow": null,
1142 |        "overflow_x": null,
1143 |        "overflow_y": null,
1144 |        "padding": null,
1145 |        "right": null,
1146 |        "top": null,
1147 |        "visibility": null,
1148 |        "width": null
1149 |       }
1150 |      },
1151 |      "f9f319b591d04a5582ba154669f8fcae": {
1152 |       "model_module": "@jupyter-widgets/base",
1153 |       "model_module_version": "1.2.0",
1154 |       "model_name": "LayoutModel",
1155 |       "state": {
1156 |        "_model_module": "@jupyter-widgets/base",
1157 |        "_model_module_version": "1.2.0",
1158 |        "_model_name": "LayoutModel",
1159 |        "_view_count": null,
1160 |        "_view_module": "@jupyter-widgets/base",
1161 |        "_view_module_version": "1.2.0",
1162 |        "_view_name": "LayoutView",
1163 |        "align_content": null,
1164 |        "align_items": null,
1165 |        "align_self": null,
1166 |        "border": null,
1167 |        "bottom": null,
1168 |        "display": null,
1169 |        "flex": null,
1170 |        "flex_flow": null,
1171 |        "grid_area": null,
1172 |        "grid_auto_columns": null,
1173 |        "grid_auto_flow": null,
1174 |        "grid_auto_rows": null,
1175 |        "grid_column": null,
1176 |        "grid_gap": null,
1177 |        "grid_row": null,
1178 |        "grid_template_areas": null,
1179 |        "grid_template_columns": null,
1180 |        "grid_template_rows": null,
1181 |        "height": null,
1182 |        "justify_content": null,
1183 |        "justify_items": null,
1184 |        "left": null,
1185 |        "margin": null,
1186 |        "max_height": null,
1187 |        "max_width": null,
1188 |        "min_height": null,
1189 |        "min_width": null,
1190 |        "object_fit": null,
1191 |        "object_position": null,
1192 |        "order": null,
1193 |        "overflow": null,
1194 |        "overflow_x": null,
1195 |        "overflow_y": null,
1196 |        "padding": null,
1197 |        "right": null,
1198 |        "top": null,
1199 |        "visibility": null,
1200 |        "width": null
1201 |       }
1202 |      }
1203 |     },
1204 |     "version_major": 2,
1205 |     "version_minor": 0
1206 |    }
1207 |   }
1208 |  },
1209 |  "nbformat": 4,
1210 |  "nbformat_minor": 5
1211 | }
1212 | 


--------------------------------------------------------------------------------
/Hyperspectral Image Classification/HSI_Confusion Matrix.png:
--------------------------------------------------------------------------------
https://raw.githubusercontent.com/dipuk0506/SpinalNet-Applications/08912f00345b1912476b6f62cd8a3b7823ed0fa5/Hyperspectral Image Classification/HSI_Confusion Matrix.png


--------------------------------------------------------------------------------
/Hyperspectral Image Classification/HSI_KSC.py:
--------------------------------------------------------------------------------
  1 | # -*- coding: utf-8 -*-
  2 | """KSC_PCA.ipynb
  3 | 
  4 | Automatically generated by Colaboratory.
  5 | 
  6 | Original file is located at
  7 |     https://colab.research.google.com/drive/13bVyVdv-yBFo30C4Ce_Q4eKjwpb5Sm8y
  8 | """
  9 | 
 10 | # Commented out IPython magic to ensure Python compatibility.
 11 | import numpy as np
 12 | from sklearn.decomposition import PCA
 13 | import scipy.io as sio
 14 | from sklearn.model_selection import train_test_split
 15 | from sklearn import preprocessing
 16 | import os
 17 | import random
 18 | from random import shuffle
 19 | from skimage.transform import rotate
 20 | import scipy.ndimage
 21 | import matplotlib.pyplot as plt
 22 | # %matplotlib inline
 23 | 
 24 | import torch
 25 | 
 26 | from torch.utils.data import Dataset, DataLoader
 27 | 
 28 | hsi_data= sio.loadmat('KSC.mat')['KSC']
 29 | labels = sio.loadmat('KSC_gt.mat')['KSC_gt']
 30 | 
 31 | [height,width,depth]=hsi_data.shape
 32 | 
 33 | def splitTrainTestSet(X, y, testRatio=0.10):
 34 |     X_train, X_test, y_train, y_test = train_test_split(X, y, test_size=testRatio, random_state=345,
 35 |                                                         stratify=y)
 36 |     return X_train, X_test, y_train, y_test
 37 | 
 38 | def oversampleWeakClasses(X, y):
 39 |     uniqueLabels, labelCounts = np.unique(y, return_counts=True)
 40 |     maxCount = np.max(labelCounts)
 41 |     labelInverseRatios = maxCount / labelCounts  
 42 |     # repeat for every label and concat
 43 |     newX = X[y == uniqueLabels[0], :, :, :].repeat(round(labelInverseRatios[0]), axis=0)
 44 |     newY = y[y == uniqueLabels[0]].repeat(round(labelInverseRatios[0]), axis=0)
 45 |     for label, labelInverseRatio in zip(uniqueLabels[1:], labelInverseRatios[1:]):
 46 |         cX = X[y== label,:,:,:].repeat(round(labelInverseRatio), axis=0)
 47 |         cY = y[y == label].repeat(round(labelInverseRatio), axis=0)
 48 |         newX = np.concatenate((newX, cX))
 49 |         newY = np.concatenate((newY, cY))
 50 |     np.random.seed(seed=42)
 51 |     rand_perm = np.random.permutation(newY.shape[0])
 52 |     newX = newX[rand_perm, :, :, :]
 53 |     newY = newY[rand_perm]
 54 |     return newX, newY
 55 | 
 56 | def standartizeData(X):
 57 |     newX = np.reshape(X, (-1, X.shape[2]))
 58 |     scaler = preprocessing.StandardScaler().fit(newX)  
 59 |     newX = scaler.transform(newX)
 60 |     newX = np.reshape(newX, (X.shape[0],X.shape[1],X.shape[2]))
 61 |     return newX, scaler
 62 | 
 63 | def applyPCA(X, numComponents=75):
 64 |     newX = np.reshape(X, (-1, X.shape[2]))
 65 |     pca = PCA(n_components=numComponents, whiten=True)
 66 |     newX = pca.fit_transform(newX)
 67 |     newX = np.reshape(newX, (X.shape[0],X.shape[1], numComponents))
 68 |     return newX, pca
 69 | 
 70 | def padWithZeros(X, margin=2):
 71 |     newX = np.zeros((X.shape[0] + 2 * margin, X.shape[1] + 2* margin, X.shape[2]))
 72 |     x_offset = margin
 73 |     y_offset = margin
 74 |     newX[x_offset:X.shape[0] + x_offset, y_offset:X.shape[1] + y_offset, :] = X
 75 |     return newX
 76 | 
 77 | def createPatches(X, y, windowSize=11, removeZeroLabels = True):
 78 |     margin = int((windowSize - 1) / 2)
 79 |     zeroPaddedX = padWithZeros(X, margin=margin)
 80 |     # split patches
 81 |     patchesData = np.zeros((X.shape[0] * X.shape[1], windowSize, windowSize, X.shape[2]))
 82 |     patchesLabels = np.zeros((X.shape[0] * X.shape[1]))
 83 |     patchIndex = 0
 84 |     for r in range(margin, zeroPaddedX.shape[0] - margin):
 85 |         for c in range(margin, zeroPaddedX.shape[1] - margin):
 86 |             patch = zeroPaddedX[r - margin:r + margin + 1, c - margin:c + margin + 1]   
 87 |             patchesData[patchIndex, :, :, :] = patch
 88 |             patchesLabels[patchIndex] = y[r-margin, c-margin]
 89 |             patchIndex = patchIndex + 1
 90 |     if removeZeroLabels:
 91 |         patchesData = patchesData[patchesLabels>0,:,:,:]
 92 |         patchesLabels = patchesLabels[patchesLabels>0]
 93 |         patchesLabels -= 1
 94 |     return patchesData, patchesLabels
 95 | 
 96 | 
 97 | def AugmentData(X_train):
 98 |     for i in range(int(X_train.shape[0]/2)):
 99 |         patch = X_train[i,:,:,:]
100 |         num = random.randint(0,2)
101 |         if (num == 0):
102 |             
103 |             flipped_patch = np.flipud(patch)
104 |         if (num == 1):
105 |             
106 |             flipped_patch = np.fliplr(patch)
107 |         if (num == 2):
108 |             
109 |             no = random.randrange(-180,180,30)
110 |             flipped_patch = scipy.ndimage.interpolation.rotate(patch, no,axes=(1, 0),
111 |                                                                reshape=False, output=None, order=3, mode='constant', cval=0.0, prefilter=False)
112 |     
113 |     
114 |     patch2 = flipped_patch
115 |     X_train[i,:,:,:] = patch2
116 |     
117 |     return X_train
118 | 
119 | import numpy as np
120 | import scipy
121 | import os
122 | from keras.models import Sequential
123 | from keras.layers import Dense, Dropout, Flatten
124 | from keras.layers import Conv2D, MaxPooling2D,BatchNormalization
125 | from tensorflow.keras.callbacks import EarlyStopping
126 | from tensorflow.keras.optimizers import SGD
127 | from keras import backend as K
128 | from keras.utils import np_utils
129 | from keras.utils.vis_utils import plot_model
130 | 
131 | weight_of_size=10
132 | 
133 | X=hsi_data
134 | y=labels
135 | 
136 | X,pca = applyPCA(X,30)
137 | 
138 | X.shape
139 | 
140 | XPatches, yPatches = createPatches(X, y, windowSize=15)
141 | 
142 | XPatches.shape
143 | 
144 | X_train, X_test, y_train, y_test = splitTrainTestSet(XPatches, yPatches, testRatio=0.2)
145 | 
146 | X_train=np.reshape(X_train,(X_train.shape[0],X_train.shape[3],X_train.shape[1],X_train.shape[1]))
147 | 
148 | X_test=np.reshape(X_test,(X_test.shape[0],X_test.shape[3],X_test.shape[1],X_test.shape[1]))
149 | 
150 | X_train.shape
151 | 
152 | X_train.shape
153 | 
154 | y_train
155 | 
156 | class MyDataset(Dataset):
157 |     def __init__(self, data, target, transform=None):
158 |         self.data = torch.from_numpy(data).float()
159 |         self.target = torch.from_numpy(target).int()
160 |         self.transform = transform
161 |         
162 |     def __getitem__(self, index):
163 |         x = self.data[index]
164 |         y = self.target[index]
165 |         
166 |         if self.transform:
167 |             x = self.transform(x)
168 |             
169 |         return x, y
170 |     
171 |     def __len__(self):
172 |         return len(self.data)
173 | 
174 | data_train = MyDataset(X_train, y_train)
175 | 
176 | input_shape= X_train[0].shape
177 | print(input_shape)
178 | 
179 | data_train.__getitem__(0)[0].shape
180 | 
181 | n_epochs = 8
182 | batch_size_train = 16
183 | batch_size_test = 10
184 | learning_rate = 0.01
185 | momentum = 0.5
186 | log_interval = 100
187 | first_HL = 8
188 | 
189 | import torch
190 | import torchvision
191 | 
192 | ## Call the Dataset Class 
193 | #data_train = torchvision.datasets.IndianPines('./data',download=True,PATCH_LENGTH=2)
194 | 
195 | ## Check the shapes
196 | print(data_train.__getitem__(0)[0].shape)
197 | print(data_train.__len__())
198 | 
199 | 
200 | ## Wrap it around a Torch Dataloader
201 | train_loader = torch.utils.data.DataLoader(data_train,batch_size=16,shuffle=True, num_workers=2)
202 | 
203 | len(data_train)
204 | 
205 | data_test=MyDataset(X_test, y_test)
206 | 
207 | import torch
208 | import torchvision
209 | 
210 | ## Call the Dataset Class 
211 | #data_test = torchvision.datasets.IndianPines('./data',download=True,PATCH_LENGTH=2)
212 | 
213 | ## Check the shapes
214 | print(data_test.__getitem__(0)[0].shape)
215 | print(data_test.__len__())
216 | 
217 | test_loader = torch.utils.data.DataLoader(data_test,batch_size=10,shuffle=False, num_workers=2)
218 | 
219 | examples = enumerate(test_loader)
220 | batch_idx, (example_data, example_targets) = next(examples)
221 | 
222 | print(example_data.shape)
223 | 
224 | import matplotlib.pyplot as plt
225 | 
226 | fig = plt.figure()
227 | for i in range(6):
228 |   plt.subplot(2,3,i+1)
229 |   plt.tight_layout()
230 |   plt.imshow(example_data[i][0], interpolation='none')
231 |   plt.title("Ground Truth: {}".format(example_targets[i]))
232 |   plt.xticks([])
233 |   plt.yticks([])
234 | fig
235 | 
236 | import torch
237 | import torch.nn as nn
238 | import torchvision
239 | import torchvision.transforms as transforms
240 | 
241 | import random
242 | 
243 | device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
244 | 
245 | # Hyper-parameters
246 | num_epochs = 16
247 | learning_rate = 0.001
248 | 
249 | torch.manual_seed(0)
250 | random.seed(0)
251 | 
252 | Half_width =60
253 | layer_width = 20
254 | 
255 | class SpinalCNN(nn.Module):
256 |     """CNN."""
257 | 
258 |     def __init__(self):
259 |         """CNN Builder."""
260 |         super(SpinalCNN, self).__init__()
261 | 
262 |         self.conv_layer = nn.Sequential(
263 | 
264 |             # Conv Layer block 1
265 |             nn.Conv2d(in_channels=30, out_channels=15, kernel_size=3, padding=1),
266 |             nn.BatchNorm2d(15),
267 |             nn.ReLU(inplace=True),
268 |             nn.Conv2d(in_channels=15, out_channels=30, kernel_size=3, padding=1),
269 |             nn.ReLU(inplace=True),
270 |             nn.MaxPool2d(kernel_size=2, stride=2),
271 | 
272 |             # Conv Layer block 2
273 |             nn.Conv2d(in_channels=30, out_channels=60, kernel_size=3, padding=1),
274 |             nn.BatchNorm2d(60),
275 |             nn.ReLU(inplace=True),
276 |             nn.Conv2d(in_channels=60, out_channels=60, kernel_size=3, padding=1),
277 |             nn.ReLU(inplace=True),
278 |             nn.MaxPool2d(kernel_size=2, stride=2),
279 |             nn.Dropout2d(p=0.05),
280 | 
281 |             # Conv Layer block 3
282 |             nn.Conv2d(in_channels=60, out_channels=120, kernel_size=3, padding=1),
283 |             nn.BatchNorm2d(120),
284 |             nn.ReLU(inplace=True),
285 |             nn.Conv2d(in_channels=120, out_channels=120, kernel_size=3, padding=1),
286 |             nn.ReLU(inplace=True),
287 |             nn.MaxPool2d(kernel_size=2, stride=2),
288 |         )
289 |         
290 |         self.fc_spinal_layer1 = nn.Sequential(
291 |             nn.Dropout(p=0.1), nn.Linear(Half_width, layer_width),
292 |             nn.ReLU(inplace=True),
293 |             )
294 |         self.fc_spinal_layer2 = nn.Sequential(
295 |             nn.Dropout(p=0.1), nn.Linear(Half_width + layer_width, layer_width),
296 |             nn.ReLU(inplace=True),
297 |             )
298 |         self.fc_spinal_layer3 = nn.Sequential(
299 |             nn.Dropout(p=0.1), nn.Linear(Half_width + layer_width, layer_width),
300 |             nn.ReLU(inplace=True),
301 |             )
302 |         self.fc_spinal_layer4 = nn.Sequential(
303 |             nn.Dropout(p=0.1), nn.Linear(Half_width + layer_width, layer_width),
304 |             nn.ReLU(inplace=True),
305 |             )
306 |         self.fc_out = nn.Sequential(
307 |             nn.Dropout(p=0.1), nn.Linear(layer_width*4, 16)            
308 |             )
309 | 
310 | 
311 |     def forward(self, x):
312 |         """Perform forward."""
313 |         
314 |         # conv layers
315 |         x = self.conv_layer(x)
316 |         
317 |         # flatten
318 |         x = x.view(x.size(0), -1)
319 |         
320 |         x1 = self.fc_spinal_layer1(x[:, 0:Half_width])
321 |         x2 = self.fc_spinal_layer2(torch.cat([ x[:,Half_width:2*Half_width], x1], dim=1))
322 |         x3 = self.fc_spinal_layer3(torch.cat([ x[:,0:Half_width], x2], dim=1))
323 |         x4 = self.fc_spinal_layer4(torch.cat([ x[:,Half_width:2*Half_width], x3], dim=1))
324 |         
325 |         x = torch.cat([x1, x2], dim=1)
326 |         x = torch.cat([x, x3], dim=1)
327 |         x = torch.cat([x, x4], dim=1)
328 | 
329 |         x = self.fc_out(x)
330 | 
331 |         return x
332 | 
333 | from tensorflow.keras.optimizers import Adam
334 | model = SpinalCNN().to(device)
335 | # defining the optimizer
336 | optimizer = Adam(model.parameters(), lr=0.07)
337 | # defining the loss function
338 | criterion = nn.CrossEntropyLoss()
339 | # checking if GPU is available
340 | if torch.cuda.is_available():
341 |     model = model.cuda()
342 |     criterion = criterion.cuda()
343 |     
344 | print(model)
345 | 
346 | from torchvision import models
347 | from torchsummary import summary
348 | model = SpinalCNN().to(device)
349 | summary(model, (30, 15, 15))
350 | 
351 | model = SpinalCNN().to(device)
352 | 
353 | 
354 | 
355 | # Loss and optimizer
356 | criterion = nn.CrossEntropyLoss()
357 | optimizer = torch.optim.Adam(model.parameters(), lr=learning_rate)
358 | 
359 | # For updating learning rate
360 | def update_lr(optimizer, lr):    
361 |     for param_group in optimizer.param_groups:
362 |         param_group['lr'] = lr
363 | 
364 | # Train the model
365 | total_step = len(train_loader)
366 | curr_lr = learning_rate
367 | for epoch in range(num_epochs):
368 |     for i, (images, labels) in enumerate(train_loader):
369 |         images = images.to(device)
370 |         labels = labels.to(device)
371 |        
372 |         #print(images.shape)
373 |         # Forward pass
374 |         outputs = model(images)
375 |         labels = torch.tensor(labels, dtype=torch.long, device=device)
376 |         loss = criterion(outputs, labels)
377 | 
378 |         # Backward and optimize
379 |         optimizer.zero_grad()
380 |         loss.backward()
381 |         optimizer.step()
382 | 
383 | 
384 |         if (i+1) % 500 == 0:
385 |           print("Epoch [{}/{}], Step [{}/{}] Loss: {:.4f}"
386 |                 .format(epoch+1, num_epochs, i+1, total_step, loss.item()))
387 |             
388 | 
389 |     # Decay learning rate
390 |     if (epoch) == 1 or epoch>20:
391 |         curr_lr /= 3
392 |         update_lr(optimizer, curr_lr)
393 |         
394 |         # Test the model
395 |     model.eval()
396 |     with torch.no_grad():
397 |       correct = 0
398 |       total = 0
399 |       predicted_numpy=[]
400 |       for images, labels in test_loader:
401 | 
402 |         images = images.to(device)
403 |         labels = labels.to(device)
404 |         outputs = model(images)
405 |         _, predicted = torch.max(outputs.data, 1)
406 |         predicted_numpy.append(predicted.cpu().numpy())
407 | 
408 |         total += labels.size(0)
409 |         correct += (predicted == labels).sum().item()
410 |         
411 |       print('Accuracy of the model on the test images: {} %'.format(100 * correct / total))
412 |                 
413 |       model.train()
414 | 
415 | len(predicted_numpy)
416 | 
417 | predicted_numpy = np.concatenate(predicted_numpy)
418 | 
419 | predicted_numpy.shape
420 | 
421 | from sklearn.metrics import confusion_matrix
422 | from sklearn.metrics import plot_confusion_matrix
423 | y_true = y_test
424 | y_pred = predicted_numpy
425 | print(confusion_matrix(y_true, y_pred), end='\n')
426 | 
427 | from sklearn.metrics import confusion_matrix
428 | from sklearn.metrics import plot_confusion_matrix
429 | import seaborn as sn
430 | y_true = y_test
431 | y_pred = predicted_numpy
432 | plt.figure(figsize = (10,7))
433 | print(sn.heatmap(confusion_matrix(y_true, y_pred),annot=True,cmap="OrRd",fmt='d'))
434 | plt.savefig('KSC_pca_Confusion Matrix')
435 | 
436 | from sklearn.metrics import cohen_kappa_score
437 | print(cohen_kappa_score(y_true, y_pred, labels=None, weights=None, sample_weight=None))
438 | 
439 | from sklearn.metrics import classification_report
440 | 
441 | from sklearn.metrics import confusion_matrix
442 | from sklearn.metrics import plot_confusion_matrix
443 | y_true = y_test
444 | y_pred = predicted_numpy
445 | target_names = ['class 0', 'class 1', 'class 2', 'class 3', 'class 4', 'class 5', 'class 6', 'class 7', 'class 8', 'class 9', 'class 10', 'class 11', 'class 12']
446 | print(classification_report(y_true, y_pred, target_names=target_names, digits=4))
447 | 
448 | f = sio.loadmat('KSC.mat')['KSC']
449 | g = sio.loadmat('KSC_gt.mat')['KSC_gt']
450 | 
451 | F,pca = applyPCA(f,30)
452 | 
453 | FPatches, gPatches = createPatches(F,g, windowSize=15)
454 | 
455 | import itertools
456 | 
457 | def classified_pixels(FPatches,gPatches,g):
458 |   FPatches=np.reshape(FPatches,(FPatches.shape[0],FPatches.shape[3],FPatches.shape[1],FPatches.shape[2]))
459 |   data_test=MyDataset(FPatches, gPatches)
460 |   test_loader = torch.utils.data.DataLoader(data_test,batch_size=10,shuffle=False, num_workers=2)
461 |   with torch.no_grad():
462 |     correct = 0
463 |     total = 0
464 |     predicted_numpy=[]
465 |     for images, labels in test_loader:
466 |       images = images.to(device)
467 |       labels = labels.to(device)
468 |       outputs = model(images)
469 |       _, predicted = torch.max(outputs.data, 1)
470 |       predicted_numpy.append(predicted.cpu().numpy())
471 |       total += labels.size(0)
472 |       correct += (predicted == labels).sum().item()
473 |   classification_map=np.array(predicted_numpy)
474 |   cm=[]
475 |   for arr in classification_map:
476 |     cm.append(arr.tolist())
477 |   cm=list(itertools.chain.from_iterable(cm))
478 |   classification_map=np.array(cm)
479 | 
480 |   height=g.shape[0]
481 |   width=g.shape[1]
482 |   outputs = np.zeros((height,width))
483 |   k=0
484 |   for i in range(height):
485 |     for j in range(width):
486 |        target = g[i][j]
487 |        if target == 0 :
488 |          continue
489 |        else :
490 |         outputs[i][j]=classification_map[k]
491 |        k=k+1
492 |   return classification_map,outputs
493 | 
494 | cma,out=classified_pixels(FPatches,gPatches,g)
495 | 
496 | plt.figure(figsize=(7,7))
497 | a=plt.imshow(out)
498 | plt.savefig('KSC_pca_cmap')
499 | 
500 | plt.figure(figsize=(7,7))
501 | plt.imshow(g)
502 | plt.savefig('KSC_pca_gt')
503 | 


--------------------------------------------------------------------------------
/Hyperspectral Image Classification/HSI_cmap.png:
--------------------------------------------------------------------------------
https://raw.githubusercontent.com/dipuk0506/SpinalNet-Applications/08912f00345b1912476b6f62cd8a3b7823ed0fa5/Hyperspectral Image Classification/HSI_cmap.png


--------------------------------------------------------------------------------
/Hyperspectral Image Classification/HSI_gt.png:
--------------------------------------------------------------------------------
https://raw.githubusercontent.com/dipuk0506/SpinalNet-Applications/08912f00345b1912476b6f62cd8a3b7823ed0fa5/Hyperspectral Image Classification/HSI_gt.png


--------------------------------------------------------------------------------
/README.md:
--------------------------------------------------------------------------------
 1 | # [SpinalNet](https://github.com/dipuk0506/SpinalNet): New Applications and Improvements
 2 | 
 3 | ## Steps to share codes.
 4 | ### 1. Fork this repository.
 5 | ### 2. Upload codes, and other files as a folder in your repository. 
 6 | #### a. The folder name needs to be meaningful.
 7 | #### b. The folder name needs to be unique to avoid overwriting.
 8 | ### 3. Send us a pull request.
 9 | 
10 | #### To know about the process, please visit: [Collaborating with pull requests](https://docs.github.com/en/github/collaborating-with-pull-requests)
11 | 


--------------------------------------------------------------------------------