Poster Binary Model v5.ipynb

{
 "cells": [
  {
   "cell_type": "code",
   "execution_count": 1,
   "metadata": {},
   "outputs": [],
   "source": [
    "import numpy as np\n",
    "import pickle\n",
    "import matplotlib.pyplot as plt\n",
    "\n",
    "import torch\n",
    "import torch.nn as nn\n",
    "import torch.optim as optim\n",
    "from torch.utils.data import DataLoader\n",
    "from torch.utils.data import sampler\n",
    "\n",
    "import torchvision.datasets as dset\n",
    "import torchvision.transforms as T\n",
    "\n",
    "%load_ext autoreload\n",
    "%autoreload 2"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 2,
   "metadata": {},
   "outputs": [],
   "source": [
    "with open(\"data/pkl_files/trimmed_inputs.pkl\", \"rb\") as fp:\n",
    "    complete_inputs = pickle.load(fp)\n",
    "with open(\"data/pkl_files/trimmed_outputs.pkl\", \"rb\") as fp:\n",
    "    outputs = pickle.load(fp)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 3,
   "metadata": {},
   "outputs": [],
   "source": [
    "inputs = np.reshape(complete_inputs[:,0:3,:,:], (complete_inputs.shape[0], 3, complete_inputs.shape[2], complete_inputs.shape[3]))\n"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 4,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "(25650, 3, 32, 32)\n",
      "(25650, 1, 32, 32)\n"
     ]
    }
   ],
   "source": [
    "# Sanity check\n",
    "print(inputs.shape)\n",
    "print(outputs.shape)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 5,
   "metadata": {},
   "outputs": [],
   "source": [
    "outputs[outputs > 0] = 1"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "# Sanity Check\n",
    "idx = np.random.randint(0, inputs.shape[0])\n",
    "def plot_sampled_images(idx):    \n",
    "    cropped_seg_img = outputs[idx,0,:,:]\n",
    "    plt.imshow(cropped_seg_img)\n",
    "    plt.show()\n",
    "    cropped_flair_img = inputs[idx,0,:,:]\n",
    "    plt.imshow(cropped_flair_img)\n",
    "    plt.show()\n",
    "plot_sampled_images(idx)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 6,
   "metadata": {},
   "outputs": [],
   "source": [
    "np.random.seed(0)\n",
    "shuffled_idxs = list(range(inputs.shape[0]))\n",
    "np.random.shuffle(shuffled_idxs)\n",
    "train_cutoff = int(inputs.shape[0]*80/100)\n",
    "val_cutoff = int(inputs.shape[0]*90/100)\n",
    "\n",
    "train_inputs = []\n",
    "train_outputs = []\n",
    "val_inputs = []\n",
    "val_outputs = []\n",
    "test_inputs = []\n",
    "test_outputs = []\n",
    "\n",
    "for iteration, idx in enumerate(shuffled_idxs):\n",
    "    if iteration < train_cutoff:\n",
    "        train_inputs.append(inputs[idx])\n",
    "        train_outputs.append(outputs[idx])\n",
    "    elif iteration < val_cutoff:\n",
    "        val_inputs.append(inputs[idx])\n",
    "        val_outputs.append(outputs[idx])\n",
    "    else:\n",
    "        test_inputs.append(inputs[idx])\n",
    "        test_outputs.append(outputs[idx])"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 7,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "20520\n",
      "20520\n",
      "2565\n",
      "2565\n",
      "2565\n",
      "2565\n"
     ]
    }
   ],
   "source": [
    "# sanity check\n",
    "print(len(train_inputs))\n",
    "print(len(train_outputs))\n",
    "print(len(val_inputs))\n",
    "print(len(val_outputs))\n",
    "print(len(test_inputs))\n",
    "print(len(test_outputs))"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 8,
   "metadata": {},
   "outputs": [],
   "source": [
    "train_inputs = np.asarray(train_inputs).astype(np.float32)\n",
    "val_inputs = np.asarray(val_inputs).astype(np.float32)\n",
    "test_inputs = np.asarray(test_inputs).astype(np.float32)\n",
    "\n",
    "train_inputs_mean = np.mean(train_inputs, axis=0)\n",
    "train_inputs_std = np.std(train_inputs, axis=0)\n",
    "\n",
    "train_inputs -= train_inputs_mean\n",
    "train_inputs /= train_inputs_std\n",
    "val_inputs -= train_inputs_mean\n",
    "val_inputs /= train_inputs_std\n",
    "test_inputs -= train_inputs_mean\n",
    "test_inputs /= train_inputs_std"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 9,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "0.06759054618969298\n"
     ]
    }
   ],
   "source": [
    "train_outputs = np.asarray(train_outputs).astype(np.float32)\n",
    "n_tumor_pixels = np.count_nonzero(train_outputs)\n",
    "total_output_pixels = train_outputs.shape[0]*train_outputs.shape[2]*train_outputs.shape[3]\n",
    "frac_tumor = n_tumor_pixels / total_output_pixels\n",
    "print(frac_tumor)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 10,
   "metadata": {},
   "outputs": [],
   "source": [
    "train_data = []\n",
    "val_data = []\n",
    "test_data = []\n",
    "\n",
    "for idx in range(train_inputs.shape[0]):\n",
    "    train_data.append([train_inputs[idx], train_outputs[idx]])\n",
    "for idx in range(val_inputs.shape[0]):\n",
    "    val_data.append([val_inputs[idx], val_outputs[idx]])\n",
    "for idx in range(test_inputs.shape[0]):\n",
    "    test_data.append([test_inputs[idx], test_outputs[idx]])"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 11,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "642\n",
      "20520\n"
     ]
    }
   ],
   "source": [
    "train_loader = torch.utils.data.DataLoader(train_data, shuffle=True, batch_size=32)\n",
    "val_loader = torch.utils.data.DataLoader(val_data, shuffle=True, batch_size=32)\n",
    "test_loader = torch.utils.data.DataLoader(test_data, shuffle=True, batch_size=32)\n",
    "\n",
    "print(len(train_loader))\n",
    "print(len(train_data))"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 12,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "True\n",
      "using device: cuda\n"
     ]
    }
   ],
   "source": [
    "USE_GPU = True\n",
    "\n",
    "dtype = torch.float32 # we will be using float throughout this tutorial\n",
    "print(torch.cuda.is_available())\n",
    "\n",
    "if USE_GPU and torch.cuda.is_available():\n",
    "    device = torch.device('cuda')\n",
    "else:\n",
    "    device = torch.device('cpu')\n",
    "\n",
    "# Constant to control how frequently we print train loss\n",
    "print_every = 30\n",
    "\n",
    "print('using device:', device)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 13,
   "metadata": {},
   "outputs": [],
   "source": [
    "from models.unet import *"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 15,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "torch.Size([32, 2, 32, 32])\n"
     ]
    }
   ],
   "source": [
    "\"\"\"\n",
    "in_channels=1,\n",
    "        n_classes=2,\n",
    "        depth=5,\n",
    "        wf=6,\n",
    "        padding=False,\n",
    "        batch_norm=False,\n",
    "        up_mode='upconv',\n",
    "\"\"\"\n",
    "def test_UNet():\n",
    "    channels_in = 3\n",
    "    n_classes = 2\n",
    "    x = torch.zeros((32, 3, 32, 32), dtype=dtype)\n",
    "    model = UNet(channel_in=channels_in, classes=n_classes)\n",
    "    scores = model(x)\n",
    "    print(scores.size())\n",
    "test_UNet()"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 38,
   "metadata": {},
   "outputs": [],
   "source": [
    "# Defining all parameters \n",
    "print_every = 25\n",
    "loss_history = []\n",
    "learning_rate = 0.5e-2\n",
    "input_channels = 3\n",
    "n_classes = 2\n",
    "model = UNet(channel_in=input_channels, classes=n_classes)\n",
    "optimizer = optim.SGD(model.parameters(), lr=learning_rate, momentum=0.9, nesterov=True)\n",
    "# criterion = torch.nn.CrossEntropyLoss(weight=torch.FloatTensor([1.0, n_blank_pixel/float(n_tumor_pixels)]).cuda()) \n",
    "criterion = torch.nn.CrossEntropyLoss(weight=torch.FloatTensor([1.0, (total_output_pixels-n_tumor_pixels)/float(n_tumor_pixels)]).cuda()) \n",
    "# could try dice loss? not sure how it would work nor how it expects its parameters though lol\n",
    "\n",
    "if USE_GPU and torch.cuda.is_available():\n",
    "    model.cuda()"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 39,
   "metadata": {},
   "outputs": [],
   "source": [
    "def create_masks(x):\n",
    "    mask = np.zeros((n_classes, x.shape[0], x.shape[1]))\n",
    "    for c in range(n_classes):\n",
    "        mask[c] = x == c\n",
    "    return mask"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 40,
   "metadata": {},
   "outputs": [],
   "source": [
    "def compute_iou(scores, output):  \n",
    "    scores = scores.astype(np.bool)\n",
    "    output = output.astype(np.bool)\n",
    "    overlap = scores*output\n",
    "    union = scores+output # Logical OR\n",
    "    IOU = (overlap.sum())/(float(union.sum())+1e-10) \n",
    "    return IOU"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 41,
   "metadata": {},
   "outputs": [],
   "source": [
    "def iou_metric(preds, labels):\n",
    "    # preds and labels are of size (64, 32, 32)\n",
    "    # return average iou value for each class\n",
    "    numpy_preds = preds.cpu().numpy()  \n",
    "    numpy_labels = labels.cpu().numpy()\n",
    "    \n",
    "    iou_tracker = np.zeros((numpy_preds.shape[0], n_classes))\n",
    "    # get the masks\n",
    "    for idx in range(numpy_preds.shape[0]):\n",
    "        \"\"\"plt.title('Segmented Output')\n",
    "        plt.imshow(numpy_preds[idx,:,:])\n",
    "        plt.show()\n",
    "        plt.title('Ground Truth')\n",
    "        plt.imshow(numpy_labels[idx,:,:])\n",
    "        plt.show()\"\"\"\n",
    "        pred_mask = create_masks(numpy_preds[idx])\n",
    "        label_mask = create_masks(numpy_labels[idx])\n",
    "        for c in range(n_classes):\n",
    "            iou_tracker[idx, c] = compute_iou(pred_mask[c], label_mask[c])\n",
    "        # print(iou_tracker[idx])\n",
    "    # print(np.mean(iou_tracker, axis=0))\n",
    "    return np.mean(iou_tracker, axis=0)     "
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 42,
   "metadata": {},
   "outputs": [],
   "source": [
    "def check_iou(loader, model):\n",
    "    all_iou = np.zeros((len(loader),n_classes))\n",
    "\n",
    "    with torch.no_grad():\n",
    "        for t, (x,y) in enumerate(loader):\n",
    "            x = x.to(device=device, dtype=dtype)  # move to device, e.g. GPU\n",
    "            y = y.to(device=device, dtype=torch.float).view(y.shape[0], 32, 32)\n",
    "            scores = model(x)\n",
    "            _, preds = scores.max(1)\n",
    "            \n",
    "            all_iou[t] = iou_metric(preds, y)\n",
    "    # print(np.mean(all_iou, axis=0))\n",
    "    return np.mean(all_iou, axis=0)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 43,
   "metadata": {},
   "outputs": [
    {
     "name": "stderr",
     "output_type": "stream",
     "text": [
      "/opt/anaconda3/lib/python3.7/site-packages/torch/nn/functional.py:2539: UserWarning: Default upsampling behavior when mode=bilinear is changed to align_corners=False since 0.4.0. Please specify align_corners=True if the old behavior is desired. See the documentation of nn.Upsample for details.\n",
      "  \"See the documentation of nn.Upsample for details.\".format(mode))\n"
     ]
    },
    {
     "data": {
      "text/plain": [
       "array([0.        , 0.06215225])"
      ]
     },
     "execution_count": 43,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "check_iou(val_loader, model)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 44,
   "metadata": {},
   "outputs": [],
   "source": [
    "def plot_output(scores, y):\n",
    "    with torch.no_grad():\n",
    "        _, preds = scores.max(1)\n",
    "        numpy_scores = preds.cpu().numpy().astype(np.uint8)\n",
    "        numpy_truth = y.cpu().numpy().astype(np.uint8)\n",
    "        idx = np.random.randint(1, y.shape[0])\n",
    "        while(np.count_nonzero(numpy_truth[idx]) < int(0.10*32*32)):\n",
    "            idx = np.random.randint(1, y.shape[0])\n",
    "        plt.title('Segmented Output')\n",
    "        plt.imshow(numpy_scores[idx,:,:])\n",
    "        plt.show()\n",
    "        plt.title('Ground Truth')\n",
    "        plt.imshow(numpy_truth[idx,:,:])\n",
    "        plt.show()"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 45,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "Epoch 0, Iteration 0, loss = 0.7237\n",
      "Validation IOU:  [0.         0.06310033]\n",
      "Training IOU:  [0.         0.06763351]\n",
      "Epoch 0, Iteration 25, loss = 0.6328\n",
      "Validation IOU:  [0.         0.06215225]\n",
      "Training IOU:  [0.         0.06751159]\n",
      "Epoch 0, Iteration 50, loss = 0.7106\n",
      "Validation IOU:  [0.         0.06215225]\n",
      "Training IOU:  [0.         0.06761369]\n",
      "Epoch 0, Iteration 75, loss = 0.6669\n",
      "Validation IOU:  [0.02786961 0.06252606]\n",
      "Training IOU:  [0.02713037 0.06805595]\n",
      "Epoch 0, Iteration 100, loss = 0.7450\n",
      "Validation IOU:  [0.16471436 0.06592142]\n",
      "Training IOU:  [0.16547081 0.07130487]\n",
      "Epoch 0, Iteration 125, loss = 0.6746\n",
      "Validation IOU:  [0.23758264 0.06690081]\n",
      "Training IOU:  [0.24006084 0.07280316]\n",
      "Epoch 0, Iteration 150, loss = 0.7156\n",
      "Validation IOU:  [0.32428283 0.06468786]\n",
      "Training IOU:  [0.32447414 0.07115551]\n",
      "Epoch 0, Iteration 175, loss = 0.6709\n",
      "Validation IOU:  [0.52386132 0.05510607]\n",
      "Training IOU:  [0.52159991 0.05976416]\n",
      "Epoch 0, Iteration 200, loss = 0.6580\n",
      "Validation IOU:  [0.67426421 0.0462375 ]\n",
      "Training IOU:  [0.67126494 0.05072684]\n",
      "Epoch 0, Iteration 225, loss = 0.6738\n",
      "Validation IOU:  [0.72757282 0.03964026]\n",
      "Training IOU:  [0.72714604 0.04493176]\n",
      "Epoch 0, Iteration 250, loss = 0.6585\n",
      "Validation IOU:  [0.74421767 0.04135651]\n",
      "Training IOU:  [0.74238878 0.0456546 ]\n",
      "Epoch 0, Iteration 275, loss = 0.6371\n",
      "Validation IOU:  [0.77483428 0.03789676]\n",
      "Training IOU:  [0.77360635 0.04270625]\n",
      "Epoch 0, Iteration 300, loss = 0.6373\n",
      "Validation IOU:  [0.76332614 0.03926626]\n",
      "Training IOU:  [0.76410802 0.04515966]\n",
      "Epoch 0, Iteration 325, loss = 0.6761\n",
      "Validation IOU:  [0.78290609 0.03880052]\n",
      "Training IOU:  [0.78276271 0.04328344]\n",
      "Epoch 0, Iteration 350, loss = 0.5994\n",
      "Validation IOU:  [0.73761682 0.04570307]\n",
      "Training IOU:  [0.73735645 0.05219038]\n",
      "Epoch 0, Iteration 375, loss = 0.6349\n",
      "Validation IOU:  [0.71058316 0.0506186 ]\n",
      "Training IOU:  [0.71282465 0.05693713]\n",
      "Epoch 0, Iteration 400, loss = 0.5989\n",
      "Validation IOU:  [0.69856594 0.05248944]\n",
      "Training IOU:  [0.69906742 0.05938101]\n",
      "Epoch 0, Iteration 425, loss = 0.6303\n",
      "Validation IOU:  [0.72776427 0.04959493]\n",
      "Training IOU:  [0.73145634 0.055973  ]\n",
      "Epoch 0, Iteration 450, loss = 0.6130\n",
      "Validation IOU:  [0.74333505 0.05236629]\n",
      "Training IOU:  [0.74425683 0.05705412]\n",
      "Epoch 0, Iteration 475, loss = 0.6137\n",
      "Validation IOU:  [0.68500121 0.0530295 ]\n",
      "Training IOU:  [0.68769742 0.05973121]\n",
      "Epoch 0, Iteration 500, loss = 0.6024\n",
      "Validation IOU:  [0.65466167 0.05560491]\n",
      "Training IOU:  [0.66017975 0.06309134]\n",
      "Epoch 0, Iteration 525, loss = 0.6356\n",
      "Validation IOU:  [0.68938498 0.05900878]\n",
      "Training IOU:  [0.69268131 0.06580866]\n",
      "Epoch 0, Iteration 550, loss = 0.6345\n",
      "Validation IOU:  [0.61818259 0.06239416]\n",
      "Training IOU:  [0.62073513 0.06994154]\n",
      "Epoch 0, Iteration 575, loss = 0.6332\n",
      "Validation IOU:  [0.65745321 0.06031415]\n",
      "Training IOU:  [0.65938857 0.06629283]\n",
      "Epoch 0, Iteration 600, loss = 0.5260\n",
      "Validation IOU:  [0.70067656 0.06054958]\n",
      "Training IOU:  [0.7048448  0.06600683]\n",
      "Epoch 0, Iteration 625, loss = 0.6306\n",
      "Validation IOU:  [0.71542233 0.05810368]\n",
      "Training IOU:  [0.71938424 0.06632798]\n",
      "Epoch 1, Iteration 0, loss = 0.5753\n",
      "Validation IOU:  [0.64281406 0.06616   ]\n",
      "Training IOU:  [0.64198933 0.07195577]\n",
      "Epoch 1, Iteration 25, loss = 0.5735\n",
      "Validation IOU:  [0.6765713  0.06416857]\n",
      "Training IOU:  [0.68032586 0.07095346]\n",
      "Epoch 1, Iteration 50, loss = 0.5278\n",
      "Validation IOU:  [0.69079309 0.06322497]\n",
      "Training IOU:  [0.69509319 0.07006326]\n",
      "Epoch 1, Iteration 75, loss = 0.5228\n",
      "Validation IOU:  [0.70780944 0.06272019]\n",
      "Training IOU:  [0.70946336 0.06846599]\n",
      "Epoch 1, Iteration 100, loss = 0.5699\n",
      "Validation IOU:  [0.67423639 0.06455716]\n",
      "Training IOU:  [0.67591473 0.07125355]\n",
      "Epoch 1, Iteration 125, loss = 0.6209\n",
      "Validation IOU:  [0.6799263  0.06789975]\n",
      "Training IOU:  [0.68236269 0.07300968]\n",
      "Epoch 1, Iteration 150, loss = 0.5935\n",
      "Validation IOU:  [0.69346503 0.06386447]\n",
      "Training IOU:  [0.70069547 0.06973693]\n",
      "Epoch 1, Iteration 175, loss = 0.6194\n",
      "Validation IOU:  [0.70253845 0.0652285 ]\n",
      "Training IOU:  [0.70723652 0.0705541 ]\n",
      "Epoch 1, Iteration 200, loss = 0.5518\n",
      "Validation IOU:  [0.70593477 0.06315567]\n",
      "Training IOU:  [0.71246271 0.06865881]\n",
      "Epoch 1, Iteration 225, loss = 0.6254\n",
      "Validation IOU:  [0.70537325 0.06174252]\n",
      "Training IOU:  [0.71156893 0.0682509 ]\n",
      "Epoch 1, Iteration 250, loss = 0.5507\n",
      "Validation IOU:  [0.69621526 0.06550587]\n",
      "Training IOU:  [0.6969142  0.07002329]\n",
      "Epoch 1, Iteration 275, loss = 0.5645\n",
      "Validation IOU:  [0.7306615  0.06187315]\n",
      "Training IOU:  [0.73595749 0.06855532]\n",
      "Epoch 1, Iteration 300, loss = 0.5636\n",
      "Validation IOU:  [0.72943452 0.06416602]\n",
      "Training IOU:  [0.73788044 0.07122936]\n",
      "Epoch 1, Iteration 325, loss = 0.5048\n",
      "Validation IOU:  [0.74148976 0.06410516]\n",
      "Training IOU:  [0.74525982 0.07135672]\n",
      "Epoch 1, Iteration 350, loss = 0.5753\n",
      "Validation IOU:  [0.72832206 0.06432391]\n",
      "Training IOU:  [0.73274633 0.07155364]\n",
      "Epoch 1, Iteration 375, loss = 0.5431\n",
      "Validation IOU:  [0.70433007 0.06711311]\n",
      "Training IOU:  [0.70543064 0.07283557]\n",
      "Epoch 1, Iteration 400, loss = 0.5443\n",
      "Validation IOU:  [0.71735032 0.0689233 ]\n",
      "Training IOU:  [0.72259785 0.07523424]\n",
      "Epoch 1, Iteration 425, loss = 0.5392\n",
      "Validation IOU:  [0.6745944  0.06783366]\n",
      "Training IOU:  [0.67534402 0.07366745]\n",
      "Epoch 1, Iteration 450, loss = 0.6560\n",
      "Validation IOU:  [0.71199613 0.06623448]\n",
      "Training IOU:  [0.7150828  0.07347082]\n",
      "Epoch 1, Iteration 475, loss = 0.5687\n",
      "Validation IOU:  [0.70135612 0.06783674]\n",
      "Training IOU:  [0.70731291 0.07491842]\n",
      "Epoch 1, Iteration 500, loss = 0.5399\n",
      "Validation IOU:  [0.68105208 0.06942278]\n",
      "Training IOU:  [0.68804679 0.07551768]\n",
      "Epoch 1, Iteration 525, loss = 0.4264\n",
      "Validation IOU:  [0.71353122 0.07016994]\n",
      "Training IOU:  [0.71616392 0.07460377]\n",
      "Epoch 1, Iteration 550, loss = 0.6121\n",
      "Validation IOU:  [0.71335415 0.06666815]\n",
      "Training IOU:  [0.71431185 0.07375472]\n",
      "Epoch 1, Iteration 575, loss = 0.6393\n",
      "Validation IOU:  [0.73655539 0.07121308]\n",
      "Training IOU:  [0.7369726  0.07362426]\n",
      "Epoch 1, Iteration 600, loss = 0.4576\n",
      "Validation IOU:  [0.7275956  0.06644101]\n",
      "Training IOU:  [0.7326966  0.07390982]\n",
      "Epoch 1, Iteration 625, loss = 0.5362\n",
      "Validation IOU:  [0.73732946 0.06753784]\n",
      "Training IOU:  [0.74079698 0.07247941]\n",
      "Epoch 2, Iteration 0, loss = 0.5555\n",
      "Validation IOU:  [0.72521337 0.06555191]\n",
      "Training IOU:  [0.72841699 0.07362122]\n",
      "Epoch 2, Iteration 25, loss = 0.5911\n",
      "Validation IOU:  [0.71703548 0.06580055]\n",
      "Training IOU:  [0.71894607 0.07370167]\n",
      "Epoch 2, Iteration 50, loss = 0.5354\n",
      "Validation IOU:  [0.68894828 0.06806464]\n",
      "Training IOU:  [0.69410471 0.07553487]\n",
      "Epoch 2, Iteration 75, loss = 0.4444\n",
      "Validation IOU:  [0.70750735 0.06952713]\n",
      "Training IOU:  [0.71118922 0.07630516]\n",
      "Epoch 2, Iteration 100, loss = 0.5661\n",
      "Validation IOU:  [0.73363278 0.06809052]\n",
      "Training IOU:  [0.73548706 0.07136933]\n",
      "Epoch 2, Iteration 125, loss = 0.5288\n",
      "Validation IOU:  [0.73159394 0.0651806 ]\n",
      "Training IOU:  [0.73347173 0.06964834]\n",
      "Epoch 2, Iteration 150, loss = 0.4929\n",
      "Validation IOU:  [0.70256579 0.06316564]\n",
      "Training IOU:  [0.70213976 0.07206378]\n",
      "Epoch 2, Iteration 175, loss = 0.4907\n",
      "Validation IOU:  [0.71988545 0.06912078]\n",
      "Training IOU:  [0.72292212 0.07509531]\n",
      "Epoch 2, Iteration 200, loss = 0.4802\n",
      "Validation IOU:  [0.68679672 0.06794858]\n",
      "Training IOU:  [0.68756161 0.07603439]\n",
      "Epoch 2, Iteration 225, loss = 0.5186\n"
     ]
    },
    {
     "ename": "KeyboardInterrupt",
     "evalue": "",
     "output_type": "error",
     "traceback": [
      "\u001b[0;31m---------------------------------------------------------------------------\u001b[0m",
      "\u001b[0;31mKeyboardInterrupt\u001b[0m                         Traceback (most recent call last)",
      "\u001b[0;32m<ipython-input-45-1df2b8a587bb>\u001b[0m in \u001b[0;36m<module>\u001b[0;34m\u001b[0m\n\u001b[1;32m     36\u001b[0m             \u001b[0;31m# if t % plot_every == 0:\u001b[0m\u001b[0;34m\u001b[0m\u001b[0;34m\u001b[0m\u001b[0;34m\u001b[0m\u001b[0m\n\u001b[1;32m     37\u001b[0m                 \u001b[0;31m# plot_output(scores, labels)\u001b[0m\u001b[0;34m\u001b[0m\u001b[0;34m\u001b[0m\u001b[0;34m\u001b[0m\u001b[0m\n\u001b[0;32m---> 38\u001b[0;31m \u001b[0mtrain\u001b[0m\u001b[0;34m(\u001b[0m\u001b[0mmodel\u001b[0m\u001b[0;34m,\u001b[0m \u001b[0moptimizer\u001b[0m\u001b[0;34m,\u001b[0m \u001b[0mepochs\u001b[0m\u001b[0;34m=\u001b[0m\u001b[0;36m250\u001b[0m\u001b[0;34m)\u001b[0m\u001b[0;34m\u001b[0m\u001b[0;34m\u001b[0m\u001b[0m\n\u001b[0m",
      "\u001b[0;32m<ipython-input-45-1df2b8a587bb>\u001b[0m in \u001b[0;36mtrain\u001b[0;34m(model, optimizer, epochs)\u001b[0m\n\u001b[1;32m     30\u001b[0m                 \u001b[0mval_iou\u001b[0m \u001b[0;34m=\u001b[0m \u001b[0mcheck_iou\u001b[0m\u001b[0;34m(\u001b[0m\u001b[0mval_loader\u001b[0m\u001b[0;34m,\u001b[0m \u001b[0mmodel\u001b[0m\u001b[0;34m)\u001b[0m\u001b[0;34m\u001b[0m\u001b[0;34m\u001b[0m\u001b[0m\n\u001b[1;32m     31\u001b[0m                 \u001b[0mval_iou_history\u001b[0m\u001b[0;34m.\u001b[0m\u001b[0mappend\u001b[0m\u001b[0;34m(\u001b[0m\u001b[0mval_iou\u001b[0m\u001b[0;34m)\u001b[0m\u001b[0;34m\u001b[0m\u001b[0;34m\u001b[0m\u001b[0m\n\u001b[0;32m---> 32\u001b[0;31m                 \u001b[0mtrain_iou\u001b[0m \u001b[0;34m=\u001b[0m \u001b[0mcheck_iou\u001b[0m\u001b[0;34m(\u001b[0m\u001b[0mtrain_loader\u001b[0m\u001b[0;34m,\u001b[0m \u001b[0mmodel\u001b[0m\u001b[0;34m)\u001b[0m\u001b[0;34m\u001b[0m\u001b[0;34m\u001b[0m\u001b[0m\n\u001b[0m\u001b[1;32m     33\u001b[0m                 \u001b[0mtrain_iou_history\u001b[0m\u001b[0;34m.\u001b[0m\u001b[0mappend\u001b[0m\u001b[0;34m(\u001b[0m\u001b[0mtrain_iou\u001b[0m\u001b[0;34m)\u001b[0m\u001b[0;34m\u001b[0m\u001b[0;34m\u001b[0m\u001b[0m\n\u001b[1;32m     34\u001b[0m                 \u001b[0mprint\u001b[0m\u001b[0;34m(\u001b[0m\u001b[0;34m'Validation IOU: '\u001b[0m\u001b[0;34m,\u001b[0m \u001b[0mval_iou\u001b[0m\u001b[0;34m)\u001b[0m\u001b[0;34m\u001b[0m\u001b[0;34m\u001b[0m\u001b[0m\n",
      "\u001b[0;32m<ipython-input-42-1efe0f204942>\u001b[0m in \u001b[0;36mcheck_iou\u001b[0;34m(loader, model)\u001b[0m\n\u001b[1;32m      9\u001b[0m             \u001b[0m_\u001b[0m\u001b[0;34m,\u001b[0m \u001b[0mpreds\u001b[0m \u001b[0;34m=\u001b[0m \u001b[0mscores\u001b[0m\u001b[0;34m.\u001b[0m\u001b[0mmax\u001b[0m\u001b[0;34m(\u001b[0m\u001b[0;36m1\u001b[0m\u001b[0;34m)\u001b[0m\u001b[0;34m\u001b[0m\u001b[0;34m\u001b[0m\u001b[0m\n\u001b[1;32m     10\u001b[0m \u001b[0;34m\u001b[0m\u001b[0m\n\u001b[0;32m---> 11\u001b[0;31m             \u001b[0mall_iou\u001b[0m\u001b[0;34m[\u001b[0m\u001b[0mt\u001b[0m\u001b[0;34m]\u001b[0m \u001b[0;34m=\u001b[0m \u001b[0miou_metric\u001b[0m\u001b[0;34m(\u001b[0m\u001b[0mpreds\u001b[0m\u001b[0;34m,\u001b[0m \u001b[0my\u001b[0m\u001b[0;34m)\u001b[0m\u001b[0;34m\u001b[0m\u001b[0;34m\u001b[0m\u001b[0m\n\u001b[0m\u001b[1;32m     12\u001b[0m     \u001b[0;31m# print(np.mean(all_iou, axis=0))\u001b[0m\u001b[0;34m\u001b[0m\u001b[0;34m\u001b[0m\u001b[0;34m\u001b[0m\u001b[0m\n\u001b[1;32m     13\u001b[0m     \u001b[0;32mreturn\u001b[0m \u001b[0mnp\u001b[0m\u001b[0;34m.\u001b[0m\u001b[0mmean\u001b[0m\u001b[0;34m(\u001b[0m\u001b[0mall_iou\u001b[0m\u001b[0;34m,\u001b[0m \u001b[0maxis\u001b[0m\u001b[0;34m=\u001b[0m\u001b[0;36m0\u001b[0m\u001b[0;34m)\u001b[0m\u001b[0;34m\u001b[0m\u001b[0;34m\u001b[0m\u001b[0m\n",
      "\u001b[0;32m<ipython-input-41-f89f8357680b>\u001b[0m in \u001b[0;36miou_metric\u001b[0;34m(preds, labels)\u001b[0m\n\u001b[1;32m     17\u001b[0m         \u001b[0mlabel_mask\u001b[0m \u001b[0;34m=\u001b[0m \u001b[0mcreate_masks\u001b[0m\u001b[0;34m(\u001b[0m\u001b[0mnumpy_labels\u001b[0m\u001b[0;34m[\u001b[0m\u001b[0midx\u001b[0m\u001b[0;34m]\u001b[0m\u001b[0;34m)\u001b[0m\u001b[0;34m\u001b[0m\u001b[0;34m\u001b[0m\u001b[0m\n\u001b[1;32m     18\u001b[0m         \u001b[0;32mfor\u001b[0m \u001b[0mc\u001b[0m \u001b[0;32min\u001b[0m \u001b[0mrange\u001b[0m\u001b[0;34m(\u001b[0m\u001b[0mn_classes\u001b[0m\u001b[0;34m)\u001b[0m\u001b[0;34m:\u001b[0m\u001b[0;34m\u001b[0m\u001b[0;34m\u001b[0m\u001b[0m\n\u001b[0;32m---> 19\u001b[0;31m             \u001b[0miou_tracker\u001b[0m\u001b[0;34m[\u001b[0m\u001b[0midx\u001b[0m\u001b[0;34m,\u001b[0m \u001b[0mc\u001b[0m\u001b[0;34m]\u001b[0m \u001b[0;34m=\u001b[0m \u001b[0mcompute_iou\u001b[0m\u001b[0;34m(\u001b[0m\u001b[0mpred_mask\u001b[0m\u001b[0;34m[\u001b[0m\u001b[0mc\u001b[0m\u001b[0;34m]\u001b[0m\u001b[0;34m,\u001b[0m \u001b[0mlabel_mask\u001b[0m\u001b[0;34m[\u001b[0m\u001b[0mc\u001b[0m\u001b[0;34m]\u001b[0m\u001b[0;34m)\u001b[0m\u001b[0;34m\u001b[0m\u001b[0;34m\u001b[0m\u001b[0m\n\u001b[0m\u001b[1;32m     20\u001b[0m         \u001b[0;31m# print(iou_tracker[idx])\u001b[0m\u001b[0;34m\u001b[0m\u001b[0;34m\u001b[0m\u001b[0;34m\u001b[0m\u001b[0m\n\u001b[1;32m     21\u001b[0m     \u001b[0;31m# print(np.mean(iou_tracker, axis=0))\u001b[0m\u001b[0;34m\u001b[0m\u001b[0;34m\u001b[0m\u001b[0;34m\u001b[0m\u001b[0m\n",
      "\u001b[0;32m<ipython-input-40-a0fbd233bce6>\u001b[0m in \u001b[0;36mcompute_iou\u001b[0;34m(scores, output)\u001b[0m\n\u001b[1;32m      4\u001b[0m     \u001b[0moverlap\u001b[0m \u001b[0;34m=\u001b[0m \u001b[0mscores\u001b[0m\u001b[0;34m*\u001b[0m\u001b[0moutput\u001b[0m\u001b[0;34m\u001b[0m\u001b[0;34m\u001b[0m\u001b[0m\n\u001b[1;32m      5\u001b[0m     \u001b[0munion\u001b[0m \u001b[0;34m=\u001b[0m \u001b[0mscores\u001b[0m\u001b[0;34m+\u001b[0m\u001b[0moutput\u001b[0m \u001b[0;31m# Logical OR\u001b[0m\u001b[0;34m\u001b[0m\u001b[0;34m\u001b[0m\u001b[0m\n\u001b[0;32m----> 6\u001b[0;31m     \u001b[0mIOU\u001b[0m \u001b[0;34m=\u001b[0m \u001b[0;34m(\u001b[0m\u001b[0moverlap\u001b[0m\u001b[0;34m.\u001b[0m\u001b[0msum\u001b[0m\u001b[0;34m(\u001b[0m\u001b[0;34m)\u001b[0m\u001b[0;34m)\u001b[0m\u001b[0;34m/\u001b[0m\u001b[0;34m(\u001b[0m\u001b[0mfloat\u001b[0m\u001b[0;34m(\u001b[0m\u001b[0munion\u001b[0m\u001b[0;34m.\u001b[0m\u001b[0msum\u001b[0m\u001b[0;34m(\u001b[0m\u001b[0;34m)\u001b[0m\u001b[0;34m)\u001b[0m\u001b[0;34m+\u001b[0m\u001b[0;36m1e-10\u001b[0m\u001b[0;34m)\u001b[0m\u001b[0;34m\u001b[0m\u001b[0;34m\u001b[0m\u001b[0m\n\u001b[0m\u001b[1;32m      7\u001b[0m     \u001b[0;32mreturn\u001b[0m \u001b[0mIOU\u001b[0m\u001b[0;34m\u001b[0m\u001b[0;34m\u001b[0m\u001b[0m\n",
      "\u001b[0;32m/opt/anaconda3/lib/python3.7/site-packages/numpy/core/_methods.py\u001b[0m in \u001b[0;36m_sum\u001b[0;34m(a, axis, dtype, out, keepdims, initial)\u001b[0m\n\u001b[1;32m     34\u001b[0m def _sum(a, axis=None, dtype=None, out=None, keepdims=False,\n\u001b[1;32m     35\u001b[0m          initial=_NoValue):\n\u001b[0;32m---> 36\u001b[0;31m     \u001b[0;32mreturn\u001b[0m \u001b[0mumr_sum\u001b[0m\u001b[0;34m(\u001b[0m\u001b[0ma\u001b[0m\u001b[0;34m,\u001b[0m \u001b[0maxis\u001b[0m\u001b[0;34m,\u001b[0m \u001b[0mdtype\u001b[0m\u001b[0;34m,\u001b[0m \u001b[0mout\u001b[0m\u001b[0;34m,\u001b[0m \u001b[0mkeepdims\u001b[0m\u001b[0;34m,\u001b[0m \u001b[0minitial\u001b[0m\u001b[0;34m)\u001b[0m\u001b[0;34m\u001b[0m\u001b[0;34m\u001b[0m\u001b[0m\n\u001b[0m\u001b[1;32m     37\u001b[0m \u001b[0;34m\u001b[0m\u001b[0m\n\u001b[1;32m     38\u001b[0m def _prod(a, axis=None, dtype=None, out=None, keepdims=False,\n",
      "\u001b[0;31mKeyboardInterrupt\u001b[0m: "
     ]
    }
   ],
   "source": [
    "import torch.nn.functional as F\n",
    "\n",
    "val_iou_history = []\n",
    "train_iou_history = []\n",
    "plot_every = 100\n",
    "def train(model, optimizer, epochs=1):\n",
    "    model = model.to(device=device)\n",
    "    loss_history.clear()\n",
    "    val_iou_history.clear()\n",
    "    train_iou_history.clear()\n",
    "    for e in range(epochs):\n",
    "        for t, (x,y) in enumerate(train_loader):\n",
    "            model.train()\n",
    "            x = x.to(device=device, dtype=dtype)\n",
    "            labels = y.to(device=device, dtype=torch.long).view(y.shape[0], 32, 32)\n",
    "            # print(labels.shape)\n",
    "\n",
    "            scores = model(x)\n",
    "            # print(scores.shape)\n",
    "\n",
    "            loss = criterion(scores, labels)\n",
    "            # loss = F.cross_entropy(scores, labels, weight=torch.FloatTensor([1/(float(26265600-1762336)), 1/(float(1762336))]).cuda())\n",
    "            loss_history.append(loss.item())\n",
    "            optimizer.zero_grad()\n",
    "            loss.backward()\n",
    "            optimizer.step()\n",
    "            # print(t)\n",
    "            if t % print_every == 0:\n",
    "                print('Epoch %d, Iteration %d, loss = %.4f' % (e, t, loss.item()))\n",
    "                val_iou = check_iou(val_loader, model)\n",
    "                val_iou_history.append(val_iou)\n",
    "                train_iou = check_iou(train_loader, model)\n",
    "                train_iou_history.append(train_iou)\n",
    "                print('Validation IOU: ', val_iou)\n",
    "                print('Training IOU: ', train_iou)\n",
    "            # if t % plot_every == 0:\n",
    "                # plot_output(scores, labels)\n",
    "train(model, optimizer, epochs=250)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 50,
   "metadata": {
    "scrolled": true
   },
   "outputs": [
    {
     "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": [
    "fig = plt.figure()\n",
    "plt.title('Loss over time')\n",
    "plt.xlabel('Iteration')\n",
    "plt.ylabel('Loss')\n",
    "plt.plot(range(len(loss_history))[0::225], loss_history[0::225])\n",
    "# plt.plot(range(len(loss_history)), loss_history)\n",
    "fig.savefig('Binary Problem Loss - Poster Session.png')"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 25,
   "metadata": {},
   "outputs": [],
   "source": [
    "val_hist = np.asarray(val_iou_history)\n",
    "train_hist = np.asarray(train_iou_history)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 26,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "(2299, 2)\n",
      "(2298, 2)\n"
     ]
    }
   ],
   "source": [
    "print(val_hist.shape)\n",
    "print(train_hist.shape)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 33,
   "metadata": {},
   "outputs": [
    {
     "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": [
    "fig = plt.figure()\n",
    "plt.title('IOU for tumor over time')\n",
    "plt.xlabel('Iteration')\n",
    "plt.ylabel('IOU')\n",
    "plt.plot(range(val_hist.shape[0]), val_hist[:,1], label='val')\n",
    "plt.plot(range(train_hist.shape[0]), train_hist[:,1], label='train')\n",
    "plt.legend()\n",
    "fig.savefig('IOU for tumor - Poster Session.png')"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 34,
   "metadata": {},
   "outputs": [
    {
     "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": [
    "fig = plt.figure()\n",
    "plt.title('IOU for no tumor over time')\n",
    "plt.xlabel('Iteration')\n",
    "plt.ylabel('IOU')\n",
    "plt.plot(range(val_hist.shape[0]), val_hist[:,0], label='val')\n",
    "plt.plot(range(train_hist.shape[0]), train_hist[:,0], label='train')\n",
    "plt.legend()\n",
    "fig.savefig('IOU for no tumor - Poster Session.png')"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 32,
   "metadata": {},
   "outputs": [
    {
     "name": "stderr",
     "output_type": "stream",
     "text": [
      "/opt/anaconda3/lib/python3.7/site-packages/torch/nn/functional.py:2539: UserWarning: Default upsampling behavior when mode=bilinear is changed to align_corners=False since 0.4.0. Please specify align_corners=True if the old behavior is desired. See the documentation of nn.Upsample for details.\n",
      "  \"See the documentation of nn.Upsample for details.\".format(mode))\n"
     ]
    },
    {
     "data": {
      "text/plain": [
       "array([0.88586248, 0.06868494])"
      ]
     },
     "execution_count": 32,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "check_iou(test_loader, model)"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "Overfitting small dataset"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "small_inputs = inputs[:40]\n",
    "small_outputs = outputs[:40]\n",
    "\n",
    "np.random.seed(0)\n",
    "shuffled_idxs = list(range(small_inputs.shape[0]))\n",
    "np.random.shuffle(shuffled_idxs)\n",
    "train_cutoff = int(small_inputs.shape[0]*80/100)\n",
    "\n",
    "small_train_inputs = []\n",
    "small_train_outputs = []\n",
    "small_val_inputs = []\n",
    "small_val_outputs = []\n",
    "\n",
    "\n",
    "for iteration, idx in enumerate(shuffled_idxs):\n",
    "    if iteration < train_cutoff:\n",
    "        small_train_inputs.append(small_inputs[idx])\n",
    "        small_train_outputs.append(small_outputs[idx])\n",
    "    else:\n",
    "        small_val_inputs.append(small_inputs[idx])\n",
    "        small_val_outputs.append(small_outputs[idx])\n",
    "\n",
    "n_tumor_pixels = np.count_nonzero(np.asarray(small_train_outputs))\n",
    "total_output_pixels = 32*32*32\n",
    "frac_tumor = n_tumor_pixels / total_output_pixels\n"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "small_train_inputs = np.asarray(small_train_inputs).astype(np.float32)\n",
    "small_val_inputs = np.asarray(small_val_inputs).astype(np.float32)\n",
    "\n",
    "small_train_inputs_mean = np.mean(small_train_inputs, axis=0)\n",
    "small_train_inputs_std = np.std(small_train_inputs, axis=0)\n",
    "\n",
    "small_train_inputs -= small_train_inputs_mean\n",
    "small_train_inputs /= small_train_inputs_std\n",
    "small_val_inputs -= small_train_inputs_mean\n",
    "small_val_inputs /= small_train_inputs_std\n"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "small_train_data = []\n",
    "small_val_data = []\n",
    "test_data = []\n",
    "\n",
    "for idx in range(small_train_inputs.shape[0]):\n",
    "    small_train_data.append([small_train_inputs[idx], small_train_outputs[idx]])\n",
    "for idx in range(small_val_inputs.shape[0]):\n",
    "    small_val_data.append([small_val_inputs[idx], small_val_outputs[idx]])"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "print(len(small_train_data))\n",
    "print(len(small_val_data))"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "small_train_loader = torch.utils.data.DataLoader(small_train_data, shuffle=True, batch_size=32)\n",
    "small_val_loader = torch.utils.data.DataLoader(small_val_data, shuffle=True, batch_size=32)\n"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "print_every = 25\n",
    "learning_rate = 1e-1\n",
    "input_channels = 3\n",
    "n_classes = 2\n",
    "model = UNet(channel_in=input_channels, classes=n_classes)\n",
    "optimizer = optim.SGD(model.parameters(), lr=learning_rate, momentum=0.9, nesterov=True)\n",
    "# criterion = torch.nn.CrossEntropyLoss(weight=torch.FloatTensor([1.0, n_blank_pixel/float(n_tumor_pixels)]).cuda()) \n",
    "# criterion = torch.nn.CrossEntropyLoss(weight=torch.FloatTensor([1.0, 35]).cuda()) \n",
    "criterion = torch.nn.CrossEntropyLoss(weight=torch.FloatTensor([1.0, (total_output_pixels-n_tumor_pixels)/float(n_tumor_pixels)]).cuda()) \n",
    "\n",
    "# could try dice loss? not sure how it would work nor how it expects its parameters though lol"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "val_iou_history = []\n",
    "loss_history = []\n",
    "train_iou_history = []\n",
    "plot_every = 100\n",
    "def train(model, optimizer, epochs=1):\n",
    "    model = model.to(device=device)\n",
    "    loss_history.clear()\n",
    "    val_iou_history.clear()\n",
    "    train_iou_history.clear()\n",
    "    for e in range(epochs):\n",
    "        for t, (x,y) in enumerate(small_train_loader):\n",
    "            model.train()\n",
    "            x = x.to(device='cuda', dtype=dtype)\n",
    "            labels = y.to(device=device, dtype=torch.long).view(y.shape[0], 32, 32)\n",
    "            # print(labels.shape)\n",
    "\n",
    "            scores = model(x)\n",
    "            # print(scores.shape)\n",
    "\n",
    "            loss = criterion(scores, labels)\n",
    "            # loss = F.cross_entropy(scores, labels, weight=torch.FloatTensor([1/(float(26265600-1762336)), 1/(float(1762336))]).cuda())\n",
    "            loss_history.append(loss.item())\n",
    "            optimizer.zero_grad()\n",
    "            loss.backward()\n",
    "            optimizer.step()\n",
    "            # print(t)\n",
    "            if t % print_every == 0:\n",
    "                print('Epoch %d, Iteration %d, loss = %.4f' % (e, t, loss.item()))\n",
    "                val_iou = check_iou(small_val_loader, model)\n",
    "                val_iou_history.append(val_iou)\n",
    "                train_iou = check_iou(small_train_loader, model)\n",
    "                train_iou_history.append(train_iou)\n",
    "                print('Validation IOU: ', val_iou)\n",
    "                print('Training IOU: ', train_iou)\n",
    "            # if t % plot_every == 0:\n",
    "                # plot_output(scores, labels)\n",
    "train(model, optimizer, epochs=5000)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "plt.title('Loss over time')\n",
    "plt.xlabel('Iteration')\n",
    "plt.ylabel('Loss')\n",
    "plt.plot(range(len(loss_history)), loss_history)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "val_hist = np.asarray(val_iou_history)\n",
    "train_hist = np.asarray(train_iou_history)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "plt.title('IOU for tumor over time')\n",
    "plt.xlabel('Iteration')\n",
    "plt.ylabel('IOU')\n",
    "plt.plot(range(val_hist.shape[0]), val_hist[:,1], label='val')\n",
    "plt.plot(range(train_hist.shape[0]), train_hist[:,1], label='train')\n",
    "plt.legend()"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "plt.title('IOU for no tumor over time')\n",
    "plt.xlabel('Iteration')\n",
    "plt.ylabel('IOU')\n",
    "plt.plot(range(val_hist.shape[0]), val_hist[:,0], label='val')\n",
    "plt.plot(range(train_hist.shape[0]), train_hist[:,0], label='train')\n",
    "plt.legend()"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": []
  }
 ],
 "metadata": {
  "kernelspec": {
   "display_name": "Python 3",
   "language": "python",
   "name": "python3"
  },
  "language_info": {
   "codemirror_mode": {
    "name": "ipython",
    "version": 3
   },
   "file_extension": ".py",
   "mimetype": "text/x-python",
   "name": "python",
   "nbconvert_exporter": "python",
   "pygments_lexer": "ipython3",
   "version": "3.7.3"
  }
 },
 "nbformat": 4,
 "nbformat_minor": 2
}