From 12805762cc4c98bd348b6169eb1074067c6f878a Mon Sep 17 00:00:00 2001
From: Luca Coviello <luca.coviello@gmail.com>
Date: Thu, 26 Jul 2018 14:37:36 +0200
Subject: [PATCH] Perform data augmentation in training and normalize correctly
 at test time

---
 image.py  |  4 ++--
 val.ipynb | 25 +++++++++++--------------
 2 files changed, 13 insertions(+), 16 deletions(-)

diff --git a/image.py b/image.py
index c8506d49..c4a6a848 100644
--- a/image.py
+++ b/image.py
@@ -11,7 +11,7 @@ def load_data(img_path,train = True):
     img = Image.open(img_path).convert('RGB')
     gt_file = h5py.File(gt_path)
     target = np.asarray(gt_file['density'])
-    if False:
+    if train:
         crop_size = (img.size[0]/2,img.size[1]/2)
         if random.randint(0,9)<= -1:
             
@@ -40,4 +40,4 @@ def load_data(img_path,train = True):
     target = cv2.resize(target,(target.shape[1]/8,target.shape[0]/8),interpolation = cv2.INTER_CUBIC)*64
     
     
-    return img,target
\ No newline at end of file
+    return img,target
diff --git a/val.ipynb b/val.ipynb
index ba73c72a..e6475ccd 100644
--- a/val.ipynb
+++ b/val.ipynb
@@ -18,11 +18,12 @@
     "from scipy.ndimage.filters import gaussian_filter \n",
     "import scipy\n",
     "import json\n",
-    "import torchvision.transforms.functional as F\n",
+    "from torchvision import datasets, transforms\n",
     "from matplotlib import cm as CM\n",
     "from image import *\n",
     "from model import CSRNet\n",
     "import torch\n",
+    "\n",
     "%matplotlib inline"
    ]
   },
@@ -34,11 +35,11 @@
    },
    "outputs": [],
    "source": [
-    "from torchvision import datasets, transforms\n",
-    "transform=transforms.Compose([\n",
-    "                       transforms.ToTensor(),transforms.Normalize(mean=[0.485, 0.456, 0.406],\n",
-    "                                     std=[0.229, 0.224, 0.225]),\n",
-    "                   ])"
+    "transform = transforms.Compose([\n",
+    "    transforms.ToTensor(),\n",
+    "    transforms.Normalize(mean=[0.485, 0.456, 0.406],\n",
+    "                         std=[0.229, 0.224, 0.225])\n",
+    "])"
    ]
   },
   {
@@ -326,17 +327,13 @@
    "source": [
     "mae = 0\n",
     "for i in xrange(len(img_paths)):\n",
-    "    img = 255.0 * F.to_tensor(Image.open(img_paths[i]).convert('RGB'))\n",
-    "\n",
-    "    img[0,:,:]=img[0,:,:]-92.8207477031\n",
-    "    img[1,:,:]=img[1,:,:]-95.2757037428\n",
-    "    img[2,:,:]=img[2,:,:]-104.877445883\n",
-    "    img = img.cuda()\n",
-    "    #img = transform(Image.open(img_paths[i]).convert('RGB')).cuda()\n",
+    "    img = transform(Image.open(img_paths[i]).convert('RGB')).cuda()\n",
     "    gt_file = h5py.File(img_paths[i].replace('.jpg','.h5').replace('images','ground_truth'),'r')\n",
     "    groundtruth = np.asarray(gt_file['density'])\n",
+    "    \n",
     "    output = model(img.unsqueeze(0))\n",
     "    mae += abs(output.detach().cpu().sum().numpy()-np.sum(groundtruth))\n",
+    "    \n",
     "    print i,mae\n",
     "print mae/len(img_paths)"
    ]
@@ -358,7 +355,7 @@
    "name": "python",
    "nbconvert_exporter": "python",
    "pygments_lexer": "ipython2",
-   "version": "2.7.13"
+   "version": "2.7.15"
   }
  },
  "nbformat": 4,