objectness thresholding is done. Todo: nms

Aniket-ind · Feb 23, 2019 · 449d693 · 449d693
1 parent 1872084
commit 449d693
Show file tree

Hide file tree

Showing 2 changed files with 17 additions and 20 deletions.
diff --git a/face_detector/experiments/yolo/Untitled.ipynb b/face_detector/experiments/yolo/Untitled.ipynb
@@ -108,7 +108,7 @@
   },
   {
    "cell_type": "code",
-   "execution_count": 6,
+   "execution_count": 55,
    "metadata": {},
    "outputs": [
     {
@@ -126,32 +126,29 @@
     "# todo nms\n",
     "x = torch.randn((2, 3, 416, 416))\n",
     "predictions = darknet.forward(x, 'cpu')\n",
-    "objectness_thres = 0.8\n",
+    "predictions[:, :, 4] *= 1000 # to be removed. just for testing\n",
+    "objectness_thres = 0.2\n",
     "nms_thres = 0.4\n",
     "classes = 80 # darknet.layers_list[-1][0].classes"
    ]
   },
   {
    "cell_type": "code",
-   "execution_count": 11,
+   "execution_count": 70,
    "metadata": {},
-   "outputs": [
-    {
-     "data": {
-      "text/plain": [
-       "tensor(0)"
-      ]
-     },
-     "execution_count": 11,
-     "metadata": {},
-     "output_type": "execute_result"
-    }
-   ],
+   "outputs": [],
    "source": [
-    "objectness_mask = predictions[:, :, 4] > objectness_thres\n",
-    "(predictions[:, :, 4] == 0).sum()\n",
-    "# todo nms. check whether > < 0 in the previous line is \n",
-    "# expectable behaviour"
+    "# Non-max Suppression:\n",
+    "# 1. Filter out predictions with low objectness score\n",
+    "# 2. \n",
+    "# objectiveness filtering:\n",
+    "# replace all boxes with predicted probability lower than objectness_thres \n",
+    "# with zeros: calculate 0/1 mask and apply it to the prediction tensor\n",
+    "# Note: '>' returns Byte but '*' needs Float unsqueezed back to 3D\n",
+    "objectness_mask = (predictions[:, :, 4] > objectness_thres).float().unsqueeze(2)\n",
+    "predictions = predictions * objectness_mask\n",
+    "\n",
+    "# "
    ]
   },
   {

diff --git a/face_detector/experiments/yolo/darknet.py b/face_detector/experiments/yolo/darknet.py
@@ -111,7 +111,7 @@ def forward(self, x, device):
                 pwh = anchors_tens.repeat(w * w, 1).unsqueeze(0)
                 pwh = pwh.to(device)
 
-                # transform the predictions
+                # transform the predictions (center, size, objectness, class scores)
                 x[:, :, 0:2] = (torch.sigmoid(x[:, :, 0:2]) + cxy) * stride
                 x[:, :, 2:4] = (pwh * torch.exp(x[:, :, 2:4])) * stride
                 x[:, :, 4] = torch.sigmoid(x[:, :, 4])