fix: use alpha_t in FocalLoss as per standard formula

Ch14_Computer_Vision/Single_Shot_Multibox_Detection.ipynb

@@ -640,12 +640,25 @@
   },
   {
    "cell_type": "code",
-   "execution_count": 21,
+   "execution_count": null,
    "metadata": {},
    "outputs": [],
    "source": [
     "id_cat = dict()\n",
-    "id_cat[0] = 'pikachu'\n",
+    "id_cat[0] = 'pikachu'"
+   ]
+  },
+  {
+   "metadata": {
+    "jupyter": {
+     "is_executing": true
+    }
+   },
+   "cell_type": "code",
+   "source": [
+    "\n",
+    "import torch\n",
+    "import torch.nn as nn\n",
     "\n",
     "class FocalLoss(nn.Module):\n",
     "    def __init__(self, alpha=0.25, gamma=2, device=\"cuda:0\", eps=1e-10):\n",
@@ -658,19 +671,29 @@
     "    def forward(self, input, target):\n",
     "        p = torch.sigmoid(input)\n",
     "        pt = p * target.float() + (1.0 - p) * (1 - target).float()\n",
-    "        alpha_t = (1.0 - self.alpha) * target.float() + self.alpha * (1 - target).float()\n",
-    "        loss = - 1.0 * torch.pow((1 - pt), self.gamma) * torch.log(pt + self.eps)\n",
-    "        return loss.sum()\n",
-    "    \n",
+    "        alpha_t = self.alpha * target.float() + (1.0 - self.alpha) * (1 - target).float()\n",
+    "        loss = - alpha_t * torch.pow((1 - pt), self.gamma) * torch.log(pt + self.eps)\n",
+    "        return loss.sum()\n"
+   ],
+   "outputs": [],
+   "execution_count": null
+  },
+  {
+   "metadata": {},
+   "cell_type": "code",
+   "outputs": [],
+   "execution_count": 21,
+   "source": [
+    "\n",
     "class SSDLoss(nn.Module):\n",
     "    def __init__(self, loc_factor, jaccard_overlap, device = \"cuda:0\", **kwargs):\n",
     "        super().__init__()\n",
     "        self.fl = FocalLoss(**kwargs)\n",
     "        self.loc_factor = loc_factor\n",
     "        self.jaccard_overlap = jaccard_overlap\n",
-    "        \n",
     "\n",
-    "        \n",
+    "\n",
+    "\n",
     "        self.device = device\n",
     "\n",
     "    def one_hot_encoding(labels, num_classes):\n",
@@ -684,23 +707,23 @@
     "                          torch.log((x[:, 3:4] / anchors[overlap_indicies, 3:4]))\n",
     "                         ], dim=1)\n",
     "\n",
-    "    def forward(self, class_hat, bb_hat, class_true, bb_true, anchors):        \n",
+    "    def forward(self, class_hat, bb_hat, class_true, bb_true, anchors):\n",
     "        loc_loss = 0.0\n",
     "        class_loss = 0.0\n",
-    "    \n",
-    "        \n",
+    "\n",
+    "\n",
     "        for i in range(len(class_true)):  # Batch level\n",
     "\n",
     "            class_hat_i = class_hat[i, :, :]\n",
     "\n",
     "            bb_true_i = bb_true[i].float()\n",
-    "            \n",
-    "            class_true_i = class_true[i]            \n",
-    "            \n",
+    "\n",
+    "            class_true_i = class_true[i]\n",
+    "\n",
     "            class_target = torch.zeros(class_hat_i.shape[0]).long().to(self.device)\n",
-    "            \n",
+    "\n",
     "            overlap_list = d2l.find_overlap(bb_true_i.squeeze(0), anchors, self.jaccard_overlap)\n",
-    "            \n",
+    "\n",
     "            temp_loc_loss = 0.0\n",
     "            for j in range(len(overlap_list)):  # BB level\n",
     "                overlap = overlap_list[j]\n",