Upload semanticallysegmentdeezglaciers.ipynb

semanticallysegmentdeezglaciers.ipynb

@@ -333,8 +333,11 @@
         "def compute_metrics(eval_pred):\n",
         "    # Ensure that gradient computation is turned off, as it is not needed for evaluation\n",
         "    with torch.no_grad():\n",
+        "        # This computes the final logits tensor by interpolating the output logits to the size of the labels tensor from an input of size (batch_size, num_labels, height, width)\n",
+        "        # This is input that has gone through the model's forward pass\n",
         "        logits, labels = eval_pred\n",
         "        logits_tensor = torch.from_numpy(logits)\n",
+        "        # this can lead to very high ram usage for the upscaling\n",
         "        logits_tensor = nn.functional.interpolate(\n",
         "            logits_tensor,\n",
         "            size=labels.shape[-2:],\n",
@@ -343,8 +346,10 @@
         "        )\n",
         "        # Take the argmax of the logits tensor along dimension 1 to get the predicted labels\n",
         "        logits_tensor = logits_tensor.argmax(dim=1)\n",
-        "        # Detach the predicted labels from the computation graph and move them to the CPU to save memory and to use numpy features\n",
+        "        # Detach the predicted labels from the computation graph and move them to the CPU \n",
+        "        # (although they are already on the CPU) to save memory and to use numpy features like the metrics module\n",
         "        pred_labels = logits_tensor.detach().cpu().numpy()\n",
+        "        # Computes metrics\n",
         "        metrics = metric.compute(\n",
         "            predictions=pred_labels,\n",
         "            references=labels,\n",