TTA

Author: rijobro

modules/inverse_transforms_and_test_time_augmentations.ipynb

@@ -11,7 +11,7 @@
     "\n",
     "### What are transforms?\n",
     "\n",
-    "- We use transforms to modify data. In MONAI, we use them to (for exampl) load images from file, add a channel component, normalise the intensities and reshape the image.\n",
+    "- We use transforms to modify data. In MONAI, we use them to (for example) load images from file, add a channel component, normalise the intensities and reshape the image.\n",
     "- We can also use transforms as a method of data augmentation – we have a finite amount of data so to avoid overfitting, we can apply random transforms to modify our data each epoch.\n",
     "- Examples of random transformations might be randomly flipping, rotating, cropping, padding, zooming, as well as applying non-rigid deformations.\n",
     "\n",
@@ -229,8 +229,8 @@
     "    def __call__(self, data):\n",
     "        d = dict(data)\n",
     "        im = d[self.label_key]\n",
-    "        q = np.sum((im > 0).reshape(-1, im.shape[-1]), axis=0)\n",
-    "        _slice = np.where(q == np.max(q))[0][0]\n",
+    "        q = (im > 0).reshape(-1, im.shape[-1]).sum(dim=0)\n",
+    "        _slice = q.argmax(dim=0)\n",
     "        for key in self.keys:\n",
     "            d[key] = d[key][..., _slice]\n",
     "        return d\n",
@@ -244,10 +244,9 @@
     "    def __call__(self, data):\n",
     "        d = {}\n",
     "        for key in self.keys:\n",
-    "            fname = os.path.basename(\n",
-    "                data[key + \"_meta_dict\"][\"filename_or_obj\"])\n",
+    "            fname = os.path.basename(data[key].meta[\"filename_or_obj\"])\n",
     "            path = os.path.join(self.path, key, fname)\n",
-    "            nib.save(nib.Nifti1Image(data[key], np.eye(4)), path)\n",
+    "            nib.save(nib.Nifti1Image(data[key].numpy(), np.eye(4)), path)\n",
     "            d[key] = path\n",
     "        return d\n",
     "\n",
@@ -258,12 +257,12 @@
     "    os.makedirs(os.path.join(data_dir, key), exist_ok=True)\n",
     "transform_2d_slice = Compose([\n",
     "    LoadImaged(keys),\n",
-    "    FromMetaTensord(keys),\n",
     "    AsChannelFirstd(\"image\"),\n",
     "    AddChanneld(\"label\"),\n",
     "    SliceWithMaxNumLabelsd(keys, \"label\"),\n",
     "    SaveSliced(keys, data_dir),\n",
     "])\n",
+    "\n",
     "# Running the whole way through the dataset will create the 2D slices and save to file\n",
     "ds_2d = Dataset(data_dicts, transform_2d_slice)\n",
     "dl_2d = DataLoader(ds_2d, batch_size=1, num_workers=10)\n",
@@ -306,8 +305,8 @@
     "    [\n",
     "        LoadImaged(keys),\n",
     "        FromMetaTensord(keys),\n",
-    "        Lambdad(\"label\", lambda x: (x > 0).astype(\n",
-    "            np.float64)),  # make label binary\n",
+    "        Lambdad(\"label\", lambda x: (x > 0).to(\n",
+    "            torch.float32)),  # make label binary\n",
     "        RandAffined(\n",
     "            keys,\n",
     "            prob=1.0,\n",
@@ -1627,8 +1626,8 @@
     "minimal_transforms = Compose([\n",
     "    LoadImaged(keys),\n",
     "    FromMetaTensord(keys),\n",
-    "    Lambdad(\"label\", lambda x: (x > 0).astype(\n",
-    "        np.float64)),  # make label binary\n",
+    "    Lambdad(\"label\", lambda x: (x > 0).to(\n",
+    "        torch.float32)),  # make label binary\n",
     "    ScaleIntensityd(\"image\"),\n",
     "    EnsureTyped(keys),\n",
     "])\n",