chore: nb cleanup and more transformations

eschmidt42 · Feb 20, 2024 · 2c79664 · 2c79664
1 parent 46f1ead
commit 2c79664
Showing 1 changed file with 56 additions and 12 deletions.
diff --git a/nbs/unet-isbi2012.ipynb b/nbs/unet-isbi2012.ipynb
@@ -33,6 +33,7 @@
     "\n",
     "import matplotlib.pyplot as plt\n",
     "import numpy as np\n",
+    "import seaborn as sns\n",
     "import tifffile as tiff\n",
     "import torch\n",
     "import torch.nn as nn\n",
@@ -64,7 +65,6 @@
    "metadata": {},
    "outputs": [],
    "source": [
-    "# X = Image.open(path_em_data / \"train-volume.tif\")\n",
     "X = tiff.imread(path_em_data / \"train-volume.tif\")\n",
     "Y = tiff.imread(path_em_data / \"train-labels.tif\")"
    ]
@@ -95,6 +95,15 @@
     "    plt.tight_layout()"
    ]
   },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "_x.mean(), _x.std()"
+   ]
+  },
   {
    "cell_type": "code",
    "execution_count": null,
@@ -152,13 +161,12 @@
     "        return self.n\n",
     "\n",
     "    def __getitem__(self, idx: int) -> T.Tuple[torch.Tensor, torch.Tensor]:\n",
-    "        # img = torch.from_numpy(self.X[idx])\n",
+    "\n",
     "        img = tv_tensors.Image(self.X[idx])\n",
-    "        labels = tv_tensors.Mask(self.Y[idx] == 255)  # , dtype=torch.bool\n",
+    "        labels = tv_tensors.Mask(self.Y[idx] == 255)\n",
     "\n",
     "        if self.transform:\n",
     "            img, labels = self.transform(img, labels)\n",
-    "            # labels = self.transform(labels)\n",
     "\n",
     "        return img, labels\n",
     "\n",
@@ -186,7 +194,16 @@
    "metadata": {},
    "outputs": [],
    "source": [
-    "trafos = vision_trafos_v2.RandomCrop(size=(64, 64))"
+    "trafos = vision_trafos_v2.Compose(\n",
+    "    [\n",
+    "        vision_trafos_v2.RandomAffine(degrees=0, shear=5),\n",
+    "        vision_trafos_v2.RandomCrop(size=(64, 64)),\n",
+    "        vision_trafos_v2.RandomVerticalFlip(),\n",
+    "        vision_trafos_v2.RandomHorizontalFlip(),\n",
+    "        vision_trafos_v2.ToDtype(torch.float32),\n",
+    "        vision_trafos_v2.Normalize(mean=[0.0], std=[255.0]),\n",
+    "    ]\n",
+    ")"
    ]
   },
   {
@@ -206,10 +223,16 @@
    "outputs": [],
    "source": [
     "def show_image_and_labels(image: torch.Tensor, labels: torch.Tensor):\n",
-    "    fig, axs = plt.subplots(ncols=2)\n",
+    "    fig, axs = plt.subplots(ncols=2, nrows=2)\n",
     "\n",
-    "    axs[0].imshow(img[0])\n",
-    "    axs[1].imshow(labels)\n",
+    "    axs[0, 0].imshow(image[0])\n",
+    "    axs[0, 1].imshow(labels)\n",
+    "\n",
+    "    sns.histplot(x=image[0].ravel(), ax=axs[1, 0])\n",
+    "    sns.histplot(x=labels.ravel(), ax=axs[1, 1])\n",
+    "\n",
+    "    print(image[0].ravel().mean(), image[0].ravel().std())\n",
+    "    print(image[0].ravel().min(), image[0].ravel().max())\n",
     "\n",
     "    plt.tight_layout()\n",
     "\n",
@@ -438,7 +461,7 @@
    "metadata": {},
    "outputs": [],
    "source": [
-    "learning_rate = 1e-4\n",
+    "learning_rate = 1e-3\n",
     "\n",
     "scheduler = optim.lr_scheduler.OneCycleLR(\n",
     "    optimizer=optimizer,\n",
@@ -529,7 +552,7 @@
    "metadata": {},
    "outputs": [],
    "source": [
-    "losses_simple = loss_callback.get_losses().dropna(subset=[\"loss_valid\"])\n",
+    "losses_simple = loss_callback.get_losses_valid()\n",
     "\n",
     "losses_simple"
    ]
@@ -776,7 +799,7 @@
    "metadata": {},
    "outputs": [],
    "source": [
-    "model = Model3Layers(res=True)"
+    "model = Model3Layers(res=False)"
    ]
   },
   {
@@ -922,10 +945,31 @@
    "metadata": {},
    "outputs": [],
    "source": [
-    "losses_shallow = loss_callback.get_losses().dropna(subset=[\"loss_valid\"])\n",
+    "losses_shallow = loss_callback.get_losses_valid()\n",
     "display(losses_simple.tail(), losses_shallow.tail())"
    ]
   },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "fig, ax = plt.subplots()\n",
+    "\n",
+    "sns.lineplot(\n",
+    "    data=losses_simple, x=\"iteration\", y=\"loss_valid\", label=\"simple\", ax=ax\n",
+    ")\n",
+    "sns.lineplot(\n",
+    "    data=losses_shallow, x=\"iteration\", y=\"loss_valid\", label=\"u-net\", ax=ax\n",
+    ")\n",
+    "\n",
+    "ax.legend(title=\"model\")\n",
+    "ax.set(yscale=\"log\")\n",
+    "\n",
+    "plt.tight_layout()"
+   ]
+  },
   {
    "cell_type": "code",
    "execution_count": null,