Add ViT to examples

nbs/examples/vit.torch.ipynb

@@ -31,6 +31,7 @@
     "\n",
     "# utils\n",
     "import functools as ft\n",
+    "import matplotlib.pyplot as plt\n",
     "from dataclasses import dataclass\n",
     "from tqdm.auto import tqdm"
    ]
@@ -250,17 +251,19 @@
     "    xs, ys = next(iter(data_loader))\n",
     "    params = model.init(init_key, xs, train=False)\n",
     "    opt_state = optimizer.init(params)\n",
-    "    losses = []\n",
+    "    losses, steps = [], 0\n",
     "\n",
     "    for epoch in range(epochs):\n",
-    "        with tqdm(data_loader, desc=f\"Epoch {epoch}\", unit='batch') as pbar:\n",
-    "            for batch in pbar:\n",
-    "                rng_key, key = jrand.split(rng_key)\n",
-    "                params, opt_state, loss = step(\n",
-    "                    params, model, optimizer, opt_state, batch, key\n",
-    "                )\n",
-    "                losses.append(loss)\n",
-    "                pbar.set_postfix({\"loss\": loss})\n",
+    "        for batch in data_loader:\n",
+    "            rng_key, key = jrand.split(rng_key)\n",
+    "            params, opt_state, loss = step(\n",
+    "                params, model, optimizer, opt_state, batch, key\n",
+    "            )\n",
+    "            losses.append(loss)\n",
+    "            steps += 1\n",
+    "\n",
+    "            if steps % 500 == 0:\n",
+    "                print(f\"Epoch: {epoch}, Step: {steps}, Loss: {loss}\")\n",
     "    return params, losses"
    ]
   },
@@ -271,11 +274,11 @@
    "outputs": [],
    "source": [
     "# Hyperparameters\n",
-    "lr = 1e-3\n",
+    "lr = 3e-4\n",
     "dropout_rate = 0.1\n",
     "beta1 = 0.9\n",
     "beta2 = 0.99\n",
-    "batch_size = 64 * 2\n",
+    "batch_size = 64 * 2 * 2\n",
     "patch_size = 4\n",
     "num_patches = 64\n",
     "num_steps = 100000\n",
@@ -297,7 +300,21 @@
      "name": "stdout",
      "output_type": "stream",
      "text": [
-      "Files already downloaded and verified\n",
+      "Downloading https://www.cs.toronto.edu/~kriz/cifar-10-python.tar.gz to /tmp/CIFAR/cifar-10-python.tar.gz\n"
+     ]
+    },
+    {
+     "name": "stderr",
+     "output_type": "stream",
+     "text": [
+      "100%|██████████| 170498071/170498071 [00:02<00:00, 63361930.39it/s]\n"
+     ]
+    },
+    {
+     "name": "stdout",
+     "output_type": "stream",
+     "text": [
+      "Extracting /tmp/CIFAR/cifar-10-python.tar.gz to /tmp/CIFAR/\n",
       "Files already downloaded and verified\n"
      ]
     }
@@ -364,7 +381,33 @@
    "metadata": {},
    "outputs": [],
    "source": [
-    "_, losses = train(vit, opt, dl, 100)"
+    "params, losses = train(vit, opt, dl, 500)"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "metadata": {},
+   "outputs": [
+    {
+     "name": "stdout",
+     "output_type": "stream",
+     "text": [
+      "Accuracy: 0.8149\n"
+     ]
+    }
+   ],
+   "source": [
+    "corrects = []\n",
+    "\n",
+    "dl = jdl.DataLoader(test_dataset, 'pytorch', batch_size=batch_size * 4, shuffle=True)\n",
+    "for batch in dl:\n",
+    "    img, label = batch\n",
+    "    logits = vit.apply(params, img, rngs={'dropout': jrand.PRNGKey(0)}, train=False)\n",
+    "    preds = jnp.argmax(logits, axis=-1)\n",
+    "    corrects.append((preds == label))\n",
+    "\n",
+    "print(f\"Accuracy: {np.concatenate(corrects).mean()}\")"
    ]
   }
  ],

nbs/sidebar.yml

@@ -2,6 +2,9 @@ website:
   sidebar:
     contents:
       - index.ipynb
+      - section: Examples
+        contents:
+        - examples/vit.torch.ipynb
       - section: API
         contents:
         - core.ipynb