Merge pull request #153 from april-tools/fix_notebook

Fix notebook to align with previous PRs
april-tools · Nov 27, 2023 · 7e1aa99 · 7e1aa99
2 parents 3a96e96 + 077383a
commit 7e1aa99
Showing 1 changed file with 55 additions and 24 deletions.
diff --git a/notebooks/digits.ipynb b/notebooks/digits.ipynb
@@ -28,7 +28,7 @@
    "metadata": {},
    "outputs": [],
    "source": [
-    "device = torch.device(\"cpu\")  # The device to use, e.g., \"cpu\", \"cuda\", \"cuda:1\""
+    "device = torch.device(\"cuda\")  # The device to use, e.g., \"cpu\", \"cuda\", \"cuda:1\""
    ]
   },
   {
@@ -231,30 +231,60 @@
      "output_type": "stream",
      "text": [
       "TensorizedPC(\n",
-      "  (input_layer): CategoricalLayer()\n",
+      "  (input_layer): CategoricalLayer(\n",
+      "    (params): ReparamEFCategorical()\n",
+      "  )\n",
       "  (scope_layer): ScopeLayer()\n",
       "  (inner_layers): ModuleList(\n",
-      "    (0-1): 2 x CPLayer()\n",
-      "    (2): MixingLayer()\n",
-      "    (3): CPLayer()\n",
-      "    (4): MixingLayer()\n",
-      "    (5): CPLayer()\n",
-      "    (6): MixingLayer()\n",
-      "    (7): CPLayer()\n",
-      "    (8): MixingLayer()\n",
-      "    (9): CPLayer()\n",
-      "    (10): MixingLayer()\n",
-      "    (11-12): 2 x CPLayer()\n",
-      "    (13): MixingLayer()\n",
-      "    (14-15): 2 x CPLayer()\n",
-      "    (16): MixingLayer()\n",
+      "    (0-1): 2 x CollapsedCPLayer(\n",
+      "      (params_in): ReparamExp()\n",
+      "    )\n",
+      "    (2): SumLayer(\n",
+      "      (params): ReparamExp()\n",
+      "    )\n",
+      "    (3): CollapsedCPLayer(\n",
+      "      (params_in): ReparamExp()\n",
+      "    )\n",
+      "    (4): SumLayer(\n",
+      "      (params): ReparamExp()\n",
+      "    )\n",
+      "    (5): CollapsedCPLayer(\n",
+      "      (params_in): ReparamExp()\n",
+      "    )\n",
+      "    (6): SumLayer(\n",
+      "      (params): ReparamExp()\n",
+      "    )\n",
+      "    (7): CollapsedCPLayer(\n",
+      "      (params_in): ReparamExp()\n",
+      "    )\n",
+      "    (8): SumLayer(\n",
+      "      (params): ReparamExp()\n",
+      "    )\n",
+      "    (9): CollapsedCPLayer(\n",
+      "      (params_in): ReparamExp()\n",
+      "    )\n",
+      "    (10): SumLayer(\n",
+      "      (params): ReparamExp()\n",
+      "    )\n",
+      "    (11-12): 2 x CollapsedCPLayer(\n",
+      "      (params_in): ReparamExp()\n",
+      "    )\n",
+      "    (13): SumLayer(\n",
+      "      (params): ReparamExp()\n",
+      "    )\n",
+      "    (14-15): 2 x CollapsedCPLayer(\n",
+      "      (params_in): ReparamExp()\n",
+      "    )\n",
+      "    (16): SumLayer(\n",
+      "      (params): ReparamExp()\n",
+      "    )\n",
       "  )\n",
       ")\n"
      ]
     }
    ],
    "source": [
-    "from cirkit.utils.reparams import reparam_exp\n",
+    "from cirkit.reparams.leaf import ReparamExp\n",
     "from cirkit.models.tensorized_circuit import TensorizedPC\n",
     "pc = TensorizedPC.from_region_graph(\n",
     "    region_graph,\n",
@@ -265,7 +295,7 @@
     "    layer_cls=layer_cls,\n",
     "    layer_kwargs=layer_kwargs,\n",
     "    num_classes=1,\n",
-    "    reparam=reparam_exp\n",
+    "    reparam=ReparamExp\n",
     ")\n",
     "pc.to(device)\n",
     "print(pc)"
@@ -351,8 +381,9 @@
     "for epoch_idx in range(num_epochs):\n",
     "    running_loss = 0.0\n",
     "    for batch, _ in train_dataloader:\n",
-    "        log_score = pc.forward(batch)\n",
-    "        log_pf = pc_pf()          # Compute the partition function\n",
+    "        batch = batch.to(device).unsqueeze(dim=-1)  # Add a channel dimension\n",
+    "        log_score = pc(batch)\n",
+    "        log_pf = pc_pf(batch)     # Compute the partition function\n",
     "        lls = log_score - log_pf  # Compute the log-likelihood\n",
     "        loss = -torch.mean(lls)   # The loss is the negative average log-likelihood\n",
     "        loss.backward()\n",
@@ -385,18 +416,18 @@
      "name": "stdout",
      "output_type": "stream",
      "text": [
-      "Average test LL: -833.413\n",
-      "Bits per dimension: 1.533623705697752\n"
+      "Average test LL: -833.414\n",
+      "Bits per dimension: 1.5336248522113194\n"
      ]
     }
    ],
    "source": [
     "with torch.no_grad():\n",
     "    pc.eval()\n",
-    "    log_pf = pc_pf()  # Compute the partition function once for testing\n",
+    "    log_pf = pc_pf(torch.empty((), device=device))  # Compute the partition function once for testing\n",
     "    test_lls = 0.0\n",
     "    for batch, _ in test_dataloader:\n",
-    "        log_score = pc.forward(batch)\n",
+    "        log_score = pc(batch.to(device).unsqueeze(dim=-1))\n",
     "        lls = log_score - log_pf\n",
     "        test_lls += lls.sum().item()\n",
     "    average_ll = test_lls / len(data_test)\n",