new plotting strategy

.gitignore

@@ -210,6 +210,7 @@ __marimo__/
 logs/
 *.log
 models/
+logs*
 
 # Images
 *.png

ThermalSolver/autoregressive_module.py

@@ -15,52 +15,102 @@ def import_class(class_path: str):
     return cls
 
 
-def _plot_mesh(pos_, y_, y_pred_, y_true_, batch, i, batch_idx):
-    for j in [0, 10, 20, 30]:
+def _plot_mesh(pos_, y_, y_pred_, y_true_, batch, cells, i, batch_idx):
+    # print(pos_.shape, y_.shape, y_pred_.shape, y_true_.shape)
+    for j in [0]:
         idx = (batch == j).nonzero(as_tuple=True)[0]
         y = y_[idx].detach().cpu()
         y_pred = y_pred_[idx].detach().cpu()
         pos = pos_[idx].detach().cpu()
+        # print(pos.shape, y.shape, y_pred.shape)
         y_true = y_true_[idx].detach().cpu()
         y_true = torch.clamp(y_true, min=0)
-        folder = f"{j:02d}_images"
+        folder = f"{batch_idx:02d}_images"
         if os.path.exists(folder) is False:
             os.makedirs(folder)
-        pos = pos.detach().cpu()
-        tria = Triangulation(pos[:, 0], pos[:, 1])
-        plt.figure(figsize=(24, 5))
+        triangles = torch.vstack([cells[:, [0, 1, 2]], cells[:, [0, 2, 3]]])
+        tria = Triangulation(pos[:, 0], pos[:, 1], triangles=triangles)
+        plt.figure(figsize=(24, 6))
+        # plt.subplot(1, 4, 1)
+        # plt.tricontourf(tria, y.squeeze().numpy(), levels=100)
+        # plt.colorbar()
+        # plt.title("Step t-1")
+        # plt.tripcolor(tria, y_pred.squeeze().numpy()
+        # plt.savefig("test_scatter_step_before.png", dpi=72)
+        # x = z
         plt.subplot(1, 4, 1)
-        plt.tricontourf(tria, y.squeeze().numpy(), levels=100)
-        plt.colorbar()
-        plt.title("Step t-1")
-        plt.subplot(1, 4, 2)
         plt.tricontourf(tria, y_pred.squeeze().numpy(), levels=100)
+        # plt.scatter(pos[:, 0], pos[:, 1], c=y_pred.squeeze().numpy(), s=20, cmap="viridis",)
         plt.colorbar()
-        plt.title("Step t Predicted")
-        plt.subplot(1, 4, 3)
+        plt.title(f"Prediction at timestep {i:03d}")
+        plt.subplot(1, 4, 2)
         plt.tricontourf(tria, y_true.squeeze().numpy(), levels=100)
+        # plt.scatter(pos[:, 0], pos[:, 1], c=y_true.squeeze().numpy(), s=20, cmap="viridis")
         plt.colorbar()
-        plt.title("t True")
+        plt.title("Ground Truth Steady State")
+        plt.subplot(1, 4, 3)
+        per_element_relative_error = torch.abs(y_pred - y_true) / (
+            y_true + 1e-6
+        )
+        per_element_relative_error = torch.clamp(
+            per_element_relative_error, max=1.0, min=0.0
+        )
+        plt.tricontourf(
+            tria,
+            per_element_relative_error.squeeze(),
+            levels=100,
+            vmin=0,
+            vmax=1.0,
+        )
+        # plt.scatter(pos[:, 0], pos[:, 1], c=per_element_relative_error.squeeze().numpy(), s=20, cmap="viridis", vmin=0, vmax=1.0)
+        plt.colorbar()
+        plt.title("Relative Error")
         plt.subplot(1, 4, 4)
-        plt.tricontourf(tria, (y_true - y_pred).squeeze().numpy(), levels=100)
+        absolute_error = torch.abs(y_pred - y_true)
+        plt.tricontourf(tria, absolute_error.squeeze(), levels=100)
+        # plt.scatter(pos[:, 0], pos[:, 1], c=absolute_error.squeeze().numpy(), s=20, cmap="viridis")
         plt.colorbar()
-        plt.title("Error")
+        plt.title("Absolute Error")
         plt.suptitle("GNO", fontsize=16)
         name = f"{folder}/{j:04d}_graph_iter_{i:04d}.png"
         plt.savefig(name, dpi=72)
         plt.close()
 
 
-def _plot_losses(losses, batch_idx):
-    folder = f"{batch_idx:02d}_images"
-    plt.figure()
-    plt.plot(losses)
+def _plot_losses(relative_errors, test_losses, relative_update, batch_idx):
+    # folder = f"{batch_idx:02d}_images"
+    plt.figure(figsize=(18, 6))
+    plt.subplot(1, 3, 1)
+    for i, losses in enumerate(test_losses):
+        plt.plot(losses)
+        if i == 3:
+            break
     plt.yscale("log")
     plt.xlabel("Iteration")
-    plt.ylabel("Loss")
+    plt.ylabel("Test Loss")
     plt.title("Test Loss over Iterations")
     plt.grid(True)
-    file_name = f"{folder}/test_loss.png"
+    plt.subplot(1, 3, 2)
+    for i, losses in enumerate(relative_errors):
+        plt.plot(losses)
+        if i == 3:
+            break
+    plt.yscale("log")
+    plt.xlabel("Iteration")
+    plt.ylabel("Relative Error")
+    plt.title("Relative error over Iterations")
+    plt.grid(True)
+    plt.subplot(1, 3, 3)
+    for i, updates in enumerate(relative_update):
+        plt.plot(updates)
+        if i == 3:
+            break
+    plt.yscale("log")
+    plt.xlabel("Iteration")
+    plt.ylabel("Relative Update")
+    plt.title("Relative update over Iterations")
+    plt.grid(True)
+    file_name = f"test_errors.png"
     plt.savefig(file_name, dpi=300)
     plt.close()
 
@@ -80,6 +130,9 @@ class GraphSolver(LightningModule):
         # print(f"Param: {param[0]}")
         self.loss = loss if loss is not None else torch.nn.MSELoss()
         self.unrolling_steps = unrolling_steps
+        self.test_losses = []
+        self.test_relative_errors = []
+        self.test_relative_updates = []
 
     def _compute_loss(self, x, y):
         return self.loss(x, y)
@@ -116,13 +169,12 @@ class GraphSolver(LightningModule):
         return out
 
     def _preprocess_batch(self, batch: Batch):
-        x, y, c, edge_index, edge_attr, nodal_area = (
+        x, y, c, edge_index, edge_attr = (
             batch.x,
             batch.y,
             batch.c,
             batch.edge_index,
             batch.edge_attr,
-            batch.nodal_area,
         )
         edge_attr = 1 / edge_attr
         conductivity = self._compute_c_ij(c, edge_index)
@@ -133,35 +185,9 @@ class GraphSolver(LightningModule):
         x, y, edge_index, edge_attr, conductivity = self._preprocess_batch(
             batch
         )
-        # deg = self._compute_deg(edge_index, edge_attr, x.size(0))
         losses = []
-        # print(x.shape, y.shape)
-        # # print(torch.max(edge_index), torch.min(edge_index))
-        # plt.figure()
-        # plt.subplot(2,3,1)
-        # plt.scatter(batch.pos[:,0].cpu(), batch.pos[:,1].cpu(), c=x.squeeze().cpu())
-        # plt.subplot(2,3,2)
-        # plt.scatter(batch.pos[:,0].cpu(), batch.pos[:,1].cpu(), c=y[:,0,:].squeeze().cpu())
-        # plt.subplot(2,3,3)
-        # plt.scatter(batch.pos[:,0].cpu(), batch.pos[:,1].cpu(), c=y[:,1,:].squeeze().cpu())
-        # plt.subplot(2,3,4)
-        # plt.scatter(batch.pos[:,0].cpu(), batch.pos[:,1].cpu(), c=y[:,2,:].squeeze().cpu())
-        # plt.subplot(2,3,5)
-        # plt.scatter(batch.pos[:,0].cpu(), batch.pos[:,1].cpu(), c=y[:,3,:].squeeze().cpu())
-        # plt.subplot(2,3,6)
-        # plt.scatter(batch.pos[:,0].cpu(), batch.pos[:,1].cpu(), c=y[:,4,:].squeeze().cpu())
-        # plt.suptitle("Training Batch Visualization", fontsize=16)
-        # plt.savefig("training_batch_visualization.png", dpi=300)
-        # plt.close()
-        # y = z
-        pos = batch.pos
-        boundary_mask = batch.boundary_mask
-        boundary_values = batch.boundary_values
-        # plt.scatter(pos[boundary_mask,0].cpu(), pos[boundary_mask,1].cpu(), c=boundary_values.cpu(), s=1)
-        # plt.savefig("boundary_nodes.png", dpi=300)
-        # y = z
-        scale = 50
         for i in range(self.unrolling_steps):
+            # print(f"Training step {i+1}/{self.unrolling_steps}")
             out = self._compute_model_steps(
                 x,
                 edge_index,
@@ -172,15 +198,26 @@ class GraphSolver(LightningModule):
                 conductivity,
             )
             x = out
-            # print(out.shape, y[:, i, :].shape)
             losses.append(self.loss(out.flatten(), y[:, i, :].flatten()))
-        # print(self.model.scale_edge_attr.item())
-
         loss = torch.stack(losses).mean()
-        # for param in self.model.parameters():
-        #     print(f"Param: {param.shape}, Grad: {param.grad}")
-        #     print(f"Param: {param[0]}")
         self._log_loss(loss, batch, "train")
+        for i, layer in enumerate(self.model.layers):
+            self.log(
+                f"{i:03d}_alpha",
+                layer.alpha,
+                prog_bar=True,
+                on_epoch=True,
+                on_step=False,
+                batch_size=int(batch.num_graphs),
+            )
+        self.log(
+            "dt",
+            self.model.dt,
+            prog_bar=True,
+            on_epoch=True,
+            on_step=False,
+            batch_size=int(batch.num_graphs),
+        )
         return loss
 
     def validation_step(self, batch: Batch, batch_idx):
@@ -201,20 +238,20 @@ class GraphSolver(LightningModule):
                 batch.boundary_values,
                 conductivity,
             )
-            if (
-                batch_idx == 0
-                and self.current_epoch % 10 == 0
-                and self.current_epoch > 0
-            ):
-                _plot_mesh(
-                    batch.pos,
-                    x,
-                    out,
-                    y[:, i, :],
-                    batch.batch,
-                    i,
-                    self.current_epoch,
-                )
+            # if (
+            #     batch_idx == 0
+            #     and self.current_epoch % 10 == 0
+            #     and self.current_epoch > 0
+            # ):
+            #     _plot_mesh(
+            #         batch.pos,
+            #         x,
+            #         out,
+            #         y[:, i, :],
+            #         batch.batch,
+            #         i,
+            #         self.current_epoch,
+            #     )
             x = out
             losses.append(self.loss(out, y[:, i, :]))
 
@@ -222,8 +259,81 @@ class GraphSolver(LightningModule):
         self._log_loss(loss, batch, "val")
         return loss
 
+    def _check_convergence(self, y_new, y_old, tol=1e-4):
+        l2_norm = torch.norm(y_new - y_old, p=2)
+        y_old_norm = torch.norm(y_old, p=2)
+        rel_error = l2_norm / (y_old_norm)
+        return rel_error.item() < tol, rel_error.item()
+
     def test_step(self, batch: Batch, batch_idx):
-        pass
+        x, y, edge_index, edge_attr, conductivity = self._preprocess_batch(
+            batch
+        )
+        # deg = self._compute_deg(edge_index, edge_attr, x.size(0))
+        losses = []
+        all_losses = []
+        norms = []
+        s = []
+        relative_updates = []
+        sequence_length = y.size(1)
+        y = y[:, -1, :].unsqueeze(1)
+        _plot_mesh(
+            batch.pos, x, x, y[:, -1, :], batch.batch, batch.cells, 0, batch_idx
+        )
+        for i in range(200):
+            out = self._compute_model_steps(
+                # torch.cat([x,pos], dim=-1),
+                x,
+                edge_index,
+                edge_attr,
+                # deg,
+                batch.boundary_mask,
+                batch.boundary_values,
+                conductivity,
+            )
+            norms.append(torch.norm(out - x, p=2).item())
+            converged, relative_update = self._check_convergence(out, x)
+            relative_updates.append(relative_update)
+            if batch_idx <= 4:
+                print(f"Plotting iteration {i}, norm diff: {norms[-1]}")
+                _plot_mesh(
+                    batch.pos,
+                    x,
+                    out,
+                    y[:, -1, :],
+                    batch.batch,
+                    batch.cells,
+                    i + 1,
+                    batch_idx,
+                )
+            x = out
+            loss = self.loss(out, y[:, -1, :])
+            relative_error = torch.abs(out - y[:, -1, :]) / (
+                torch.abs(y[:, -1, :]) + 1e-6
+            )
+            mean_relative_error = relative_error.mean()
+            all_losses.append(mean_relative_error.item())
+            losses.append(loss)
+            if converged:
+                print(
+                    f"Test step converged at iteration {i} for batch {batch_idx}"
+                )
+                break
+        loss = torch.stack(losses).mean()
+        self.test_losses.append(losses)
+        self.test_relative_errors.append(all_losses)
+        self.test_relative_updates.append(relative_updates)
+        self._log_loss(loss, batch, "test")
+        return loss
+
+    def on_test_end(self):
+        if len(self.test_losses) > 0:
+            _plot_losses(
+                self.test_relative_errors,
+                self.test_losses,
+                self.test_relative_updates,
+                batch_idx=0,
+            )
 
     def configure_optimizers(self):
         optimizer = torch.optim.AdamW(self.parameters(), lr=1e-3)

ThermalSolver/graph_datamodule.py

@@ -82,7 +82,9 @@ class GraphDataModule(LightningDataModule):
         conductivity = torch.tensor(
             snapshot["conductivity"], dtype=torch.float32
         )
-        temperature = torch.tensor(snapshot["temperature"], dtype=torch.float32)
+        temperature = torch.tensor(
+            snapshot["temperature"], dtype=torch.float32
+        )[:50]
 
         pos = torch.tensor(geometry["points"], dtype=torch.float32)[:, :2]
 

ThermalSolver/graph_datamodule_unsteady.py

@@ -1,50 +1,20 @@
 import torch
 from tqdm import tqdm
 from lightning import LightningDataModule
-from datasets import load_dataset
+from datasets import load_dataset, concatenate_datasets
 from torch_geometric.data import Data
 from torch_geometric.loader import DataLoader
 from torch_geometric.utils import to_undirected
 from .mesh_data import MeshData
-
-# from torch.utils.data import Dataset
-from torch_geometric.utils import scatter
-
-
-def compute_nodal_area(edge_index, edge_attr, num_nodes):
-    """
-    1. Calculates Area ~ (Min Edge Length)^2
-    2. Scales by Mean so average cell has size 1.0
-    """
-    row, col = edge_index
-    dist = edge_attr.squeeze()
-
-    # 1. Get 'h' (Closest neighbor distance)
-    # Using 'min' filters out diagonal connections in the quad mesh
-    h = scatter(dist, col, dim=0, dim_size=num_nodes, reduce="min")
-
-    # 2. Estimate Raw Area
-    raw_area = h.pow(2)
-
-    # 3. Mean Scaling (The Best Normalization)
-    # This keeps values near 1.0, preserving stability AND physics ratios.
-    # We detach to ensure no gradients flow here (it's static data).
-    mean_val = raw_area.mean().detach()
-
-    # Result:
-    # Small cells -> approx 0.1
-    # Large cells -> approx 5.0
-    # Average -> 1.0
-    # nodal_area = (raw_area / mean_val).unsqueeze(-1) + 1e-6
-    nodal_area = raw_area
-    return nodal_area.unsqueeze(-1)
+from typing import List, Union
 
 
 class GraphDataModule(LightningDataModule):
     def __init__(
         self,
         hf_repo: str,
-        split_name: str,
+        split_name: Union[str, List[str]],
+        n_elements: int = None,
         train_size: float = 0.2,
         val_size: float = 0.1,
         test_size: float = 0.1,
@@ -52,18 +22,24 @@ class GraphDataModule(LightningDataModule):
         remove_boundary_edges: bool = False,
         build_radial_graph: bool = False,
         radius: float = None,
-        start_unrolling_steps: int = 1,
+        unrolling_steps: int = 1,
+        aggregate_timesteps: int = 1,
+        min_normalized_diff: float = 1e-3,
     ):
         super().__init__()
         self.hf_repo = hf_repo
         self.split_name = split_name
+        self.n_elements = n_elements
         self.dataset_dict = {}
         self.train_dataset, self.val_dataset, self.test_dataset = (
             None,
             None,
             None,
         )
-        self.unrolling_steps = start_unrolling_steps
+        self.unrolling_steps = unrolling_steps
+        self.aggregate_timesteps = aggregate_timesteps
+        self.min_normalized_diff = min_normalized_diff
+
         self.geometry_dict = {}
         self.train_size = train_size
         self.val_size = val_size
@@ -74,8 +50,33 @@ class GraphDataModule(LightningDataModule):
         self.radius = radius
 
     def prepare_data(self):
-        dataset = load_dataset(self.hf_repo, name="snapshots")[self.split_name]
-        geometry = load_dataset(self.hf_repo, name="geometry")[self.split_name]
+        if isinstance(self.split_name, list):
+            dataset_list = []
+            geometry_list = []
+            for split in self.split_name:
+                dataset_list.append(
+                    load_dataset(self.hf_repo, name="snapshots")[split]
+                )
+                geometry_list.append(
+                    load_dataset(self.hf_repo, name="geometry")[split]
+                )
+
+            dataset = concatenate_datasets(dataset_list)
+            geometry = concatenate_datasets(geometry_list)
+            idx = torch.randperm(len(dataset))
+            dataset = dataset.select(idx.tolist())
+            geometry = geometry.select(idx.tolist())
+        else:
+            dataset = load_dataset(self.hf_repo, name="snapshots")[
+                self.split_name
+            ]
+            geometry = load_dataset(self.hf_repo, name="geometry")[
+                self.split_name
+            ]
+
+        if self.n_elements is not None:
+            dataset = dataset.select(range(self.n_elements))
+            geometry = geometry.select(range(self.n_elements))
 
         total_len = len(dataset)
         train_len = int(self.train_size * total_len)
@@ -91,63 +92,39 @@ class GraphDataModule(LightningDataModule):
             "test": geometry.select(range(train_len + valid_len, total_len)),
         }
 
-    def _compute_boundary_mask(
-        self, bottom_ids, right_ids, top_ids, left_ids, temperature
-    ):
-        left_ids = left_ids[~torch.isin(left_ids, bottom_ids)]
-        right_ids = right_ids[~torch.isin(right_ids, bottom_ids)]
-        left_ids = left_ids[~torch.isin(left_ids, top_ids)]
-        right_ids = right_ids[~torch.isin(right_ids, top_ids)]
-
-        bottom_bc = temperature[bottom_ids].median()
-        bottom_bc_mask = torch.ones(len(bottom_ids)) * bottom_bc
-        left_bc = temperature[left_ids].median()
-        left_bc_mask = torch.ones(len(left_ids)) * left_bc
-        right_bc = temperature[right_ids].median()
-        right_bc_mask = torch.ones(len(right_ids)) * right_bc
-
-        boundary_values = torch.cat(
-            [bottom_bc_mask, right_bc_mask, left_bc_mask], dim=0
-        )
-        boundary_mask = torch.cat([bottom_ids, right_ids, left_ids], dim=0)
-
-        return boundary_mask, boundary_values
-
     def _build_dataset(
         self,
         snapshot: dict,
         geometry: dict,
+        test: bool = False,
     ) -> Data:
         conductivity = torch.tensor(
             geometry["conductivity"], dtype=torch.float32
         )
-        temperatures = torch.tensor(
-            snapshot["temperatures"], dtype=torch.float32
-        )[:40]
-        times = torch.tensor(snapshot["times"], dtype=torch.float32)
-
+        temperatures = (
+            torch.tensor(snapshot["unsteady"], dtype=torch.float32)
+            if not test
+            else torch.stack(
+                [
+                    torch.tensor(snapshot["unsteady"], dtype=torch.float32)[
+                        0, ...
+                    ],
+                    torch.tensor(snapshot["steady"], dtype=torch.float32),
+                ],
+                dim=0,
+            )
+        )
+        if not test:
+            for t in range(1, temperatures.size(0)):
+                diff = temperatures[t, :] - temperatures[t - 1, :]
+                norm_diff = torch.norm(diff, p=2) / torch.norm(
+                    temperatures[t - 1], p=2
+                )
+                if norm_diff < self.min_normalized_diff:
+                    temperatures = temperatures[: t + 1, :]
+                    break
         pos = torch.tensor(geometry["points"], dtype=torch.float32)[:, :2]
-
-        bottom_ids = torch.tensor(
-            geometry["bottom_boundary_ids"], dtype=torch.long
-        )
-        top_ids = torch.tensor(geometry["top_boundary_ids"], dtype=torch.long)
-        left_ids = torch.tensor(geometry["left_boundary_ids"], dtype=torch.long)
-        right_ids = torch.tensor(
-            geometry["right_boundary_ids"], dtype=torch.long
-        )
-
         if self.build_radial_graph:
-            # from pina.graph import RadiusGraph
-
-            # if self.radius is None:
-            #     raise ValueError("Radius must be specified for radial graph.")
-            # edge_index = RadiusGraph.compute_radius_graph(
-            #     pos, radius=self.radius
-            # )
-            # from torch_geometric.utils import remove_self_loops
-
-            # edge_index, _ = remove_self_loops(edge_index)
             raise NotImplementedError(
                 "Radial graph building not implemented yet."
             )
@@ -157,18 +134,39 @@ class GraphDataModule(LightningDataModule):
             ).T
             edge_index = to_undirected(edge_index, num_nodes=pos.size(0))
 
-        boundary_mask, boundary_values = self._compute_boundary_mask(
-            bottom_ids, right_ids, top_ids, left_ids, temperatures[0, :]
+        boundary_mask = torch.tensor(
+            geometry["constraints_mask"], dtype=torch.int64
         )
+        boundary_values = temperatures[0, boundary_mask]
+
         edge_attr = torch.norm(pos[edge_index[0]] - pos[edge_index[1]], dim=1)
-        nodal_area = compute_nodal_area(edge_index, edge_attr, pos.size(0))
         if self.remove_boundary_edges:
             boundary_idx = torch.unique(boundary_mask)
             edge_index_mask = ~torch.isin(edge_index[1], boundary_idx)
             edge_index = edge_index[:, edge_index_mask]
             edge_attr = edge_attr[edge_index_mask]
-        n_data = temperatures.size(0) - self.unrolling_steps
+
+        n_data = max(temperatures.size(0) - self.unrolling_steps, 1)
         data = []
+
+        if test:
+            cells = geometry.get("cells", None)
+            if cells is not None:
+                cells = torch.tensor(cells, dtype=torch.int64)
+            data.append(
+                MeshData(
+                    x=temperatures[0, :].unsqueeze(-1),
+                    y=temperatures[1:2, :].unsqueeze(-1).permute(1, 0, 2),
+                    c=conductivity.unsqueeze(-1),
+                    edge_index=edge_index,
+                    pos=pos,
+                    edge_attr=edge_attr,
+                    boundary_mask=boundary_mask,
+                    boundary_values=boundary_values,
+                    cells=cells,
+                )
+            )
+            return data
         for i in range(n_data):
             x = temperatures[i, :].unsqueeze(-1)
             y = (
@@ -186,7 +184,6 @@ class GraphDataModule(LightningDataModule):
                     edge_attr=edge_attr,
                     boundary_mask=boundary_mask,
                     boundary_values=boundary_values,
-                    nodal_area=nodal_area,
                 )
             )
         return data
@@ -213,7 +210,7 @@ class GraphDataModule(LightningDataModule):
             ]
         if stage == "test" or stage is None:
             self.test_data = [
-                self._build_dataset(snap, geom)
+                self._build_dataset(snap, geom, test=True)
                 for snap, geom in tqdm(
                     zip(self.dataset_dict["test"], self.geometry_dict["test"]),
                     desc="Building test graphs",
@@ -234,33 +231,36 @@ class GraphDataModule(LightningDataModule):
         # self.train_dataset = ds
         # print(type(self.train_data[0]))
         ds = [i for data in self.train_data for i in data]
-        # print(type(ds[0]))
+        print(
+            f"\nLoading training data, using {self.unrolling_steps} unrolling steps..."
+        )
         return DataLoader(
             ds,
             batch_size=self.batch_size,
             shuffle=True,
             num_workers=8,
-            pin_memory=True,
+            pin_memory=False,
         )
 
     def val_dataloader(self):
+        print(
+            f"\nLoading validation data, using {self.unrolling_steps} unrolling steps..."
+        )
         ds = [i for data in self.val_data for i in data]
         return DataLoader(
             ds,
             batch_size=128,
             shuffle=False,
             num_workers=8,
-            pin_memory=True,
+            pin_memory=False,
         )
 
     def test_dataloader(self):
-        ds = self.create_autoregressive_datasets(
-            dataset="test", no_unrolling=True
-        )
+        ds = [i for data in self.test_data for i in data]
         return DataLoader(
             ds,
-            batch_size=self.batch_size,
+            batch_size=1,
             shuffle=False,
             num_workers=8,
-            pin_memory=True,
+            pin_memory=False,
         )

ThermalSolver/model/diffusion_net.py

@@ -1,6 +1,28 @@
 import torch
 import torch.nn as nn
 from torch_geometric.nn import MessagePassing
+from torch.nn.utils import spectral_norm
+
+
+class LogPhysEncoder(nn.Module):
+    """
+    Processes 1/dx in log-space to handle multiple scales of geometry
+    (from micro-meshes to macro-meshes) without numerical instability.
+    """
+
+    def __init__(self, hidden_dim):
+        super().__init__()
+        self.mlp = nn.Sequential(
+            spectral_norm(nn.Linear(1, hidden_dim)),
+            nn.GELU(),
+            spectral_norm(nn.Linear(hidden_dim, 1)),
+            nn.Softplus(),  # Physical conductance must be positive
+        )
+
+    def forward(self, inv_dx):
+        # We use log(1/dx) to linearize the scale of different geometries
+        log_inv_dx = torch.log(inv_dx + 1e-9)
+        return self.mlp(log_inv_dx)
 
 
 class DiffusionLayer(MessagePassing):
@@ -13,28 +35,27 @@ class DiffusionLayer(MessagePassing):
         channels: int,
         **kwargs,
     ):
-
         super().__init__(aggr="add", **kwargs)
 
-        self.dt = nn.Parameter(torch.tensor(1e-4))
         self.conductivity_net = nn.Sequential(
-            nn.Linear(channels, channels, bias=False),
+            spectral_norm(nn.Linear(channels, channels, bias=False)),
             nn.GELU(),
-            nn.Linear(channels, channels, bias=False),
+            spectral_norm(nn.Linear(channels, channels, bias=False)),
         )
 
-        self.phys_encoder = nn.Sequential(
-            nn.Linear(1, 8, bias=False),
-            nn.Tanh(),
-            nn.Linear(8, 1, bias=False),
-            nn.Softplus(),
-        )
+        self.phys_encoder = LogPhysEncoder(hidden_dim=channels)
+
+        self.alpha_param = nn.Parameter(torch.tensor(1e-2))
+
+    @property
+    def alpha(self):
+        return torch.clamp(self.alpha_param, min=1e-7, max=1.0)
 
     def forward(self, x, edge_index, edge_weight, conductivity):
         edge_weight = edge_weight.unsqueeze(-1)
         conductance = self.phys_encoder(edge_weight)
         net_flux = self.propagate(edge_index, x=x, conductance=conductance)
-        return x + ((net_flux) * self.dt)
+        return x + self.alpha * net_flux
 
     def message(self, x_i, x_j, conductance):
         delta = x_j - x_i
@@ -44,15 +65,21 @@ class DiffusionLayer(MessagePassing):
 
 
 class DiffusionNet(nn.Module):
-    def __init__(self, input_dim=1, output_dim=1, hidden_dim=8, n_layers=4):
+    def __init__(
+        self,
+        input_dim=1,
+        output_dim=1,
+        hidden_dim=8,
+        n_layers=4,
+        shared_weights=False,
+    ):
         super().__init__()
 
         # Encoder: Projects input temperature to hidden feature space
         self.enc = nn.Sequential(
-            nn.Linear(input_dim, hidden_dim, bias=True),
-            nn.GELU(),
-            nn.Linear(hidden_dim, hidden_dim, bias=True),
+            spectral_norm(nn.Linear(input_dim, hidden_dim, bias=True)),
             nn.GELU(),
+            spectral_norm(nn.Linear(hidden_dim, hidden_dim, bias=True)),
         )
 
         self.scale_x = nn.Parameter(torch.zeros(hidden_dim))
@@ -60,27 +87,40 @@ class DiffusionNet(nn.Module):
         # Scale parameters for conditioning
         self.scale_edge_attr = nn.Parameter(torch.zeros(1))
 
-        # Stack of Diffusion Layers
-        self.layers = torch.nn.ModuleList(
-            [DiffusionLayer(hidden_dim) for _ in range(n_layers)]
-        )
+        # If shared_weights is True, use the same DiffusionLayer multiple times
+        if shared_weights:
+            diffusion_layer = DiffusionLayer(hidden_dim)
+            self.layers = torch.nn.ModuleList(
+                [diffusion_layer for _ in range(n_layers)]
+            )
+        # If shared_weights is False, use separate DiffusionLayers
+        else:
+            # Stack of Diffusion Layers
+            self.layers = torch.nn.ModuleList(
+                [DiffusionLayer(hidden_dim) for _ in range(n_layers)]
+            )
 
         # Decoder: Projects hidden features back to Temperature space
         self.dec = nn.Sequential(
-            nn.Linear(hidden_dim, hidden_dim, bias=True),
+            spectral_norm(nn.Linear(hidden_dim, hidden_dim, bias=True)),
             nn.GELU(),
-            nn.Linear(hidden_dim, output_dim, bias=True),
+            spectral_norm(nn.Linear(hidden_dim, output_dim, bias=True)),
             nn.Softplus(),  # Ensure positive temperature output
         )
 
         self.func = torch.nn.GELU()
 
+        self.dt_param = nn.Parameter(torch.tensor(1e-2))
+
+    @property
+    def dt(self):
+        return torch.clamp(self.dt_param, min=1e-5, max=0.5)
+
     def forward(self, x, edge_index, edge_attr, conductivity):
         # 1. Global Residual Connection setup
         # We save the input to add it back at the very end.
         # The network learns the correction (Delta T), not the absolute T.
         x_input = x
-
         # 2. Encode
         h = self.enc(x) * torch.exp(self.scale_x)
 
@@ -98,5 +138,5 @@ class DiffusionNet(nn.Module):
 
         # 6. Final Update (Explicit Euler Step)
         # T_new = T_old + Correction
-        # return x_input + delta_x
-        return delta_x
+        return delta_x + x_input * self.dt
+        # return delta_x

ThermalSolver/switch_dataloader_callback.py

@@ -0,0 +1,104 @@
+import torch
+from lightning.pytorch.callbacks import Callback
+import os
+
+
+class SwitchDataLoaderCallback(Callback):
+    def __init__(
+        self,
+        ckpt_path,
+        increase_unrolling_steps_by,
+        increase_unrolling_steps_every,
+        max_unrolling_steps=10,
+        patience=None,
+        last_patience=None,
+        metric="val/loss",
+    ):
+        super().__init__()
+        self.ckpt_path = ckpt_path
+        if os.path.exists(ckpt_path) is False:
+            os.makedirs(ckpt_path)
+        self.increase_unrolling_steps_by = increase_unrolling_steps_by
+        self.increase_unrolling_steps_every = increase_unrolling_steps_every
+        self.max_unrolling_steps = max_unrolling_steps
+        self.metric = metric
+        self.actual_loss = torch.inf
+        if patience is not None:
+            self.patience = patience
+        if last_patience is not None:
+            self.last_patience = last_patience
+        self.no_improvement_epochs = 0
+        self.last_step_reached = False
+
+    def on_validation_epoch_end(self, trainer, pl_module):
+        self._metric_tracker(trainer, pl_module)
+        if self.last_step_reached is False:
+            self._unrolling_steps_handler(pl_module, trainer)
+        else:
+            if self.no_improvement_epochs >= self.last_patience:
+                trainer.should_stop = True
+
+    def _metric_tracker(self, trainer, pl_module):
+        if trainer.callback_metrics.get(self.metric) < self.actual_loss:
+            self.actual_loss = trainer.callback_metrics.get(self.metric)
+            self._save_model(pl_module, trainer)
+            self.no_improvement_epochs = 0
+            print(f"\nNew best {self.metric}: {self.actual_loss:.4f}")
+        else:
+            self.no_improvement_epochs += 1
+            print(
+                f"\nNo improvement in {self.metric} for {self.no_improvement_epochs} epochs."
+            )
+
+    def _should_reload_dataloader(self, trainer):
+        if self.patience is not None:
+            print(
+                f"Checking patience: {self.no_improvement_epochs} / {self.patience}"
+            )
+            if self.no_improvement_epochs >= self.patience:
+                return True
+        elif (
+            trainer.current_epoch + 1 % self.increase_unrolling_steps_every == 0
+        ):
+            print("Reached scheduled epoch for increasing unrolling steps.")
+            return True
+        return False
+
+    def _unrolling_steps_handler(self, pl_module, trainer):
+        if self._should_reload_dataloader(trainer):
+            self._load_model(pl_module)
+            if pl_module.unrolling_steps >= self.max_unrolling_steps:
+                return
+            pl_module.unrolling_steps += self.increase_unrolling_steps_by
+            trainer.datamodule.unrolling_steps = pl_module.unrolling_steps
+            print(f"Incremented unrolling steps to {pl_module.unrolling_steps}")
+            trainer.datamodule.setup(stage="fit")
+            trainer.manual_dataloader_reload()
+            self.actual_loss = torch.inf
+            if pl_module.unrolling_steps >= self.max_unrolling_steps:
+                print(
+                    "Reached max unrolling steps. Stopping further increments."
+                )
+                self.last_step_reached = True
+
+    def _save_model(self, pl_module, trainer):
+        pt_path = os.path.join(
+            self.ckpt_path,
+            f"{pl_module.unrolling_steps}_unrolling_best_model.pt",
+        )
+        torch.save(pl_module.state_dict(), pt_path)  # <--- CHANGED THIS
+        ckpt_path = os.path.join(
+            self.ckpt_path,
+            f"{pl_module.unrolling_steps}_unrolling_best_checkpoint.ckpt",
+        )
+        trainer.save_checkpoint(ckpt_path, weights_only=False)
+
+    def _load_model(self, pl_module):
+        pt_path = os.path.join(
+            self.ckpt_path,
+            f"{pl_module.unrolling_steps}_unrolling_best_model.pt",
+        )
+        pl_module.load_state_dict(torch.load(pt_path, weights_only=True))
+        print(
+            f"Loaded model weights from {pt_path} for unrolling steps = {pl_module.unrolling_steps}"
+        )

experiments/10_steps/config_16_layer_16_hidden_adaptive_refined.yaml

@@ -0,0 +1,72 @@
+# lightning.pytorch==2.5.5
+seed_everything: 1999
+trainer:
+  accelerator: gpu
+  strategy: auto
+  devices: 1
+  num_nodes: 1
+  precision: null
+  logger:
+    - class_path: lightning.pytorch.loggers.WandbLogger
+      init_args:
+        save_dir: logs.autoregressive.wandb
+        project: "thermal-conduction-unsteady-10.steps"
+        name: "16_layer_16_hidden.adaptive_refined"
+  callbacks:
+    # - class_path: lightning.pytorch.callbacks.ModelCheckpoint
+    #   init_args:
+    #     dirpath: logs.autoregressive.wandb/16_refined.10_steps/checkpoints
+    #     monitor: val/loss
+    #     mode: min
+    #     save_top_k: 1
+    #     filename: best-checkpoint
+    # - class_path: lightning.pytorch.callbacks.EarlyStopping
+    #   init_args:
+    #     monitor: val/loss
+    #     mode: min
+    #     patience: 30
+    #     verbose: false
+    - class_path: ThermalSolver.switch_dataloader_callback.SwitchDataLoaderCallback
+      init_args:
+        increase_unrolling_steps_by: 4
+        patience: 5
+        last_patience: 15
+        max_unrolling_steps: 10
+        ckpt_path: logs.autoregressive.wandb/10_steps/basic.adaptive_refined/16_layer_16_hidden/
+  max_epochs: 1000
+  min_epochs: null
+  max_steps: -1
+  min_steps: null
+  overfit_batches: 0.0
+  log_every_n_steps: 0
+  accumulate_grad_batches: 1
+  default_root_dir: null
+  gradient_clip_val: 1.0
+
+model: 
+  class_path: ThermalSolver.autoregressive_module.GraphSolver
+  init_args:
+    model_class_path: ThermalSolver.model.diffusion_net.DiffusionNet
+    model_init_args: 
+      input_dim: 1
+      hidden_dim: 16
+      output_dim: 1
+      n_layers: 16
+    unrolling_steps: 2
+    
+data: 
+  class_path: ThermalSolver.graph_datamodule_unsteady.GraphDataModule
+  init_args:
+    hf_repo: "SISSAmathLab/thermal-conduction-unsteady"
+    split_name: "3_stripes.basic.1_adaptive_refined"
+    n_elements: 100
+    batch_size: 24
+    train_size: 0.7
+    val_size: 0.2
+    test_size: 0.1
+    build_radial_graph: false
+    remove_boundary_edges: true
+    unrolling_steps: 2
+    min_normalized_diff: 1e-4
+optimizer: null
+lr_scheduler: null

experiments/10_steps/config_16_layer_16_hidden_adaptive_refined_combined.yaml

@@ -10,30 +10,25 @@ trainer:
     - class_path: lightning.pytorch.loggers.WandbLogger
       init_args:
         save_dir: logs.autoregressive.wandb
-        project: "thermal-conduction-unsteady"
-        name: "5_step_4_layers_8_hidden"
+        project: "thermal-conduction-unsteady-10.steps"
+        name: "16_layer_16_hidden.adaptive_refined.combined"
   callbacks:
-    - class_path: lightning.pytorch.callbacks.ModelCheckpoint
+    - class_path: ThermalSolver.switch_dataloader_callback.SwitchDataLoaderCallback
       init_args:
-        dirpath: logs.autoregressive.wandb/5_step_4_layers_8_hidden_0.7_radius/checkpoints
-        monitor: val/loss
-        mode: min
-        save_top_k: 1
-        filename: best-checkpoint
-    - class_path: lightning.pytorch.callbacks.EarlyStopping
-      init_args:
-        monitor: val/loss
-        mode: min
-        patience: 10
-        verbose: false
+        increase_unrolling_steps_by: 4
+        patience: 5
+        last_patience: 10
+        max_unrolling_steps: 10
+        ckpt_path: logs.autoregressive.wandb/10_steps/basic.adaptive_refined.combined/16_layer_16_hidden/
   max_epochs: 1000
   min_epochs: null
   max_steps: -1
   min_steps: null
   overfit_batches: 0.0
-  log_every_n_steps: null
+  log_every_n_steps: 0
   accumulate_grad_batches: 1
   default_root_dir: null
+  gradient_clip_val: 1.0
 
 model: 
   class_path: ThermalSolver.autoregressive_module.GraphSolver
@@ -41,22 +36,28 @@ model:
     model_class_path: ThermalSolver.model.diffusion_net.DiffusionNet
     model_init_args: 
       input_dim: 1
-      hidden_dim: 8
+      hidden_dim: 16
       output_dim: 1
-      n_layers: 4
-    unrolling_steps: 5
-
+      n_layers: 16
+    unrolling_steps: 2
+    
 data: 
   class_path: ThermalSolver.graph_datamodule_unsteady.GraphDataModule
   init_args:
     hf_repo: "SISSAmathLab/thermal-conduction-unsteady"
-    split_name: "100_samples_easy_refined"
-    batch_size: 32
+    split_name: 
+      - "4_stripes.basic.1_adaptive_refined"
+      - "3_stripes.basic.1_adaptive_refined"
+      - "2_stripes.basic.1_adaptive_refined"
+    n_elements: 100
+    batch_size: 24
     train_size: 0.7
     val_size: 0.2
     test_size: 0.1
     build_radial_graph: false
     remove_boundary_edges: true
-    start_unrolling_steps: 5
+    unrolling_steps: 2
+    min_normalized_diff: 1e-4
+    
 optimizer: null
-lr_scheduler: null
+lr_scheduler: null

experiments/10_steps/config_16_layer_16_hidden_refined.yaml

@@ -0,0 +1,72 @@
+# lightning.pytorch==2.5.5
+seed_everything: 1999
+trainer:
+  accelerator: gpu
+  strategy: auto
+  devices: 1
+  num_nodes: 1
+  precision: null
+  logger:
+    - class_path: lightning.pytorch.loggers.WandbLogger
+      init_args:
+        save_dir: logs.autoregressive.wandb
+        project: "thermal-conduction-unsteady-10.steps"
+        name: "16_layer_16_hidden.refined"
+  callbacks:
+    # - class_path: lightning.pytorch.callbacks.ModelCheckpoint
+    #   init_args:
+    #     dirpath: logs.autoregressive.wandb/16_refined.10_steps/checkpoints
+    #     monitor: val/loss
+    #     mode: min
+    #     save_top_k: 1
+    #     filename: best-checkpoint
+    # - class_path: lightning.pytorch.callbacks.EarlyStopping
+    #   init_args:
+    #     monitor: val/loss
+    #     mode: min
+    #     patience: 30
+    #     verbose: false
+    - class_path: ThermalSolver.switch_dataloader_callback.SwitchDataLoaderCallback
+      init_args:
+        increase_unrolling_steps_by: 4
+        patience: 5
+        last_patience: 15
+        max_unrolling_steps: 10
+        ckpt_path: logs.autoregressive.wandb/10_steps/basic.refined/16_layer_16_hidden/
+  max_epochs: 1000
+  min_epochs: null
+  max_steps: -1
+  min_steps: null
+  overfit_batches: 0.0
+  log_every_n_steps: 0
+  accumulate_grad_batches: 1
+  default_root_dir: null
+  gradient_clip_val: 1.0
+
+model: 
+  class_path: ThermalSolver.autoregressive_module.GraphSolver
+  init_args:
+    model_class_path: ThermalSolver.model.diffusion_net.DiffusionNet
+    model_init_args: 
+      input_dim: 1
+      hidden_dim: 16
+      output_dim: 1
+      n_layers: 16
+    unrolling_steps: 2
+    
+data: 
+  class_path: ThermalSolver.graph_datamodule_unsteady.GraphDataModule
+  init_args:
+    hf_repo: "SISSAmathLab/thermal-conduction-unsteady"
+    split_name: "3_stripes.basic.refined"
+    n_elements: 50
+    batch_size: 24
+    train_size: 0.7
+    val_size: 0.2
+    test_size: 0.1
+    build_radial_graph: false
+    remove_boundary_edges: true
+    unrolling_steps: 2
+    min_normalized_diff: 1e-4
+optimizer: null
+lr_scheduler: null

experiments/10_steps/config_16_layer_16_hidden_refined_combined.yaml

@@ -0,0 +1,77 @@
+# lightning.pytorch==2.5.5
+seed_everything: 1999
+trainer:
+  accelerator: gpu
+  strategy: auto
+  devices: 1
+  num_nodes: 1
+  precision: null
+  logger:
+    - class_path: lightning.pytorch.loggers.WandbLogger
+      init_args:
+        save_dir: logs.autoregressive.wandb
+        project: "thermal-conduction-unsteady-10.steps"
+        name: "16_layer_16_hidden.refined.combined"
+  callbacks:
+    # - class_path: lightning.pytorch.callbacks.ModelCheckpoint
+    #   init_args:
+    #     dirpath: logs.autoregressive.wandb/16_refined.10_steps/checkpoints
+    #     monitor: val/loss
+    #     mode: min
+    #     save_top_k: 1
+    #     filename: best-checkpoint
+    # - class_path: lightning.pytorch.callbacks.EarlyStopping
+    #   init_args:
+    #     monitor: val/loss
+    #     mode: min
+    #     patience: 30
+    #     verbose: false
+    - class_path: ThermalSolver.switch_dataloader_callback.SwitchDataLoaderCallback
+      init_args:
+        increase_unrolling_steps_by: 4
+        patience: 5
+        last_patience: 10
+        max_unrolling_steps: 10
+        ckpt_path: logs.autoregressive.wandb/10_steps/basic.refined.combined/16_layer_16_hidden/
+  max_epochs: 1000
+  min_epochs: null
+  max_steps: -1
+  min_steps: null
+  overfit_batches: 0.0
+  log_every_n_steps: 0
+  accumulate_grad_batches: 1
+  default_root_dir: null
+  gradient_clip_val: 1.0
+
+model: 
+  class_path: ThermalSolver.autoregressive_module.GraphSolver
+  init_args:
+    model_class_path: ThermalSolver.model.diffusion_net.DiffusionNet
+    model_init_args: 
+      input_dim: 1
+      hidden_dim: 16
+      output_dim: 1
+      n_layers: 16
+    unrolling_steps: 10
+    
+data: 
+  class_path: ThermalSolver.graph_datamodule_unsteady.GraphDataModule
+  init_args:
+    hf_repo: "SISSAmathLab/thermal-conduction-unsteady"
+    split_name: 
+      - "4_stripes.basic.refined"
+      - "3_stripes.basic.refined"
+      - "2_stripes.basic.refined"
+    n_elements: 75
+    batch_size: 24
+    train_size: 0.7
+    val_size: 0.2
+    test_size: 0.1
+    build_radial_graph: false
+    remove_boundary_edges: true
+    unrolling_steps: 10
+    min_normalized_diff: 1e-4
+    
+optimizer: null
+lr_scheduler: null
+ckpt_path: logs.autoregressive.wandb/10_steps/basic.refined.combined/16_layer_16_hidden/6_unrolling_best_checkpoint.ckpt

experiments/10_steps/config_16_layer_16_hidden_refined_star.yaml

@@ -0,0 +1,72 @@
+# lightning.pytorch==2.5.5
+seed_everything: 1999
+trainer:
+  accelerator: gpu
+  strategy: auto
+  devices: 1
+  num_nodes: 1
+  precision: null
+  logger:
+    - class_path: lightning.pytorch.loggers.WandbLogger
+      init_args:
+        save_dir: logs.autoregressive.wandb
+        project: "thermal-conduction-unsteady-10.steps"
+        name: "16_layer_16_hidden.refined.star"
+  callbacks:
+    # - class_path: lightning.pytorch.callbacks.ModelCheckpoint
+    #   init_args:
+    #     dirpath: logs.autoregressive.wandb/16_refined.10_steps/checkpoints
+    #     monitor: val/loss
+    #     mode: min
+    #     save_top_k: 1
+    #     filename: best-checkpoint
+    # - class_path: lightning.pytorch.callbacks.EarlyStopping
+    #   init_args:
+    #     monitor: val/loss
+    #     mode: min
+    #     patience: 30
+    #     verbose: false
+    - class_path: ThermalSolver.switch_dataloader_callback.SwitchDataLoaderCallback
+      init_args:
+        increase_unrolling_steps_by: 4
+        patience: 5
+        last_patience: 15
+        max_unrolling_steps: 10
+        ckpt_path: logs.autoregressive.wandb/10_steps/basic.refined.star/16_layer_16_hidden/
+  max_epochs: 1000
+  min_epochs: null
+  max_steps: -1
+  min_steps: null
+  overfit_batches: 0.0
+  log_every_n_steps: 0
+  accumulate_grad_batches: 1
+  default_root_dir: null
+  gradient_clip_val: 1.0
+
+model: 
+  class_path: ThermalSolver.autoregressive_module.GraphSolver
+  init_args:
+    model_class_path: ThermalSolver.model.diffusion_net.DiffusionNet
+    model_init_args: 
+      input_dim: 1
+      hidden_dim: 16
+      output_dim: 1
+      n_layers: 16
+    unrolling_steps: 2
+    
+data: 
+  class_path: ThermalSolver.graph_datamodule_unsteady.GraphDataModule
+  init_args:
+    hf_repo: "SISSAmathLab/thermal-conduction-unsteady"
+    split_name: "3_stripes.star"
+    n_elements: 100
+    batch_size: 24
+    train_size: 0.7
+    val_size: 0.2
+    test_size: 0.1
+    build_radial_graph: false
+    remove_boundary_edges: true
+    unrolling_steps: 2
+    min_normalized_diff: 1e-4
+optimizer: null
+lr_scheduler: null

experiments/10_steps/config_16_layer_16_hidden_star.yaml

@@ -0,0 +1,72 @@
+# lightning.pytorch==2.5.5
+seed_everything: 1999
+trainer:
+  accelerator: gpu
+  strategy: auto
+  devices: 1
+  num_nodes: 1
+  precision: null
+  logger:
+    - class_path: lightning.pytorch.loggers.WandbLogger
+      init_args:
+        save_dir: logs.autoregressive.wandb
+        project: "thermal-conduction-unsteady-10.steps"
+        name: "16_layer_16_hidden.star"
+  callbacks:
+    # - class_path: lightning.pytorch.callbacks.ModelCheckpoint
+    #   init_args:
+    #     dirpath: logs.autoregressive.wandb/16_refined.10_steps/checkpoints
+    #     monitor: val/loss
+    #     mode: min
+    #     save_top_k: 1
+    #     filename: best-checkpoint
+    # - class_path: lightning.pytorch.callbacks.EarlyStopping
+    #   init_args:
+    #     monitor: val/loss
+    #     mode: min
+    #     patience: 30
+    #     verbose: false
+    - class_path: ThermalSolver.switch_dataloader_callback.SwitchDataLoaderCallback
+      init_args:
+        increase_unrolling_steps_by: 4
+        patience: 5
+        last_patience: 15
+        max_unrolling_steps: 10
+        ckpt_path: logs.autoregressive.wandb/10_steps/basic.star/16_layer_16_hidden/
+  max_epochs: 1000
+  min_epochs: null
+  max_steps: -1
+  min_steps: null
+  overfit_batches: 0.0
+  log_every_n_steps: 0
+  accumulate_grad_batches: 1
+  default_root_dir: null
+  gradient_clip_val: 1.0
+
+model: 
+  class_path: ThermalSolver.autoregressive_module.GraphSolver
+  init_args:
+    model_class_path: ThermalSolver.model.diffusion_net.DiffusionNet
+    model_init_args: 
+      input_dim: 1
+      hidden_dim: 16
+      output_dim: 1
+      n_layers: 16
+    unrolling_steps: 2
+    
+data: 
+  class_path: ThermalSolver.graph_datamodule_unsteady.GraphDataModule
+  init_args:
+    hf_repo: "SISSAmathLab/thermal-conduction-unsteady"
+    split_name: "3_stripes.star.refined"
+    n_elements: 100
+    batch_size: 24
+    train_size: 0.7
+    val_size: 0.2
+    test_size: 0.1
+    build_radial_graph: false
+    remove_boundary_edges: true
+    unrolling_steps: 2
+    min_normalized_diff: 1e-4
+optimizer: null
+lr_scheduler: null

experiments/10_steps/config_16_layer_16_hidden_star_combined.yaml

@@ -0,0 +1,75 @@
+# lightning.pytorch==2.5.5
+seed_everything: 1999
+trainer:
+  accelerator: gpu
+  strategy: auto
+  devices: 1
+  num_nodes: 1
+  precision: null
+  logger:
+    - class_path: lightning.pytorch.loggers.WandbLogger
+      init_args:
+        save_dir: logs.autoregressive.wandb
+        project: "thermal-conduction-unsteady-10.steps"
+        name: "16_layer_16_hidden.star.combined"
+  callbacks:
+    # - class_path: lightning.pytorch.callbacks.ModelCheckpoint
+    #   init_args:
+    #     dirpath: logs.autoregressive.wandb/16_refined.10_steps/checkpoints
+    #     monitor: val/loss
+    #     mode: min
+    #     save_top_k: 1
+    #     filename: best-checkpoint
+    # - class_path: lightning.pytorch.callbacks.EarlyStopping
+    #   init_args:
+    #     monitor: val/loss
+    #     mode: min
+    #     patience: 30
+    #     verbose: false
+    - class_path: ThermalSolver.switch_dataloader_callback.SwitchDataLoaderCallback
+      init_args:
+        increase_unrolling_steps_by: 4
+        patience: 5
+        last_patience: 15
+        max_unrolling_steps: 10
+        ckpt_path: logs.autoregressive.wandb/10_steps/basic.star.combined/16_layer_16_hidden/
+  max_epochs: 1000
+  min_epochs: null
+  max_steps: -1
+  min_steps: null
+  overfit_batches: 0.0
+  log_every_n_steps: 0
+  accumulate_grad_batches: 1
+  default_root_dir: null
+  gradient_clip_val: 1.0
+
+model: 
+  class_path: ThermalSolver.autoregressive_module.GraphSolver
+  init_args:
+    model_class_path: ThermalSolver.model.diffusion_net.DiffusionNet
+    model_init_args: 
+      input_dim: 1
+      hidden_dim: 16
+      output_dim: 1
+      n_layers: 16
+    unrolling_steps: 2
+    
+data: 
+  class_path: ThermalSolver.graph_datamodule_unsteady.GraphDataModule
+  init_args:
+    hf_repo: "SISSAmathLab/thermal-conduction-unsteady"
+    split_name: 
+      - "4_stripes.star"
+      - "3_stripes.star"
+      - "2_stripes.star"
+    n_elements: 100
+    batch_size: 24
+    train_size: 0.7
+    val_size: 0.2
+    test_size: 0.1
+    build_radial_graph: false
+    remove_boundary_edges: true
+    unrolling_steps: 2
+    min_normalized_diff: 1e-4
+optimizer: null
+lr_scheduler: null

experiments/10_steps/config_8_layer_16_hidden_star.yaml

@@ -0,0 +1,72 @@
+# lightning.pytorch==2.5.5
+seed_everything: 1999
+trainer:
+  accelerator: gpu
+  strategy: auto
+  devices: 1
+  num_nodes: 1
+  precision: null
+  logger:
+    - class_path: lightning.pytorch.loggers.WandbLogger
+      init_args:
+        save_dir: logs.autoregressive.wandb
+        project: "thermal-conduction-unsteady-10.steps"
+        name: "8_layer_16_hidden.star"
+  callbacks:
+    # - class_path: lightning.pytorch.callbacks.ModelCheckpoint
+    #   init_args:
+    #     dirpath: logs.autoregressive.wandb/16_refined.10_steps/checkpoints
+    #     monitor: val/loss
+    #     mode: min
+    #     save_top_k: 1
+    #     filename: best-checkpoint
+    # - class_path: lightning.pytorch.callbacks.EarlyStopping
+    #   init_args:
+    #     monitor: val/loss
+    #     mode: min
+    #     patience: 30
+    #     verbose: false
+    - class_path: ThermalSolver.switch_dataloader_callback.SwitchDataLoaderCallback
+      init_args:
+        increase_unrolling_steps_by: 4
+        patience: 5
+        last_patience: 15
+        max_unrolling_steps: 10
+        ckpt_path: logs.autoregressive.wandb/10_steps/basic.star/8_layer_16_hidden/
+  max_epochs: 1000
+  min_epochs: null
+  max_steps: -1
+  min_steps: null
+  overfit_batches: 0.0
+  log_every_n_steps: 0
+  accumulate_grad_batches: 1
+  default_root_dir: null
+  gradient_clip_val: 1.0
+
+model: 
+  class_path: ThermalSolver.autoregressive_module.GraphSolver
+  init_args:
+    model_class_path: ThermalSolver.model.diffusion_net.DiffusionNet
+    model_init_args: 
+      input_dim: 1
+      hidden_dim: 16
+      output_dim: 1
+      n_layers: 8
+    unrolling_steps: 2
+    
+data: 
+  class_path: ThermalSolver.graph_datamodule_unsteady.GraphDataModule
+  init_args:
+    hf_repo: "SISSAmathLab/thermal-conduction-unsteady"
+    split_name: "3_stripes.star.refined"
+    n_elements: 100
+    batch_size: 24
+    train_size: 0.7
+    val_size: 0.2
+    test_size: 0.1
+    build_radial_graph: false
+    remove_boundary_edges: true
+    unrolling_steps: 2
+    min_normalized_diff: 1e-4
+optimizer: null
+lr_scheduler: null

experiments/10_steps/test_config_16_layer_16_hidden_star.yaml

@@ -0,0 +1,76 @@
+# lightning.pytorch==2.5.5
+seed_everything: 1999
+trainer:
+  accelerator: cpu
+  strategy: auto
+  devices: 1
+  num_nodes: 1
+  precision: null
+  # logger:
+  #   - class_path: lightning.pytorch.loggers.WandbLogger
+  #     init_args:
+  #       save_dir: logs.autoregressive.wandb
+  #       project: "thermal-conduction-unsteady-10.steps"
+  #       name: "16_layer_16_hidden.adaptive_refined.combined"
+  # callbacks:
+    # - class_path: lightning.pytorch.callbacks.ModelCheckpoint
+    #   init_args:
+    #     dirpath: logs.autoregressive.wandb/16_refined.10_steps/checkpoints
+    #     monitor: val/loss
+    #     mode: min
+    #     save_top_k: 1
+    #     filename: best-checkpoint
+    # - class_path: lightning.pytorch.callbacks.EarlyStopping
+    #   init_args:
+    #     monitor: val/loss
+    #     mode: min
+    #     patience: 30
+    #     verbose: false
+    # - class_path: ThermalSolver.switch_dataloader_callback.SwitchDataLoaderCallback
+    #   init_args:
+    #     increase_unrolling_steps_by: 4
+    #     patience: 15
+    #     last_patience: 20
+    #     max_unrolling_steps: 10
+    #     ckpt_path: logs.autoregressive.wandb/10_steps/basic.adaptive_refined.combined/16_layer_16_hidden/
+  max_epochs: 1000
+  min_epochs: null
+  max_steps: -1
+  min_steps: null
+  overfit_batches: 0.0
+  log_every_n_steps: 0
+  accumulate_grad_batches: 1
+  default_root_dir: null
+  gradient_clip_val: 1.0
+
+model: 
+  class_path: ThermalSolver.autoregressive_module.GraphSolver
+  init_args:
+    model_class_path: ThermalSolver.model.diffusion_net.DiffusionNet
+    model_init_args: 
+      input_dim: 1
+      hidden_dim: 16
+      output_dim: 1
+      n_layers: 16
+    unrolling_steps: 2
+    
+data: 
+  class_path: ThermalSolver.graph_datamodule_unsteady.GraphDataModule
+  init_args:
+    hf_repo: "SISSAmathLab/thermal-conduction-unsteady"
+    split_name: 
+      # - "2_stripes.basic.refined"
+      # - "3_stripes.basic.refined"
+      # - "4_stripes.basic.1_adaptive_refined"
+      - "3_stripes.star"
+    n_elements: 50
+    batch_size: 32
+    train_size: 0.7
+    val_size: 0.2
+    test_size: 0.1
+    build_radial_graph: false
+    remove_boundary_edges: true
+    unrolling_steps: 2
+optimizer: null
+lr_scheduler: null
+ckpt_path: /home/folivo/storage/thermal-conduction-ml/logs.autoregressive.wandb/10_steps/basic.star.combined/16_layer_16_hidden/10_unrolling_best_checkpoint.ckpt

experiments/5_steps/config_16_layer_16_hidden_refined.yaml

@@ -10,8 +10,8 @@ trainer:
     - class_path: lightning.pytorch.loggers.WandbLogger
       init_args:
         save_dir: logs.autoregressive.wandb
-        project: "thermal-conduction-unsteady"
-        name: "16_refined"
+        project: "thermal-conduction-unsteady-5.steps"
+        name: "16_layer_16_hidden"
   callbacks:
     - class_path: lightning.pytorch.callbacks.ModelCheckpoint
       init_args:
@@ -24,16 +24,24 @@ trainer:
       init_args:
         monitor: val/loss
         mode: min
-        patience: 10
+        patience: 30
         verbose: false
+    # - class_path: ThermalSolver.switch_dataloader_callback.SwitchDataLoaderCallback
+    #   init_args:
+    #     increase_unrolling_steps_by: 5
+    #     patience: 10
+    #     last_patience: 15
+    #     max_unrolling_steps: 20
+    #     ckpt_path: logs.autoregressive.wandb/16_16_refined/checkpoints
   max_epochs: 1000
   min_epochs: null
   max_steps: -1
   min_steps: null
   overfit_batches: 0.0
-  log_every_n_steps: null
-  accumulate_grad_batches: 4
+  log_every_n_steps: 0
+  accumulate_grad_batches: 1
   default_root_dir: null
+  gradient_clip_val: 1.0
 
 model: 
   class_path: ThermalSolver.autoregressive_module.GraphSolver
@@ -45,18 +53,19 @@ model:
       output_dim: 1
       n_layers: 16
     unrolling_steps: 5
-
+    
 data: 
   class_path: ThermalSolver.graph_datamodule_unsteady.GraphDataModule
   init_args:
     hf_repo: "SISSAmathLab/thermal-conduction-unsteady"
-    split_name: "100_samples_easy_refined"
-    batch_size: 8
+    split_name: "basic.refined"
+    n_elements: 100
+    batch_size: 32
     train_size: 0.7
     val_size: 0.2
     test_size: 0.1
     build_radial_graph: false
     remove_boundary_edges: true
-    start_unrolling_steps: 5
+    unrolling_steps: 5
 optimizer: null
 lr_scheduler: null

experiments/5_steps/config_32_layer_16_hidden_refined.yaml

@@ -10,12 +10,12 @@ trainer:
     - class_path: lightning.pytorch.loggers.WandbLogger
       init_args:
         save_dir: logs.autoregressive.wandb
-        project: "thermal-conduction-unsteady"
-        name: "standard"
+        project: "thermal-conduction-unsteady-5.steps"
+        name: "32_layer_16_hidden"
   callbacks:
     - class_path: lightning.pytorch.callbacks.ModelCheckpoint
       init_args:
-        dirpath: logs.autoregressive.wandb/standard/checkpoints
+        dirpath: logs.autoregressive.wandb/32_refined/checkpoints
         monitor: val/loss
         mode: min
         save_top_k: 1
@@ -24,16 +24,24 @@ trainer:
       init_args:
         monitor: val/loss
         mode: min
-        patience: 10
+        patience: 30
         verbose: false
+    # - class_path: ThermalSolver.switch_dataloader_callback.SwitchDataLoaderCallback
+    #   init_args:
+    #     increase_unrolling_steps_by: 5
+    #     patience: 10
+    #     last_patience: 15
+    #     max_unrolling_steps: 20
+    #     ckpt_path: logs.autoregressive.wandb/16_16_refined/checkpoints
   max_epochs: 1000
   min_epochs: null
   max_steps: -1
   min_steps: null
   overfit_batches: 0.0
-  log_every_n_steps: null
+  log_every_n_steps: 0
   accumulate_grad_batches: 1
   default_root_dir: null
+  gradient_clip_val: 1.0
 
 model: 
   class_path: ThermalSolver.autoregressive_module.GraphSolver
@@ -41,22 +49,23 @@ model:
     model_class_path: ThermalSolver.model.diffusion_net.DiffusionNet
     model_init_args: 
       input_dim: 1
-      hidden_dim: 8
+      hidden_dim: 16
       output_dim: 1
-      n_layers: 8
+      n_layers: 32
     unrolling_steps: 5
-
+    
 data: 
   class_path: ThermalSolver.graph_datamodule_unsteady.GraphDataModule
   init_args:
     hf_repo: "SISSAmathLab/thermal-conduction-unsteady"
-    split_name: "100_samples_easy_refined"
-    batch_size: 32
+    split_name: "basic.refined"
+    n_elements: 100
+    batch_size: 24
     train_size: 0.7
     val_size: 0.2
     test_size: 0.1
     build_radial_graph: false
     remove_boundary_edges: true
-    start_unrolling_steps: 5
+    unrolling_steps: 5
 optimizer: null
 lr_scheduler: null

experiments/5_steps/config_8_layer_16_hidden_refined.yaml

@@ -10,12 +10,12 @@ trainer:
     - class_path: lightning.pytorch.loggers.WandbLogger
       init_args:
         save_dir: logs.autoregressive.wandb
-        project: "thermal-conduction-unsteady"
-        name: "refined"
+        project: "thermal-conduction-unsteady-5.steps"
+        name: "8_layer_16_hidden"
   callbacks:
     - class_path: lightning.pytorch.callbacks.ModelCheckpoint
       init_args:
-        dirpath: logs.autoregressive.wandb/refined/checkpoints
+        dirpath: logs.autoregressive.wandb/8_refined/checkpoints
         monitor: val/loss
         mode: min
         save_top_k: 1
@@ -24,7 +24,7 @@ trainer:
       init_args:
         monitor: val/loss
         mode: min
-        patience: 10
+        patience: 30
         verbose: false
   max_epochs: 1000
   min_epochs: null
@@ -32,7 +32,7 @@ trainer:
   min_steps: null
   overfit_batches: 0.0
   log_every_n_steps: null
-  accumulate_grad_batches: 2
+  accumulate_grad_batches: 1
   default_root_dir: null
 
 model: 
@@ -50,13 +50,14 @@ data:
   class_path: ThermalSolver.graph_datamodule_unsteady.GraphDataModule
   init_args:
     hf_repo: "SISSAmathLab/thermal-conduction-unsteady"
-    split_name: "100_samples_easy_refined"
-    batch_size: 16
+    split_name: "basic.refined"
+    n_elements: 100
+    batch_size: 32
     train_size: 0.7
     val_size: 0.2
     test_size: 0.1
     build_radial_graph: false
     remove_boundary_edges: true
-    start_unrolling_steps: 5
+    unrolling_steps: 5
 optimizer: null
 lr_scheduler: null

experiments/config_16_layer_8_hidden_refined.yaml

@@ -1,62 +0,0 @@
-# lightning.pytorch==2.5.5
-seed_everything: 1999
-trainer:
-  accelerator: gpu
-  strategy: auto
-  devices: 1
-  num_nodes: 1
-  precision: null
-  logger:
-    - class_path: lightning.pytorch.loggers.WandbLogger
-      init_args:
-        save_dir: logs.autoregressive.wandb
-        project: "thermal-conduction-unsteady"
-        name: "16_8_refined"
-  callbacks:
-    - class_path: lightning.pytorch.callbacks.ModelCheckpoint
-      init_args:
-        dirpath: logs.autoregressive.wandb/16_8_refined/checkpoints
-        monitor: val/loss
-        mode: min
-        save_top_k: 1
-        filename: best-checkpoint
-    - class_path: lightning.pytorch.callbacks.EarlyStopping
-      init_args:
-        monitor: val/loss
-        mode: min
-        patience: 10
-        verbose: false
-  max_epochs: 1000
-  min_epochs: null
-  max_steps: -1
-  min_steps: null
-  overfit_batches: 0.0
-  log_every_n_steps: null
-  accumulate_grad_batches: 2
-  default_root_dir: null
-
-model: 
-  class_path: ThermalSolver.autoregressive_module.GraphSolver
-  init_args:
-    model_class_path: ThermalSolver.model.diffusion_net.DiffusionNet
-    model_init_args: 
-      input_dim: 1
-      hidden_dim: 8
-      output_dim: 1
-      n_layers: 16
-    unrolling_steps: 5
-
-data: 
-  class_path: ThermalSolver.graph_datamodule_unsteady.GraphDataModule
-  init_args:
-    hf_repo: "SISSAmathLab/thermal-conduction-unsteady"
-    split_name: "100_samples_easy_refined"
-    batch_size: 16
-    train_size: 0.7
-    val_size: 0.2
-    test_size: 0.1
-    build_radial_graph: false
-    remove_boundary_edges: true
-    start_unrolling_steps: 5
-optimizer: null
-lr_scheduler: null

experiments/config_4_layer_16_hidden.yaml

@@ -1,64 +0,0 @@
-# lightning.pytorch==2.5.5
-seed_everything: 1999
-trainer:
-  accelerator: gpu
-  strategy: auto
-  devices: 1
-  num_nodes: 1
-  precision: null
-  logger:
-    - class_path: lightning.pytorch.loggers.WandbLogger
-      init_args:
-        save_dir: logs.autoregressive.wandb/wandb
-        project: "thermal-conduction-unsteady"
-        name: "5_step_4_layers_16_hidden"
-        #        retain: true
-  callbacks:
-    - class_path: lightning.pytorch.callbacks.ModelCheckpoint
-      init_args:
-        dirpath: logs.autoregressive.wandb/5_step_4_layers_16_hidden/checkpoints
-        monitor: val/loss
-        mode: min
-        save_top_k: 1
-        filename: best-checkpoint
-    - class_path: lightning.pytorch.callbacks.EarlyStopping
-      init_args:
-        monitor: val/loss
-        mode: min
-        patience: 10
-        verbose: false
-  max_epochs: 1000
-  min_epochs: null
-  max_steps: -1
-  min_steps: null
-  overfit_batches: 0.0
-  log_every_n_steps: null
-  accumulate_grad_batches: 2
-  default_root_dir: null
-
-model: 
-  class_path: ThermalSolver.autoregressive_module.GraphSolver
-  init_args:
-    model_class_path: ThermalSolver.model.diffusion_net.DiffusionNet
-    model_init_args: 
-      input_dim: 1
-      hidden_dim: 16
-      output_dim: 1
-      n_layers: 4
-    unrolling_steps: 5
-
-data: 
-  class_path: ThermalSolver.graph_datamodule_unsteady.GraphDataModule
-  init_args:
-    hf_repo: "SISSAmathLab/thermal-conduction-unsteady"
-    split_name: "100_samples_easy_refined"
-    batch_size: 32
-    train_size: 0.7
-    val_size: 0.2
-    test_size: 0.1
-    build_radial_graph: true
-    radius: 0.5
-    remove_boundary_edges: true
-    start_unrolling_steps: 5
-optimizer: null
-lr_scheduler: null

experiments/config_8_layer_16_hidden.yaml

@@ -1,62 +0,0 @@
-# lightning.pytorch==2.5.5
-seed_everything: 1999
-trainer:
-  accelerator: gpu
-  strategy: auto
-  devices: 1
-  num_nodes: 1
-  precision: null
-  logger:
-    - class_path: lightning.pytorch.loggers.WandbLogger
-      init_args:
-        save_dir: logs.autoregressive.wandb
-        project: "thermal-conduction-unsteady"
-        name: "standard"
-  callbacks:
-    - class_path: lightning.pytorch.callbacks.ModelCheckpoint
-      init_args:
-        dirpath: logs.autoregressive.wandb/standard/checkpoints
-        monitor: val/loss
-        mode: min
-        save_top_k: 1
-        filename: best-checkpoint
-    - class_path: lightning.pytorch.callbacks.EarlyStopping
-      init_args:
-        monitor: val/loss
-        mode: min
-        patience: 10
-        verbose: false
-  max_epochs: 1000
-  min_epochs: null
-  max_steps: -1
-  min_steps: null
-  overfit_batches: 0.0
-  log_every_n_steps: null
-  accumulate_grad_batches: 2
-  default_root_dir: null
-
-model: 
-  class_path: ThermalSolver.autoregressive_module.GraphSolver
-  init_args:
-    model_class_path: ThermalSolver.model.diffusion_net.DiffusionNet
-    model_init_args: 
-      input_dim: 1
-      hidden_dim: 16
-      output_dim: 1
-      n_layers: 8
-    unrolling_steps: 5
-
-data: 
-  class_path: ThermalSolver.graph_datamodule_unsteady.GraphDataModule
-  init_args:
-    hf_repo: "SISSAmathLab/thermal-conduction-unsteady"
-    split_name: "100_samples_easy"
-    batch_size: 16
-    train_size: 0.7
-    val_size: 0.2
-    test_size: 0.1
-    build_radial_graph: false
-    remove_boundary_edges: true
-    start_unrolling_steps: 5
-optimizer: null
-lr_scheduler: null

run.py

@@ -5,7 +5,11 @@ torch.set_float32_matmul_precision("medium")
 
 
 def main():
-    LightningCLI(subclass_mode_data=True, subclass_mode_model=True)
+    LightningCLI(
+        subclass_mode_data=True,
+        subclass_mode_model=True,
+        save_config_kwargs={"overwrite": True},
+    )
 
 
 if __name__ == "__main__":

submit.sh

@@ -1,8 +1,5 @@
 #!/bin/bash
-# python run.py fit --config experiments/config_4_layer_8_hidden.yaml
-# python run.py fit --config experiments/config_8_layer_8_hidden.yaml
-python run.py fit --config experiments/config_8_layer_16_hidden_refined.yaml 
-python run.py fit --config experiments/config_16_layer_8_hidden_refined.yaml
-python run.py fit --config experiments/config_16_layer_16_hidden_refined.yaml
-python run.py fit --config experiments/config_8_layer_16_hidden.yaml
-# python run.py fit --config experiments/config_4_layer_16_hidden.yaml
+export CUDA_VISIBLE_DEVICES=1
+python run.py fit --config experiments/5_steps/config_16_layer_16_hidden_refined.yaml
+python run.py fit --config experiments/5_steps/config_32_layer_16_hidden_refined.yaml
+python run.py fit --config experiments/5_steps/config_8_layer_16_hidden_refined.yaml