[Model] NASA

examples/nasa/README.md

@@ -0,0 +1,21 @@
+# Regularizing GNNs via Consistency-Diversity Graph Augmentations (NASA)
+
+This example implements the model from the paper: [Regularizing Graph Neural Networks via Consistency-Diversity Graph Augmentations](https://arxiv.org/abs/2110.07627) (AAAI 2022).
+
+The implementation includes:
+- `NR_Augmentor`: A graph transformation that implements the "Neighbor Replacement" (NR) augmentation strategy.
+- `NASA_GCN`: A GCN-based model that incorporates the neighbor-constrained regularization loss (L_CR).
+
+## How to Run
+
+You can run the training script from the root directory of the GammaGL repository:
+
+```bash
+# Run on Cora dataset
+python examples/nasa/nasa_gcn_trainer.py --dataset Cora
+
+# Run on Citeseer dataset
+python examples/nasa/nasa_gcn_trainer.py --dataset Citeseer
+
+# Run on PubMed dataset
+python examples/nasa/nasa_gcn_trainer.py --dataset PubMed

examples/nasa/nasa_gcn_trainer.py

@@ -0,0 +1,153 @@
+import tensorlayerx as tlx
+import argparse
+import time
+
+from gammagl.data import Graph
+from gammagl.utils import mask_to_index
+from gammagl.datasets import Planetoid
+from gammagl.models import NASA_GCN
+from gammagl.transforms import NR_Augmentor
+from gammagl.utils import accuracy_tlx, compute_gcn_norm
+
+def main(args):
+
+    try:
+        tlx.set_device(device='GPU', id=args.gpu_id)
+        print(f"Using GPU: {args.gpu_id}")
+    except:
+        tlx.set_device(device='CPU')
+        print("GPU not available, using CPU.")
+
+    try:
+        dataset = Planetoid(root=args.dataset_path, name=args.dataset)
+    except Exception as e:
+        print(f"Error loading dataset {args.dataset}: {e}")
+        print("Please ensure the dataset name is correct (Cora, Citeseer, PubMed) and it's downloadable.")
+        return
+
+    graph = dataset[0]
+    num_nodes = graph.num_nodes
+
+    graph_x = tlx.convert_to_tensor(graph.x, dtype=tlx.float32)
+    graph_y = tlx.convert_to_tensor(graph.y, dtype=tlx.int64)
+    graph_train_mask = tlx.convert_to_tensor(graph.train_mask, dtype=tlx.bool)
+    graph_val_mask = tlx.convert_to_tensor(graph.val_mask, dtype=tlx.bool)
+    graph_test_mask = tlx.convert_to_tensor(graph.test_mask, dtype=tlx.bool)
+
+    eval_edge_index, eval_edge_weight = compute_gcn_norm(
+        graph.edge_index, num_nodes, dtype=graph_x.dtype, add_self_loops_flag=True
+    )
+    eval_edge_index = tlx.convert_to_tensor(eval_edge_index)
+    eval_edge_weight = tlx.convert_to_tensor(eval_edge_weight)
+
+    # initialize
+    augmentor = NR_Augmentor(probability=args.nr_prob)
+
+    model = NASA_GCN(
+        feature_dim=dataset.num_node_features,
+        hidden_dim=args.hidden_dim,
+        num_classes=dataset.num_classes,
+        dropout_rate=args.dropout,
+        temp=args.temp,
+        alpha=args.alpha
+    )
+
+    optimizer = tlx.optimizers.Adam(lr=args.lr, weight_decay=args.weight_decay)
+    train_weights = model.trainable_weights
+
+    best_val_acc = 0
+    best_test_acc = 0
+    best_epoch = 0
+
+    print("Starting training...")
+    for epoch in range(args.epochs):
+        epoch_start_time = time.time()
+        model.set_train()
+
+        # 1. Dynamic Augmentation
+        temp_original_graph_for_aug = Graph(x=graph_x, edge_index=graph.edge_index, num_nodes=num_nodes)
+        augmented_graph = augmentor.augment(temp_original_graph_for_aug)
+
+        # 2. Preprocess augmented graph for GCN
+        aug_edge_index, aug_edge_weight = compute_gcn_norm(
+            augmented_graph.edge_index, 
+            tlx.get_tensor_shape(augmented_graph.x)[0],
+            dtype=augmented_graph.x.dtype,
+            add_self_loops_flag=True
+        )
+        aug_edge_index = tlx.convert_to_tensor(aug_edge_index)
+        aug_edge_weight = tlx.convert_to_tensor(aug_edge_weight)
+
+        import tensorflow as tf
+        with tf.GradientTape() as tape:
+            output_logits_aug = model(
+                augmented_graph.x,
+                aug_edge_index,
+                edge_weight=aug_edge_weight,
+                num_nodes=tlx.get_tensor_shape(augmented_graph.x)[0]
+            )
+
+            loss = model.compute_nasa_loss(
+                output_logits_aug,
+                augmented_graph,
+                graph_y,
+                graph_train_mask
+            )
+
+        gradients = tape.gradient(loss, train_weights)
+        optimizer.apply_gradients(zip(gradients, train_weights))
+
+        # evaluation
+        model.set_eval()
+
+        eval_logits = model.predict(
+            graph_x,
+            eval_edge_index,
+            edge_weight=eval_edge_weight,
+            num_nodes=num_nodes
+        )
+        eval_pred_softmax = tlx.softmax(eval_logits, axis=-1)
+
+        train_acc = accuracy_tlx(tlx.gather(eval_pred_softmax, mask_to_index(graph_train_mask)), 
+                                 tlx.gather(graph_y, mask_to_index(graph_train_mask)))
+        val_acc = accuracy_tlx(tlx.gather(eval_pred_softmax, mask_to_index(graph_val_mask)),
+                               tlx.gather(graph_y, mask_to_index(graph_val_mask)))
+        test_acc = accuracy_tlx(tlx.gather(eval_pred_softmax, mask_to_index(graph_test_mask)),
+                                tlx.gather(graph_y, mask_to_index(graph_test_mask)))
+
+        epoch_duration = time.time() - epoch_start_time
+        print(f"Epoch {epoch+1:03d}/{args.epochs} | Loss: {loss.item():.4f} | "
+              f"Train Acc: {train_acc:.4f} | Val Acc: {val_acc:.4f} | Test Acc: {test_acc:.4f} | "
+              f"Time: {epoch_duration:.2f}s")
+
+        if val_acc > best_val_acc:
+            best_val_acc = val_acc
+            best_test_acc = test_acc
+            best_epoch = epoch +1
+            # save the best model
+            # model.save_weights(f"nasa_gcn_{args.dataset}_best.npz")
+            # print(f"New best validation accuracy: {best_val_acc:.4f}, saving model.")
+
+    print("Training finished.")
+    print(f"Best Epoch: {best_epoch}, Best Val Acc: {best_val_acc:.4f}, Corresponding Test Acc: {best_test_acc:.4f}")
+
+
+if __name__ == '__main__':
+    parser = argparse.ArgumentParser(description="NASA GCN training with GammaGL")
+    parser.add_argument('--dataset', type=str, default='Cora', help='Dataset name (Cora, Citeseer, PubMed)')
+    parser.add_argument('--dataset_path', type=str, default='./data', help='Path to store/load datasets')
+    parser.add_argument('--epochs', type=int, default=500, help='Number of training epochs')
+    parser.add_argument('--lr', type=float, default=0.01, help='Learning rate')
+    parser.add_argument('--weight_decay', type=float, default=1e-3, help='Weight decay for Adam optimizer')
+    parser.add_argument('--hidden_dim', type=int, default=32, help='Number of hidden units in GCN')
+    parser.add_argument('--dropout', type=float, default=0.7, help='Dropout rate for GCN layers and input features')
+    # NASA specific hyperparameters
+    parser.add_argument('--alpha', type=float, default=1.0, help='Weight for L_CR loss component')
+    parser.add_argument('--temp', type=float, default=0.5, help='Temperature for sharpening pseudo-labels')
+    parser.add_argument('--nr_prob', type=float, default=0.5, help='Probability for Neighbor Replacement')
+
+    parser.add_argument('--gpu_id', type=int, default=0, help='GPU ID to use, -1 for CPU')
+
+    cli_args = parser.parse_args()
+    print("Arguments:", cli_args)
+    main(cli_args)

gammagl/models/__init__.py

@@ -68,6 +68,7 @@
 from .adagad import PreModel, ReModel
 from .dyfss import MoeSSL,VGAE,Discriminator,InnerProductDecoder
 from .egt import EGTModel
+from .nasa_gcn import NASA_GCN
 
 __all__ = [
     'HeCo',
@@ -149,6 +150,7 @@
     'Discriminator',
     'InnerProductDecoder',
     'EGTModel',
+    'NASA_GCN',
 ]
 
 classes = __all__

gammagl/models/nasa_gcn.py

@@ -0,0 +1,87 @@
+import tensorlayerx as tlx
+from tensorlayerx.nn import Module as Model
+from gammagl.layers.conv import GCNConv
+from gammagl.utils import mask_to_index
+# from gammagl.mpops import unsorted_segment_mean
+
+class NASA_GCN(Model):
+    def __init__(self, feature_dim, hidden_dim, num_classes, dropout_rate, temp, alpha, name=None):
+        super().__init__(name=name)
+        self.conv1 = GCNConv(in_channels=feature_dim, out_channels=hidden_dim)
+        self.conv2 = GCNConv(in_channels=hidden_dim, out_channels=num_classes)
+        self.dropout = tlx.layers.Dropout(p=dropout_rate)
+        self.temp = temp
+        self.alpha = alpha
+        self.elu = tlx.layers.ELU()
+
+    def _compute_L_CR(self, aug_graph_edge_index, aug_pred_softmax, num_nodes):
+        src, dst = aug_graph_edge_index[0], aug_graph_edge_index[1]
+
+        # 1. Compute average of neighbor predictions (ŷ_i)
+        aug_pred_softmax_src = tlx.gather(aug_pred_softmax, src)
+        #avg_pred = unsorted_segment_mean(data=aug_pred_softmax_src, segment_ids=dst, num_segments=num_nodes)
+        avg_pred = tlx.ops.unsorted_segment_mean(aug_pred_softmax_src, dst, num_segments=num_nodes)
+
+        # Handle nodes with no incoming messages if unsorted_segment_mean results in NaN/Inf
+        #avg_pred = tlx.where(tlx.is_finite(avg_pred), avg_pred, tlx.zeros_like(avg_pred))
+        is_inf_avg_pred = tlx.is_inf(avg_pred)
+        is_nan_avg_pred = tlx.is_nan(avg_pred)
+        avg_pred_finite_mask = tlx.logical_and(
+            tlx.logical_not(is_inf_avg_pred),
+            tlx.logical_not(is_nan_avg_pred)
+        )
+        avg_pred = tlx.where(avg_pred_finite_mask, avg_pred, tlx.zeros_like(avg_pred))
+
+        # 2. Sharpening (p_i)
+        avg_pred_eps = avg_pred + 1e-12
+        pow_avg_pred = tlx.pow(avg_pred_eps, 1.0 / self.temp)
+        sharp_pseudo_labels = pow_avg_pred / (tlx.reduce_sum(pow_avg_pred, axis=1, keepdims=True) + 1e-12)
+        sharp_pseudo_labels_detached = tlx.ops.stop_gradient(sharp_pseudo_labels)
+
+        # 3. Compute KL Divergence Loss
+        p_dst_detached = tlx.gather(sharp_pseudo_labels_detached, dst)
+        q_src_softmax = tlx.gather(aug_pred_softmax, src)
+
+        log_p_dst_detached = tlx.log(p_dst_detached + 1e-12)
+        log_q_src_softmax = tlx.log(q_src_softmax + 1e-12)
+
+        kl_div_elements = p_dst_detached * (log_p_dst_detached - log_q_src_softmax)
+        kl_div_per_edge = tlx.reduce_sum(kl_div_elements, axis=1)
+
+        if tlx.get_tensor_shape(kl_div_per_edge)[0] > 0:
+            loss_cr = tlx.reduce_mean(kl_div_per_edge)
+        else:
+            loss_cr = tlx.convert_to_tensor(0.0, dtype=tlx.float32)
+
+        return loss_cr
+
+    def forward(self, x, edge_index, edge_weight=None, num_nodes=None):
+        h = self.dropout(x)
+        h = self.conv1(h, edge_index, edge_weight=edge_weight, num_nodes=num_nodes)
+        h = self.elu(h)
+        h = self.dropout(h)
+        logits = self.conv2(h, edge_index, edge_weight=edge_weight, num_nodes=num_nodes)
+        return logits
+
+    def compute_nasa_loss(self, output_logits_aug, augmented_graph, original_graph_labels, original_graph_train_mask):
+        train_indices = mask_to_index(original_graph_train_mask)
+        gathered_logits_aug = tlx.gather(output_logits_aug, train_indices)
+        gathered_labels = tlx.gather(original_graph_labels, train_indices)
+
+        loss_ce = tlx.losses.softmax_cross_entropy_with_logits(gathered_logits_aug, gathered_labels)
+
+        aug_pred_softmax = tlx.softmax(output_logits_aug, axis=-1)
+        loss_cr = self._compute_L_CR(
+            augmented_graph.edge_index,
+            aug_pred_softmax,
+            tlx.get_tensor_shape(augmented_graph.x)[0]
+        )
+
+        total_loss = loss_ce + self.alpha * loss_cr
+        return total_loss
+
+    def predict(self, x, edge_index, edge_weight=None, num_nodes=None):
+        h = self.conv1(x, edge_index, edge_weight=edge_weight, num_nodes=num_nodes)
+        h = self.elu(h)
+        logits = self.conv2(h, edge_index, edge_weight=edge_weight, num_nodes=num_nodes)
+        return logits

gammagl/transforms/__init__.py

@@ -7,6 +7,7 @@
 from .random_link_split import RandomLinkSplit
 from .vgae_pre import mask_test_edges, sparse_to_tuple
 from .svd_feature_reduction import SVDFeatureReduction
+from .nr_augmentor import NR_Augmentor
 
 __all__ = [
     'BaseTransform',
@@ -18,8 +19,8 @@
     'RandomLinkSplit',
     'mask_test_edges',
     'sparse_to_tuple',
-    'SVDFeatureReduction'
-
+    'SVDFeatureReduction',
+    'NR_Augmentor'
 ]
 
 classes = __all__

gammagl/transforms/nr_augmentor.py

@@ -0,0 +1,94 @@
+# gammagl/transforms/nr_augmentor.py
+
+import tensorlayerx as tlx
+import numpy as np
+import random
+from gammagl.data import Graph
+from gammagl.utils import to_undirected, coalesce
+
+class NR_Augmentor:
+    def __init__(self, probability=0.5):
+        """
+        Neighbor Replacement Augmentor for GammaGL.
+        This augmentor implements the NeighborReplace (NR) strategy from the paper
+        "Regularizing Graph Neural Networks via Consistency-Diversity Graph Augmentations".
+
+        Args:
+            probability (float): Probability of replacing a 1-hop neighbor with a 2-hop neighbor.
+        """
+        self.probability = probability
+
+    def _get_1hop_neighbors_dict(self, edge_index, num_nodes):
+        """
+        Creates an adjacency list dictionary {node_idx: [neighbor1, neighbor2,...]}
+        from edge_index, containing unique 1-hop neighbors (excluding self-loops).
+        """
+        adj = {i: set() for i in range(num_nodes)}
+        src_nodes, dst_nodes = tlx.convert_to_numpy(edge_index[0]), tlx.convert_to_numpy(edge_index[1])
+        for i in range(len(src_nodes)):
+            u, v = src_nodes[i], dst_nodes[i]
+            if u != v:
+                adj[u].add(v)
+        return {node_idx: list(neighbors) for node_idx, neighbors in adj.items()}
+
+    def __call__(self, original_graph: Graph) -> Graph:
+        """
+        Applies Neighbor Replacement augmentation to the input graph.
+        Making the class callable is a common pattern for transforms.
+
+        Args:
+            original_graph (gammagl.data.Graph): The original graph object.
+
+        Returns:
+            gammagl.data.Graph: The augmented graph object.
+        """
+        num_nodes = original_graph.num_nodes
+
+        adj_1hop_dict = self._get_1hop_neighbors_dict(original_graph.edge_index, num_nodes)
+
+        new_edge_src_list = []
+        new_edge_dst_list = []
+
+        for u_node_idx in range(num_nodes):
+            current_1hop_neighbors_of_u = adj_1hop_dict[u_node_idx]
+            if not current_1hop_neighbors_of_u:
+                continue
+
+            for v_neighbor_idx in current_1hop_neighbors_of_u:
+                if random.random() < self.probability:
+                    potential_2hop_neighbors_of_u_via_v = [
+                        vv_node for vv_node in adj_1hop_dict.get(v_neighbor_idx, [])
+                        if vv_node != u_node_idx and vv_node != v_neighbor_idx
+                    ]
+
+                    if potential_2hop_neighbors_of_u_via_v:
+                        vv_chosen = random.choice(potential_2hop_neighbors_of_u_via_v)
+                        new_edge_src_list.append(u_node_idx)
+                        new_edge_dst_list.append(vv_chosen)
+                    else:
+                        new_edge_src_list.append(u_node_idx)
+                        new_edge_dst_list.append(v_neighbor_idx)
+                else:
+                    new_edge_src_list.append(u_node_idx)
+                    new_edge_dst_list.append(v_neighbor_idx)
+
+        if not new_edge_src_list:
+            aug_edge_index = tlx.convert_to_tensor(np.array([[],[]]), dtype=tlx.int64)
+        else:
+            temp_edge_index = tlx.stack([
+                tlx.convert_to_tensor(new_edge_src_list, dtype=tlx.int64),
+                tlx.convert_to_tensor(new_edge_dst_list, dtype=tlx.int64)
+            ])
+
+            undirected_temp_edge_index = to_undirected(temp_edge_index, num_nodes=num_nodes)
+            aug_edge_index = coalesce(undirected_temp_edge_index, num_nodes=num_nodes)
+
+        aug_graph = Graph(x=original_graph.x, edge_index=aug_edge_index, num_nodes=num_nodes)
+
+        # Copy other essential attributes if they exist
+        if hasattr(original_graph, 'y'): aug_graph.y = original_graph.y
+        if hasattr(original_graph, 'train_mask'): aug_graph.train_mask = original_graph.train_mask
+        if hasattr(original_graph, 'val_mask'): aug_graph.val_mask = original_graph.val_mask
+        if hasattr(original_graph, 'test_mask'): aug_graph.test_mask = original_graph.test_mask
+
+        return aug_graph