wip

examples/ijcai_example.py

@@ -32,7 +32,7 @@
     choices=["contextgnn", "idgnn", "shallowrhsgnn"],
 )
 parser.add_argument("--lr", type=float, default=0.001)
-parser.add_argument("--epochs", type=int, default=10)
+parser.add_argument("--epochs", type=int, default=100)
 parser.add_argument("--eval_epochs_interval", type=int, default=1)
 parser.add_argument("--batch_size", type=int, default=512)
 parser.add_argument("--channels", type=int, default=128)
@@ -137,6 +137,7 @@ def create_edge(data, behavior, beh_idx, pkey_name, pkey_idx):
 
 col_stats_dict = {}
 dst_nodes = None
+trnLabel = None
 for i in range(len(behs)):
     behavior = behs[i]
     with open(osp.join(path, 'trn_' + behavior), 'rb') as fs:
@@ -166,6 +167,7 @@ def create_edge(data, behavior, beh_idx, pkey_name, pkey_idx):
     coo_mat = sp.coo_matrix(mat)
     if behavior == 'buy':
         dst_nodes = coo_mat
+        trnLabel = 1 * (mat != 0)
     beh_idx = torch.arange(len(coo_mat.data), dtype=torch.long)
     create_edge(data, behavior, beh_idx, src_entity_table,
                 torch.tensor(coo_mat.row, dtype=torch.long))
@@ -181,14 +183,13 @@ def create_edge(data, behavior, beh_idx, pkey_name, pkey_idx):
 loader_dict: Dict[str, NeighborLoader] = {}
 dst_nodes_dict = {}
 split_date: Dict[str, int] = {}
-split_date['train'] = 1103
-split_date['val'] = 1110
+split_date['train'] = 1110
 split_date['test'] = 1111
 
 num_src_nodes = data[src_entity_table].num_nodes
 num_dst_nodes = data[dst_entity_table].num_nodes
 
-for split in ["train", "val", "test"]:
+for split in ["train", "test"]:
     dst_nodes_data = dst_nodes.data < split_date[split]
     dst_nodes_dict[split] = torch.sparse_coo_tensor(
         torch.stack([torch.tensor(dst_nodes.row),
@@ -296,8 +297,12 @@ def train() -> float:
             loss = F.binary_cross_entropy_with_logits(out, target)
             numel = out.numel()
         elif args.model in ['contextgnn', 'shallowrhsgnn']:
+            sampled_dst = torch.unique(dst_index)
             logits = model(batch, src_entity_table, dst_entity_table)
+            logits = logits[:, sampled_dst]
             edge_label_index = torch.stack([src_batch, dst_index], dim=0)
+            import pdb
+            pdb.set_trace()
             loss = sparse_cross_entropy(logits, edge_label_index)
             numel = len(batch[dst_entity_table].batch)
         loss.backward()
@@ -346,6 +351,11 @@ def test(loader: NeighborLoader, desc: str, target = None) -> np.ndarray:
             # randomly select num_item indices
             batch_user = scores.shape[0]
             num_item = scores.shape[1]
+            trnLabel[batch_user]
+            import pdb
+            # full set is the sampled rhs set
+            # you pick 99 items from the sampled rhs set. 
+            pdb.set_trace()
             random_items = torch.randint(0, num_item, (batch_user, 100)).to(scores.device)  # Shape: (batch_user, 100)
             for i in range(batch_size):
                 user_idx = batch[src_entity_table].n_id[i]
@@ -354,8 +364,6 @@ def test(loader: NeighborLoader, desc: str, target = None) -> np.ndarray:
                     random_items[i, 0] = target_item
 
             selected_scores = torch.gather(scores, 1, random_items)
-            import pdb
-            pdb.set_trace()
             _, top_k_indices = torch.topk(selected_scores, args.eval_k, dim=1)  # Shape: (num_user, args.eval_k)
 
             pred_mini = random_items[top_k_indices.tolist()]