heart mri example test pass!

Author: yifanfeng97

.idea/HyperG_package.iml

@@ -22,30 +22,30 @@ def reset_parameters(self):
 
     def gen_hyedge_ft(self, x: torch.Tensor, H: torch.Tensor, hyedge_weight=None):
         ft_dim = x.size(1)
-        node_idx, edge_idx = H
+        node_idx, hyedge_idx = H
         hyedge_num = count_hyedge(H)
 
         # a vector to normalize hyperedge feature
         hyedge_norm = 1.0 / degree_hyedge(H).float()
         if hyedge_weight is not None:
             hyedge_norm *= hyedge_weight
-        hyedge_norm = hyedge_norm[edge_idx]
+        hyedge_norm = hyedge_norm[hyedge_idx]
 
         x = x[node_idx] * hyedge_norm.unsqueeze(1)
-        x = torch.zeros(hyedge_num, ft_dim).scatter_add(0, edge_idx.unsqueeze(1).repeat(1, ft_dim), x)
+        x = torch.zeros(hyedge_num, ft_dim).to(x.device).scatter_add(0, hyedge_idx.unsqueeze(1).repeat(1, ft_dim), x)
         return x
 
     def gen_node_ft(self, x: torch.Tensor, H: torch.Tensor):
         ft_dim = x.size(1)
-        node_idx, edge_idx = H
+        node_idx, hyedge_idx = H
         node_num = count_node(H)
 
         # a vector to normalize node feature
         node_norm = 1.0 / degree_node(H).float()
-        node_norm = node_norm[edge_idx]
+        node_norm = node_norm[node_idx]
 
-        x = x[edge_idx] * node_norm.unsqueeze(1)
-        x = torch.zeros(node_num, ft_dim).scatter_add(0, node_idx.unsqueeze(1).repeat(1, ft_dim), x)
+        x = x[hyedge_idx] * node_norm.unsqueeze(1)
+        x = torch.zeros(node_num, ft_dim).to(x.device).scatter_add(0, node_idx.unsqueeze(1).repeat(1, ft_dim), x)
         return x
 
     def forward(self, x: torch.Tensor, H: torch.Tensor, hyedge_weight=None):

.idea/deployment.xml

@@ -6,10 +6,15 @@
 def degree_node(H):
     node_idx, edge_idx = H
     node_num = count_node(H)
-    return torch.zeros(node_num).scatter_add(0, node_idx, torch.ones_like(node_idx).float()).long()
+    src = torch.ones_like(node_idx).float().to(H.device)
+    out = torch.zeros(node_num).to(H.device)
+    return out.scatter_add(0, node_idx, src).long()
+    # return torch.zeros(node_num).scatter_add(0, node_idx, torch.ones_like(node_idx).float()).long()
 
 
-def degree_hyedge(H):
+def degree_hyedge(H: torch.Tensor):
     node_idx, hyedge_idx = H
     edge_num = count_hyedge(H)
-    return torch.zeros(edge_num).scatter_add(0, hyedge_idx, torch.ones_like(hyedge_idx).float()).long()
+    src = torch.ones_like(hyedge_idx).float().to(H.device)
+    out = torch.zeros(edge_num).to(H.device)
+    return out.scatter_add(0, hyedge_idx, src).long()

.idea/misc.xml

@@ -2,21 +2,22 @@
 
 import torch
 
-from HyperG.hyedge import count_hyedge, count_node
+from HyperG.hyedge import count_hyedge, count_node, contiguous_hyedge_idx
 
 
 def hyedge_concat(Hs: Union[Tuple[torch.Tensor, ...], List[torch.Tensor]], same_node=True):
     node_num = 0
     hyedge_num = 0
     Hs_new = []
     for H in Hs:
+        _H = H.clone()
         if not same_node:
-            H[0, :] += node_num
-        H[1, :] += hyedge_num
+            _H[0, :] += node_num
+        _H[1, :] += hyedge_num
 
-        Hs_new.append(H)
+        Hs_new.append(_H)
 
         hyedge_num += count_hyedge(H)
         node_num += count_node(H)
-
-    return torch.cat(Hs_new, dim=1)
+    Hs_new = torch.cat(Hs_new, dim=1)
+    return contiguous_hyedge_idx(Hs_new)

HyperG/conv/hyconv.py

@@ -16,8 +16,8 @@ def normalize(x):
 
 
 def split_train_val(root, ratio=0.8, save_dir=None, resplit=False):
-    if not resplit and save_dir is not None:
-        with open(save_dir, 'r') as f:
+    if not resplit and save_dir is not None and osp.exists(save_dir):
+        with open(save_dir, 'rb') as f:
             result = pickle.load(f)
         return result
 
@@ -44,7 +44,7 @@ def split_train_val(root, ratio=0.8, save_dir=None, resplit=False):
         save_folder = osp.split(save_dir)[0]
         if not osp.exists(save_folder):
             os.makedirs(save_folder)
-        with open(save_dir, 'w') as f:
+        with open(save_dir, 'wb') as f:
             pickle.dump(result, f)
 
     return result
@@ -70,12 +70,12 @@ def preprocess(data_list, patch_size, k_nearest):
         lbl.append(_lbl)
         mask_train.extend([0] * _node_num)
 
-    x, lbl = torch.cat(x, dim=0), torch.cat(lbl, dim=0)
+    x, lbl = torch.cat(x, dim=0), torch.cat(lbl, dim=0).long()
     x = normalize(x)
 
     H_grid = hyedge_concat(H_grid, same_node=False)
-    mask_train = torch.tensor(mask_train)
-    mask_val = 1 - mask_train
+    mask_train = torch.tensor(mask_train).bool()
+    mask_val = ~mask_train
 
     H_global = neighbor_distance(x, k_nearest)
 

HyperG/hyedge/utils/degree.py

@@ -1,19 +1,19 @@
 import torch
 import torch.nn.functional as F
+from data_helper import preprocess, split_train_val
 
 from HyperG.models import HGNN
-from .data_helper import preprocess, split_train_val
 
 device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
 
 k_nearest = 7
 patch_size = (5, 5)
-data_list = split_train_val('/repository/HyperG_example/example_data/heart_mri/processed',
-                            save_dir='/repository/HyperG_example/tmp/heart_mri', ratio=0.8)
+data_list = split_train_val('/repository/HyperG_example/example_data/heart_mri/processed', ratio=0.8,
+                            save_dir='/repository/HyperG_example/tmp/heart_mri/split.pkl', resplit=True)
 x, H, lbl, mask_train, mask_val = preprocess(data_list, patch_size, k_nearest)
 
 x_ch = x.size(1)
-n_class = lbl.max() + 1
+n_class = lbl.max().item() + 1
 model = HGNN(x_ch, n_class, hidens=[16])
 
 model, x, H, lbl, mask_train, mask_val = model.to(device), x.to(device), H.to(device), \
@@ -44,9 +44,9 @@ def val():
 
 if __name__ == '__main__':
     best_acc = 0.0
-    for epoch in range(1, 51):
+    for epoch in range(1, 21):
         train()
         train_acc, val_acc = val()
         if val_acc > best_acc:
             best_acc = val_acc
-        print(f'Epoch: {epoch}, Train:{train_acc:.4f}, Val:{val_acc:.4f}, Best Val:{best_acc}:.4f')
+        print(f'Epoch: {epoch}, Train:{train_acc:.4f}, Val:{val_acc:.4f}, Best Val:{best_acc:.4f}')

HyperG/hygraph/fusion.py

@@ -1,4 +1,5 @@
 import torch
+
 from HyperG.models import HGNN