yh-yao / fedgcn

Official Code for FedGCN [NeurIPS 2023]

Hi there 👋 I am Yuhang Yao

• 🔭 I’m currently working on federated graph learning and generative ai.
• 💬 My homepage is https://www.yuhangyao.com.
• 📫 If you have any question on my research or project, e-mail me at [email protected].

Look at my recent works:
https://github.com/FedGraph/fedgraph
https://github.com/yh-yao/FedGCN
https://github.com/yh-yao/FedRule

fedgcn's Issues

how to get x,y,tx,ty,allx,ally,graph

Hi, how should i get x,y,tx,ty,allx,ally,graph? thk

Dirichlet data distributions

尊敬的作者您好，恭喜中稿NIPS，很感谢您愿意开源代码~
有一个关于客户端数据划分的问题想请教一下：

在您的论文中提到了按Dirichlet数据分布来模拟非独立同分布的情况，但是在main.ipynb文件里我似乎没有找到对应的代码，我不确定是不是我找错位置了，或者没能理解代码含义

main.ipynb中关于client数据划分的实现似乎对应于Lhop_Block_federated_GCN函数中的如下代码：

        nclass=labels.max().item() + 1
        split_data_indexes = []
        non_iid_percent = 1 - float(iid_percent)
        iid_indexes = [] #random assign
        shuffle_labels = [] #make train data points split into different devices
        for i in range(K):
            current = torch.nonzero(labels == i).reshape(-1)
            current = current[np.random.permutation(len(current))] #shuffle
            shuffle_labels.append(current)
        average_device_of_class = K // nclass
        if K % nclass != 0: #for non-iid
            average_device_of_class += 1
        for i in range(K):  
            label_i= i // average_device_of_class    
            labels_class = shuffle_labels[label_i]
            average_num= int(len(labels_class)//average_device_of_class * non_iid_percent)
            split_data_indexes.append((labels_class[average_num * (i % average_device_of_class):average_num * (i % average_device_of_class + 1)]))
        
        if args_cuda:
            iid_indexes = setdiff1d(torch.tensor(range(len(labels))).cuda(), torch.cat(split_data_indexes))
        else:
            iid_indexes = setdiff1d(torch.tensor(range(len(labels))), torch.cat(split_data_indexes))
        iid_indexes = iid_indexes[np.random.permutation(len(iid_indexes))]
        
        for i in range(K):  #for iid
            label_i= i // average_device_of_class
            labels_class = shuffle_labels[label_i]

            average_num= int(len(labels_class)//average_device_of_class * (1 - non_iid_percent))
            split_data_indexes[i] = list(split_data_indexes[i]) + list(iid_indexes[:average_num])
                    
            iid_indexes = iid_indexes[average_num:]
            
        communicate_indexes = []
        in_com_train_data_indexes = []

但上面的代码里似乎不涉及从迪利克雷分布中采样的操作（在我的认知里大家通常会用np.random.dirichlet函数），Lhop_Block_federated_GCN函数参数也不包括论文中提到的β值。

111

Stack (most recent call first):
File "C:\1 New Files\anaconda\envs\Fed\lib\ctypes_init_.py", line 374 in init
File "C:\1 New Files\anaconda\envs\Fed\lib\site-packages\torch_ops.py", line 1032 in load_library
File "C:\1 New Files\anaconda\envs\Fed\lib\site-packages\torch_scatter_init_.py", line 16 in
File "", line 241 in _call_with_frames_removed
File "", line 883 in exec_module
File "", line 688 in _load_unlocked
File "", line 1006 in _find_and_load_unlocked
File "", line 1027 in _find_and_load
File "C:\1 New Files\anaconda\envs\Fed\lib\site-packages\torch_geometric\typing.py", line 68 in
File "", line 241 in _call_with_frames_removed
File "", line 883 in exec_module
File "", line 688 in _load_unlocked
File "", line 1006 in _find_and_load_unlocked
File "", line 1027 in _find_and_load
File "C:\1 New Files\anaconda\envs\Fed\lib\site-packages\torch_geometric_compile.py", line 6 in
File "", line 241 in _call_with_frames_removed
File "", line 883 in exec_module
File "", line 688 in _load_unlocked
File "", line 1006 in _find_and_load_unlocked
File "", line 1027 in find_and_load
File "C:\1 New Files\anaconda\envs\Fed\lib\site-packages\torch_geometric_init.py", line 1 in
File "", line 241 in _call_with_frames_removed
File "", line 883 in exec_module
File "", line 688 in _load_unlocked
File "", line 1006 in _find_and_load_unlocked
File "", line 1027 in _find_and_load
File "C:\Users\admin1\Desktop\FedGCN-master\src\data_process.py", line 11 in
File "", line 241 in _call_with_frames_removed
File "", line 883 in exec_module
File "", line 688 in _load_unlocked
File "", line 1006 in _find_and_load_unlocked
File "", line 1027 in _find_and_load
File "C:\Users\admin1\Desktop\FedGCN-master\src\fedgcn_run.py", line 25 in
C:\1 New Files\anaconda\envs\Fed\lib\site-packages\torch_geometric\typing.py:72: UserWarning: An issue occurred while importing 'torch-scatter'. Disabling its usage. Stacktrace: [WinError 127] 找不到指定的程序。
warnings.warn(f"An issue occurred while importing 'torch-scatter'. "
Windows fatal exception: code 0xc0000139