FlowNet3D模型初始化时出现TypeError的原因及解决方法

报错信息 #

Namespace(batch_size=64, cycle=False, dataset='SceneflowDataset', dataset_path='data/data_processed_maxcut_35_20k_2k_8192/', dropout=0.5, emb_dims=512, epochs=250, eval=False, exp_name='flownet3d', gaussian_noise=False, lr=0.001, model='flownet', model_path='', momentum=0.9, no_cuda=False, num_points=2048, seed=1234, test_batch_size=32, unseen=False, use_sgd=False)
    train :  20006
    test :  2007
    Traceback (most recent call last):
    File "main.py", line 282, in <module>
        main()
    File "main.py", line 254, in main
        net = FlowNet3D(args).cuda()
    File "/home/ubuntu/project/flownet3d_pytorch/model.py", line 13, in __init__
        self.sa1 = PointNetSetAbstraction(npoint=1024, radius=0.5, nsample=16, in_channel=3, mlp=[32,32,64], group_all=False)
    File "/home/ubuntu/project/flownet3d_pytorch/util.py", line 225, in __init__
        for out_channel in mlp2:
    TypeError: 'NoneType' object is not iterable

问题原因 #

PointNetSetAbstraction类的__init__方法中，mlp2参数默认值设为None，但代码直接对mlp2执行遍历操作。当实例化该类（如sa1至sa4）时未传入mlp2参数，就会触发None不可迭代的错误。

解决方案 #

方案一：修改默认参数为空列表 #

将mlp2的默认值从None改为空列表，这样即使不传入该参数，遍历操作也能正常执行：

class PointNetSetAbstraction(nn.Module):
    def __init__(self, npoint, radius, nsample, in_channel, mlp, mlp2 = [], group_all = False):
        super(PointNetSetAbstraction, self).__init__()
        self.npoint = npoint
        self.radius = radius
        self.nsample = nsample
        self.group_all = group_all
        self.mlp_convs = nn.ModuleList()
        self.mlp_bns = nn.ModuleList()
        self.mlp2_convs = nn.ModuleList()
        last_channel = in_channel+3
        for out_channel in mlp:
            self.mlp_convs.append(nn.Conv2d(last_channel, out_channel, 1, bias = False))
            self.mlp_bns.append(nn.BatchNorm2d(out_channel))
            last_channel = out_channel
        for out_channel in mlp2:
            self.mlp2_convs.append(nn.Sequential(nn.Conv1d(last_channel, out_channel, 1, bias=False),
                                                nn.BatchNorm1d(out_channel)))
            last_channel = out_channel
        # 后续代码不变

方案二：增加非空判断 #

在遍历mlp2前添加判断，仅当mlp2不为None时执行循环：

class PointNetSetAbstraction(nn.Module):
    def __init__(self, npoint, radius, nsample, in_channel, mlp, mlp2 = None, group_all = False):
        super(PointNetSetAbstraction, self).__init__()
        self.npoint = npoint
        self.radius = radius
        self.nsample = nsample
        self.group_all = group_all
        self.mlp_convs = nn.ModuleList()
        self.mlp_bns = nn.ModuleList()
        self.mlp2_convs = nn.ModuleList()
        last_channel = in_channel+3
        for out_channel in mlp:
            self.mlp_convs.append(nn.Conv2d(last_channel, out_channel, 1, bias = False))
            self.mlp_bns.append(nn.BatchNorm2d(out_channel))
            last_channel = out_channel
        # 修改此处，增加非空判断
        if mlp2 is not None:
            for out_channel in mlp2:
                self.mlp2_convs.append(nn.Sequential(nn.Conv1d(last_channel, out_channel, 1, bias=False),
                                                    nn.BatchNorm1d(out_channel)))
                last_channel = out_channel
        # 后续代码不变

两种方案均可解决问题：方案一操作简单，适配原项目多数场景；方案二更严谨，能避免意外传入None的情况。

内容的提问来源于stack exchange，提问作者kaka2u_