ValueError: Expected more than 1 value per channel when training, got input size torch.Size([1, 256, 1, 1])

[liheyoung]

ValueError: Expected more than 1 value per channel when training, got input size torch.Size([1, 256, 1, 1])

[LiheYoung]

You can change the ASPPPooling from

UniMatch/model/semseg/deeplabv3plus.py

Lines 80 to 91 in 7292bac

	class ASPPPooling(nn.Module):
	def __init__(self, in_channels, out_channels):
	super(ASPPPooling, self).__init__()
	self.gap = nn.Sequential(nn.AdaptiveAvgPool2d(1),
	nn.Conv2d(in_channels, out_channels, 1, bias=False),
	nn.BatchNorm2d(out_channels),
	nn.ReLU(True))
	def forward(self, x):
	h, w = x.shape[-2:]
	pool = self.gap(x)
	return F.interpolate(pool, (h, w), mode="bilinear", align_corners=True)

to

class ASPPPooling(nn.Module):
    def __init__(self, in_channels, out_channels):
        super(ASPPPooling, self).__init__()
        self.gap = nn.Sequential(nn.AdaptiveAvgPool2d(1),
                                 nn.Conv2d(in_channels, out_channels, 1, bias=False))
        self.post = nn.Sequential(nn.BatchNorm2d(out_channels),
                                  nn.ReLU(True))

    def forward(self, x):
        h, w = x.shape[-2:]
        pool = self.gap(x)
        return self.post(F.interpolate(pool, (h, w), mode="bilinear", align_corners=True))

[LiheYoung]

It is because the BN can not calculate the variance when the spatial size is 1x1.

[LiheYoung]

Did you use the supervised or fixmatch algorithm, instead of our unimatch? The unimatch should not encounter such a problem, because there are two concatenated strongly augmented images for each unlabeled sample. And if you used the former ones, you may solve this problem by updating the PyTorch version to 1.8.1.

[StevenLu1204]

My PyTorch version is 1.8.1+cu111, the problem remains. Do you mind sharing your log on training fixmatch.py ?

[StevenLu1204]

Thanks for the update. It works fine now. Looks like it's the problem of interpolation and BatchNorm2d?