I find the loss of ResNet composed of MetaModule is difficult to decrease, and the regular ResNet behaves normally, regardless of whether or not to do Kaiming initialization. Has anyone encountered the same problem？

I'm having problem: TypeError: cannot assign 'torch.cuda.FloatTensor' as parameter 'weight' (torch.nn.Parameter or None expected)

When I'm trying to implement resnet experiment. I saw instruction in another Issue page and follow to change nn.Module to MetaModule and linear to MetaLinear but there is still that same error. Do I need to change anything else?

Also, I don't see any line of code where those PreAct layer being used. Please correct me if I'm wrong

hello ! danieltan! when i transfer your minist image to my CWRU(1-dimension)data. I meet this problem. i rewrite the LENET model because for the feature encoder is different between image and 1d-data. and i don't use your metamodel class. when i calculate the grad for the first virtual updata . this bug happends. shoud i use your metamodel class? should i write a class of meteconv1d? i can't find the reason of this bug. expect you suggestion

If Adam optimizer is used, can it still work ? (Line 7. indicates a standard gradient desent method) Or, this just fit into the SGD based optimizer? 2. I would like to use this reweighting strategy in more complicated neural framework such as LSTM, BERT for other downstream tasks. Whether I must modify these to the 'meta-style' structures ? It seems to be trivial.

Here is the code:

    def __init__(self, num_classes, block=BasicBlock, num_blocks=[5, 5, 5]):
        super(ResNet32, self).__init__()
        self.in_planes = 16

        self.conv1 = MetaConv2d(3, 16, kernel_size=3, stride=1, padding=1, bias=False)
        self.bn1 = MetaBatchNorm2d(16)
        self.layer1 = self._make_layer(block, 16, num_blocks[0], stride=1)
        self.layer2 = self._make_layer(block, 32, num_blocks[1], stride=2)
        self.layer3 = self._make_layer(block, 64, num_blocks[2], stride=2)
        self.linear = MetaLinear(64, num_classes)

        self.apply(_weights_init)

    def _make_layer(self, block, planes, num_blocks, stride):
        strides = [stride] + [1]*(num_blocks-1)
        layers = []
        for stride in strides:
            layers.append(block(self.in_planes, planes, stride))
            self.in_planes = planes * block.expansion

        return nn.Sequential(*layers)

    def forward(self, x):
        out = F.relu(self.bn1(self.conv1(x)))
        out = self.layer1(out)
        out = self.layer2(out)
        out = self.layer3(out)
        out = F.avg_pool2d(out, out.size()[3])
        out = out.view(out.size(0), -1)
        out = self.linear(out)
        return out

model = ResNet32()```
When I use 
`for name, p in list(model.named_params()):`
it returns   'NoneType' object has no attribute '_parameters'

Thank you!

I build a resnet18 transfer learning model following the example in weblink: https://pytorch.org/tutorials/beginner/transfer_learning_tutorial.html.

My program encounters the error when running to: meta_model.update_params(hyperparameters['lr'], source_params=grads) The error is: TypeError: cannot assign 'torch.cuda.FloatTensor' as parameter 'weight' (torch.nn.Parameter or None expected)

Anyone has ideas about the error I encountered? Thanks

Could you give the exact link for: class MetaModule(nn.Module): ? I have not found it from: Adrien Ecoffet https://github.com/AdrienLE

Hi! Thanks for your implementation. But it seems that there are some problems.

1. class PreActResNetMeta(nn.Module) -> class PreActResNetMeta(MetaModule) and thus accordingly parameters() -> params() for model optimizer in main funciton
2. nn.linear() -> MeataLinear()

And for preact resnet this line out = F.adaptive_max_pool2d(out, 1), it seems that adaptive_max_pool2d() is uncommonly used.