Input to InstanceNorm1d is not in the right shape (?)

[anhquanpham]

Size & shapes

self.nn_human_x.shape: torch.Size([6, 1536])
self.nn_human_x Parameter containing:
tensor([[0., 0., 0., ..., 0., 0., 0.],
[0., 0., 0., ..., 0., 0., 0.],
[0., 0., 0., ..., 0., 0., 0.],
[0., 0., 0., ..., 0., 0., 0.],
[0., 0., 0., ..., 0., 0., 0.],
[0., 0., 0., ..., 0., 0., 0.]])
self.nn_human_x[i] tensor([0., 0., 0., ..., 0., 0., 0.])
Shape x [i]: torch.Size([1536])
trans_nn_human_x []

Errors

C:\Users\phama\AppData\Local\Temp\ipykernel_13640\996063355.py:62: UserWarning: To copy construct from a tensor, it is recommended to use sourceTensor.clone().detach() or sourceTensor.clone().detach().requires_grad_(True), rather than torch.tensor(sourceTensor).
trans_nn_human_x.append( t( torch.tensor(self.nn_human_x[i], dtype=torch.float32).view(1, -1)) )

ValueError Traceback (most recent call last)
~\AppData\Local\Temp\ipykernel_13640\1820000454.py in
77
78 model.eval()
---> 79 current_acc = evaluate_loader(model, train_loader, cce_loss)
80 model.train()
81

~\AppData\Local\Temp\ipykernel_13640\1992294143.py in evaluate_loader(model, loader, cce_loss)
8 imgs, labels = data
9 imgs, labels = imgs.to(DEVICE), labels.to(DEVICE)
---> 10 logits = model(imgs)
11 loss = cce_loss(logits, labels)
12 preds = torch.argmax(logits, dim=1)

~\anaconda3\lib\site-packages\torch\nn\modules\module.py in _call_impl(self, *input, **kwargs)
1100 if not (self._backward_hooks or self._forward_hooks or self._forward_pre_hooks or _global_backward_hooks
1101 or _global_forward_hooks or _global_forward_pre_hooks):
-> 1102 return forward_call(*input, **kwargs)
1103 # Do not call functions when jit is used
1104 full_backward_hooks, non_full_backward_hooks = [], []

~\AppData\Local\Temp\ipykernel_13640\996063355.py in forward(self, x)
60 print("trans_nn_human_x", trans_nn_human_x)
61 # trans_nn_human_x.append(t(self.nn_human_x[i]).view(1, -1))
---> 62 trans_nn_human_x.append( t( torch.tensor(self.nn_human_x[i], dtype=torch.float32).view(1, -1)) )
63 # trans_nn_human_x.append(t(self.nn_human_x[i].clone().detach().to(torch.float32).view(1, -1)))
64

~\anaconda3\lib\site-packages\torch\nn\modules\module.py in _call_impl(self, *input, **kwargs)
1100 if not (self._backward_hooks or self._forward_hooks or self._forward_pre_hooks or _global_backward_hooks
1101 or _global_forward_hooks or _global_forward_pre_hooks):
-> 1102 return forward_call(*input, **kwargs)
1103 # Do not call functions when jit is used
1104 full_backward_hooks, non_full_backward_hooks = [], []

~\anaconda3\lib\site-packages\torch\nn\modules\container.py in forward(self, input)
139 def forward(self, input):
140 for module in self:
--> 141 input = module(input)
142 return input
143

~\anaconda3\lib\site-packages\torch\nn\modules\module.py in _call_impl(self, *input, **kwargs)
1100 if not (self._backward_hooks or self._forward_hooks or self._forward_pre_hooks or _global_backward_hooks
1101 or _global_forward_hooks or _global_forward_pre_hooks):
-> 1102 return forward_call(*input, **kwargs)
1103 # Do not call functions when jit is used
1104 full_backward_hooks, non_full_backward_hooks = [], []

~\anaconda3\lib\site-packages\torch\nn\modules\instancenorm.py in forward(self, input)
54
55 def forward(self, input: Tensor) -> Tensor:
---> 56 self._check_input_dim(input)
57 return F.instance_norm(
58 input, self.running_mean, self.running_var, self.weight, self.bias,

~\anaconda3\lib\site-packages\torch\nn\modules\instancenorm.py in _check_input_dim(self, input)
130 def _check_input_dim(self, input):
131 if input.dim() == 2:
--> 132 raise ValueError(
133 'InstanceNorm1d returns 0-filled tensor to 2D tensor.'
134 'This is because InstanceNorm1d reshapes inputs to'

ValueError: InstanceNorm1d returns 0-filled tensor to 2D tensor.This is because InstanceNorm1d reshapes inputs to(1, N * C, ...) from (N, C,...) and this makesvariances 0.

Problem

It seems like the data copy part

Code

for i in range(NUM_CLASSES):
    idxs = a_train == i
    temp_x = X_train[idxs]
    mean = temp_x.mean(axis=0)
    knn = KNeighborsClassifier().fit(temp_x, list(range(len(temp_x))))
    idx = knn.kneighbors(X=mean.reshape(1,-1), n_neighbors=1, return_distance=False)
    p_idxs.append(idx.item())
    nn_human_x.append( temp_x[idx.item()].tolist() )
nn_human_x = np.array(nn_human_x)



#### Training
model = PWNet().eval()
model.nn_human_x.data.copy_( torch.tensor(nn_human_x) )

It is passing from NUM_CLASSES x LATENT_SIZE to NUM_PROTOTYPES x LATENT_SIZE. The InstanceNorm1d should be normalizing amongst the CLASSES (I think?).
When going through the transformation loop, the x[i] shape reduce to 1D, LATENT_SIZE only. Should this nn_human_x shape be NUM_PROTOTYPES x NUM_CLASSES x LATENT_SIZE instead, then passing through the transformation loop its gonna be NUM_CLASSES x LATENT_SIZE, then reaching the InstanceNorm1d it will be NUM_CLASSES x LATENT_SIZE?

Sorry if I understand it incorrectly, really look forward to your help.

[EoinKenny]

Hey sorry for the late reply! Do you still need any help?

Which script is this exactly? Which dataset and baseline? Thanks for letting me know about this.