pyTorch autoencoder for unsupervised classification: loss not changing

Question

I'm new to pyTorch. I want to use the autoencoder concept to get an unsupervised classification. It seems like you should be able to use the minimum dimension from the autoencoder as input to a softmax fx. I used a "dummy" copy of the data to implement the loss fx. However, I can't get the loss to change and I don't understand why.

x = pd.read_csv(<file.csv>)
xray=x.to_numpy()
tx = torch.from_numpy(xray)
txr = tx.float()
txr.requires_grad=True
dmmy=torch.nn.Linear(22,6)(txr)
class AE(nn.Module):
    `def __init__(self):`
        `torch.autograd
        self.L1=nn.Linear(in_features=22, out_features=20)
        self.E1=nn.ELU(alpha=9)                          
        ...
        ...
        self.L7=nn.Linear(in_features=8, out_features=6)
        self.E7=nn.ELU(alpha=9)`
    `def forward(self, x: torch.Tensor, retain_graph=True):`
         `torch.autograd
          L1 = self.L1(x)
          E1 = self.E1(L1)
          ...
          ... 
          L7 = self.L7(E6)
          E7 = self.E7(L7)
          return E7`
Lossfx = torch.nn.MSELoss()
opt=torch.optim.SGD(params=m.parameters(), lr=0.01)
epochs=2
for epoch in range(epochs):
    `m.train()
    pred=m(txr)
    loss=Lossfx(dmmy,pred)
    loss.backward(retain_graph=True)
    print(f'loss: {loss}')
    print(f'pred: {pred[0]}')
    opt.zero_grad()
    opt.step()`
print('Done')
**## Softmax after training/testing e.g. ypred=nn.Softmax(dim=1)(pred)**
## Result
Loss: 38.851
pred[0]: [ 4.6630, -8.9925, -8.9977,  6.2631,  2.2962,  2.2374]
dmmy[0]: [-0.1594, -0.2595, -0.2731,  0.2965, -0.0969,  0.5345]

Loss: 38.851
pred[0]: same
dmmy[0]: same