Hello,

I am trying a simple code like

x_t, y_t = torch.ones(4, 14).int(), torch.ones(4)
net = Dnn(feature_columns)
inner_opt = torchopt.MetaAdam(net, lr=1e-2)
net.train()
for i in range(10):
    loss = loss_func(net(x_t), y_t.view(-1, 1))
    inner_opt.step(loss)
    print(loss)

I expect that inner_opt.step(loss) would update net. However, from the loss values printed out, it seems not be the case.
The loss does not decrease. However, if I replace it with torch.optim.Adam, it seems to work fine.

I meant to have something similar to the example
https://github.com/metaopt/torchopt/blob/main/examples/few-shot/maml_omniglot.py#L155-L158

I am not sure if i have made any mistakes here

Please find this notebook for details and preproduction
https://colab.research.google.com/drive/1h_S0G8ibLFl8B2iRTQ6sIyvr01E3WJbG?usp=sharing