Is StackedLSTMCell correct implementation?

[jyh2986]

For enas, StackedLSTMCell is implemented as below

nni/src/sdk/pynni/nni/nas/pytorch/enas/mutator.py

Lines 23 to 28 in bf8be1e

	curr_c, curr_h = m(inputs, (prev_c[i], prev_h[i]))
	next_c.append(curr_c)
	next_h.append(curr_h)
	# current implementation only supports batch size equals 1,
	# but the algorithm does not necessarily have this limitation
	inputs = curr_h[-1].view(1, -1)

nn.LSTMCell outputs (hidden state, cell state) but the order of output is changed in the implementation.
As a result, the cell state is used for output of LSTM.

Is this an intention or wrong implementation?

[ultmaster]

Hi @jyh2986. Thanks for asking.

If you are right, the implementation is definitely buggy. You can raise a PR. Meanwhile, I'll check the PyTorch documentation for verification. Thanks.