empty shapes returned in RNN example

[nicklhy]

Following the code in example/rnn/rnn_cell_demo.py, I create a RNN symbol like this:

data = mx.sym.Variable('data')
rnn_h_init = mx.sym.SwapAxis(mx.sym.Variable('LSTM_init_h'), dim1=0, dim2=1)
rnn_c_init = mx.sym.SwapAxis(mx.sym.Variable('LSTM_init_c'), dim1=0, dim2=1)
rnn = mx.sym.RNN(data=data, num_layers=5, state_size=512, mode='lstm', state=rnn_h_init, state_cell=rnn_c_init)
print rnn.list_arguments()
# ['data', 'rnn0_parameters', 'LSTM_init_h', 'LSTM_init_c']
print rnn.infer_shape(data=(15, 10, 1024))[0]
# [(15L, 10L, 1024L), (11554816L,), (), ()]

The returned shape of rnn_h_init and rnn_c_init are both empty tuple "()" and produce an error when initializing the parameters in FeedForward model.
If I change the code to

rnn_h_init = mx.sym.Variable('LSTM_init_h')
rnn_c_init = mx.sym.Variable('LSTM_init_c')
# [(15L, 10L, 1024L), (11554816L,), (5L, 10L, 512L), (5L, 10L, 512L)]

the returned shapes seem to be right now. But according to the instructions in rnn_cell_demo.py, both c and h are time major rather than batch major.
I am not familiar with RNN symbol in mxnet, can anyone provide some help about this? Should I define h and c without using SwapAxis?

[yajiedesign]

This issue is closed due to lack of activity in the last 90 days. Feel free to reopen if this is still an active issue. Thanks!