How to use ncnn inference for a 4-dimensional speech model

[SEMLLYCAT]

Hello, my input is
input (1, 257, 1, 2)
state_h (1, 31, 32)
state_c (1, 32, 32)
May I ask how I should create the correct Mat data type for ncnn inference?

[magicse]

HiFi-GAN accept 1d array where input channels = num_mels and = 80 (for ncnn it present as height) and width is time stamps of mel -spectrogramm.

Input Mel spectrogram paramters:
n_fft = 1024
num_mels = 80
sampling_rate = 22050
hop_size = 256
win_size = 1024
fmin = 0
fmax = 8000

[SEMLLYCAT]

Thank you for your reply. I have a question, please check it. eg:
import torch as nn
class TestLstm(nn.Module):
def init(self, input_size, hidden_size, rnn_type='LSTM', droupt=0, bidirectional=False):
super(TestLstm, self).init()
self.rnn = nn.LSTM(input_size, hidden_size, 1, dropout=dropout, batch_first=True, bidirectional=False)

def forword(self, x, state_list):
output, hidden_state = self.rnn(x, state_list)
return output, hidden_state

if name == "main":
model = TestLstm(input_size=64, hidden_size=64, rnn_type='LSTM', dropout=0, bidirectional=False)
model.eval()
x = torch.randn(1, 64, 31, 1)
state_h0 = torch.randn(1, 31, 64, requires_grad=False)
state_c0 = torch.randn(1, 31, 64, requires_grad=False)
state_in = [(state_h0, state_c0)]
out, state_out = model(x, state_in[0])

onnx2ncnn:
.param
The lstm layer in the.param file is similar：
LSTM /lstm_t1/LSTM 3 3 x state_h0 state_c0 /lstm_t1/Transpose_1_out_0 state_out_h0 state_out_c0 0=32 1=8192 2=0

problem：The ncnn inference result is inconsistent with the onnx result, and I cannot align all the results. I don't know whether I used it correctly, and could you please help verify it, thanks