Python 使用TFLite对多个LSTM进行流式LSTM推理

我正在尝试将LSTM模型移植到TFLite，由于中间池，我的LSTM层以不同的速度运行：

Conv1d(kernel_size=5)
Pool(2)
LSTM()
Conv1d(kernel_size=5)
Pool(2)
LSTM()

此LSTM模型用于音频流推断，其中第一层获得大小为N的块。每个LSTM需要保持其自身的隐藏状态，并且只有在其缓冲区已满时才执行前向传递并将其结果转发到下一层（对于池，内核大小为卷积+1）。例如：

T=6: 6 frames of audio
LSTM1 forward -> lstm1_t0
LSTM1 hidden_state -> new_hidden_state
LSTM2 buffer: [lstm1_t0]
T=8:
LSTM1 forward -> lstm1_t1
LSTM1 hidden_state -> new_hidden_state
LSTM2 buffer: [lstm1_t0, lstm1_t1]
...
T=16
LSTM1 forward -> lstm1_t5
LSTM1 hidden_state -> new_hidden_state
LSTM2 buffer: [lstm1_t0, ..., lstm1_t5] <- length 6
LSTM2 forward -> lstm2_t0
LSTM2 hidden_state -> new_hidden_state
LSTM2 buffer: [lstm1_t1, ..., lstm1_t5] <- length 5

T=6:6帧音频
LSTM1正向->LSTM1\u t0
LSTM1隐藏状态->新隐藏状态
LSTM2缓冲区：[lstm1_t0]
T=8：
LSTM1正向->LSTM1\U t1
LSTM1隐藏状态->新隐藏状态
LSTM2缓冲区：[lstm1_t0，lstm1_t1]
...
T=16
LSTM1正向->LSTM1_t5
LSTM1隐藏状态->新隐藏状态
LSTM2缓冲区：[lstm1_t0，…，lstm1_t5]LSTM2_t0
LSTM2隐藏状态->新隐藏状态
LSTM2缓冲区：[lstm1_t1，…，lstm1_t5]