阅读pyTorch的LSTM实现源码

RNNBasee

LSTM class is inherit RNNBase , which located in torch.nn.modules.rnn.RNNBase

class RNNBase(Module):
    __constants__ = ['mode', 'input_size', 'hidden_size', 'num_layers', 'bias',
                     'batch_first', 'dropout', 'bidirectional']

    mode: str
    input_size: int
    hidden_size: int
    num_layers: int
    bias: bool
    batch_first: bool
    dropout: float
    bidirectional: bool

__constants__ is torch scripts, see stack overflow

construction function

First of all, let’s see construction function , in other words, __init__

you can see, torch consider bi-direction stacked LSTM as base RNN implemention(RNNBase class). This architecture is really generic !

if mode == 'LSTM':
    gate_size = 4 * hidden_size
elif mode == 'GRU':
    gate_size =<