LSTM training speed

1

I use a lstm code like this

class StackedLSTM(nn.Module):
def __init__(self, input_size, hidden_size, num_layers=1, dropout=0):
    super(StackedLSTM, self).__init__()
    self.input_size = input_size
    self.hidden_size = hidden_size
    self.num_layers = num_layers
    self.dropout = nn.Dropout(dropout)
    self.layers = nn.ModuleList()

    for i in range(num_layers):
        self.layers.append(nn.LSTMCell(input_size, hidden_size))
        input_size = hidden_size

def forward(self, input, hidden):
    h_0, c_0 = hidden
    h_1, c_1 = [], []
    for i, layer in enumerate(self.layers):
        h_1_i, c_1_i = layer(input, (h_0[i], c_0[i]))
        input = h_1_i
            h_1_i, c_1_i = layer(input, (h_0[i], c_0[i]))
        input = h_1_i
        if i != self.num_layers:
            input = self.dropout(input)
        h_1 += [h_1_i]
        c_1 += [c_1_i]

    h_1 = torch.stack(h_1)
    c_1 = torch.stack(c_1)

    return input, (h_1, c_1)

when i compare it to the nn.LSTM (official API) on language model, the perplexity is close but the training speed differ a lot, my code may need 170s for one epoch while use nn.LSTM will only take 50s.

Can someone tell me the reason?

2

nn.LSTM uses the optimized CUDNN LSTM kernel by default, which avoids all the overhead of building a graph that contains hundreds of separate mathematical operations and the overhead of launching those operations as separate kernels. You’ll get the same slow performance by setting torch.backends.cudnn.enabled=False.

1 Like