Question

我在用LSTM实现Pytorch时遇到数据类型问题。基于类似的问题，您尝试将Input，h和c的格式更改为ShortTensor，但是仍然出现相同的错误：

RuntimeError：标量类型的预期对象很短，但是得到了标量类型调用_th_mm
时，参数＃2'mat2'的浮动

class data(Dataset):
    def __init__(self, samples=10000, number=30):
        self.x = torch.from_numpy(np.matrix
                        (np.random.random_integers(0,9,samples*number).reshape(samples, number)))
        self.y = torch.from_numpy(np.zeros((samples))).type(torch.ShortTensor)
        for index, row in enumerate(self.x):
            self.y[index] = 1 if torch.sum(row) >= 130 else 0

class LSTM(nn.Module):
    def __init__(self,i_size, h_size, n_layer, batch_size = 30 ):
        super().__init__()
        self.lstm = nn.LSTM(input_size=i_size, hidden_size=h_size, num_layers=n_layer)
        self.h = torch.randn(n_layer, batch_size, h_size).type(torch.ShortTensor)
        self.c = torch.randn(n_layer, batch_size, h_size).type(torch.ShortTensor)
        self.hidden = (self.h, self.c)
        self.linear = nn.Linear(n_layer, 1)

    def forward(self,x):
        out, hidden = self.lstm(x.type(torch.ShortTensor), self.hidden)
        out = nn.Softmax(self.linear(out.short()))
        return out

data_set = data()
train_data = data_set.x[0:8000, :, None]
train_label = data_set.y[0:8000]
test_data = data_set.x[8000:, : , None]
test_label = data_set.y[8000:]

input_size = 1
hidden_size = 30
layer_num = 200
model_LSTM = LSTM(input_size, hidden_size, layer_num)
#model_LSTM.cuda()
y_ = model_LSTM(train_data)

Answer 1

我在int16中生成输入数据，显然nn.LSTM仅占用float32，并且错误尚不清楚。

标量类型的预期对象简短，但在调用_th_mm时，参数＃2'mat2'的标量类型为Float

1 个答案: