计算精度(在lstm模型中)

Question

我正在尝试获得我的准确性，我有这样的代码：

            num_correct = 0.0
            for inputs, labels in dataloader(
            valid_features, valid_labels, batch_size=batch_size, sequence_length=20):

                top_val, top_class = torch.exp(output).topk(1)
                num_correct += torch.sum(top_class.squeeze() == labels)
#...
            print(#...,
                  "Accuracy: {:.3f}".format(num_correct*1.0 / len(valid_labels) *1.0)

它总是打印0.000，所以我决定将原始值打印到num_correct：print(top_class.squeeze(), labels)中

tensor([ 1,  3,  3,  ...,  3,  4,  3], device='cuda:0') tensor([ 1,  1,  3,  ...,  3,  3,  3], device='cuda:0')
tensor([ 4,  3,  1,  ...,  4,  4,  3], device='cuda:0') tensor([ 4,  3,  1,  ...,  4,  4,  3], device='cuda:0')
tensor([ 2,  4,  2,  ...,  4,  4,  4], device='cuda:0') tensor([ 3,  4,  1,  ...,  4,  4,  4], device='cuda:0')
tensor([ 0,  1,  3,  ...,  2,  3,  0], device='cuda:0') tensor([ 0,  1,  3,  ...,  2,  2,  0], device='cuda:0')

这些数据看起来非常准确。所以..。我可以将其提取到numpy并完成，但有一种pytorch方法。

Answer 1

这是可行的：

num_correct += torch.sum(torch.eq(top_class.squeeze(), labels)).item()