mIou’s GPU efficient implementation algorithm

When loading data, convert semantic labels into binary matrices

Since the framework implements asynchronous loading of data, the following code in the data loading phase will not affect the GPU computing speed.

        # 图像标签的稠密矩阵
        shape = img_l.shape  # img_l为图像标签
        img_ld = numpy.zeros((19,) + shape)  # img_ld为稠密矩阵
        for i in range(19):
            img_ld[i] = (img_l == i)
        img_ld.astype('float32')
        

mIou calculation function

Applied in pytorch framework

class mIou(nn.Module):

    def __init__(self):
        super().__init__()

    def forward(self, outputs, targets):
        ma = torch.amax(outputs, dim=1)
        outputs = outputs - ma
        outputs = (outputs == 0).type_as(outputs)

        a = outputs + targets

        tt = torch.sum(a == 2, dim=[0, 2, 3])
        tf = torch.sum(a > 0, dim=[0, 2, 3])

        tt[tf == 0] = 1
        tf[tf == 0] = 1

        return torch.mean(tt / tf)

        

Join the mIou evaluation

Instantiate evaluation function

evaluate = util.loss.mIou().cuda()  # 评估函数
m = 0

Calculate miou during training

    e = evaluate(y_, z)

    m += e.item()
    if batch % 30 == 0:  # 每30个批次进行一次评估
        print('loss:', f / 30, 'mIou:', m / 30)
        m = 0

If you have any more efficient algorithms or shortcomings, please leave a comment.