关于F1 loss

对于多个类的F1 loss，是先计算出每个类的F1，然后取平均

不能求每个batch的F1，然后平均

比如，下面的写法是错误的

def f1_loss(logits, labels):
    __small_value=1e-6
    beta = 1
    batch_size = logits.size()[0]
    p = F.sigmoid(logits)
    l = labels
    num_pos = torch.sum(p, 1) + __small_value
    num_pos_hat = torch.sum(l, 1) + __small_value
    tp = torch.sum(l * p, 1)
    precise = tp / num_pos
    recall = tp / num_pos_hat
    fs = (1 + beta * beta) * precise * recall / (beta * beta * precise + recall + __small_value)
    loss = fs.sum() / batch_size
    return (1 - loss)

 

正确的解法是：

def f1_loss(predict, target):
    predict = torch.sigmoid(predict)
    predict = torch.clamp(predict * (1-target), min=0.01) + predict * target
    tp = predict * target
    tp = tp.sum(dim=0)
    precision = tp / (predict.sum(dim=0) + 1e-8)
    recall = tp / (target.sum(dim=0) + 1e-8)
    f1 = 2 * (precision * recall / (precision + recall + 1e-8))
    return 1 - f1.mean()