建议这里把errors的维度打印出来看下，topk op里面k是对最后一维求topk。

您好，不好意思，我上面写错了，应该是长度是103041,维度应该是一维,errors.shape打印出来是(103041L,)这样的

创建网络的时候，第一维对应batch size，fluid里面会把batch size初始值置为-1
比如我这个代码：
a = fluid.layers.data(name='a', shape=[3,100,100], dtype='float32')
print a.shape
输出是：
(-1L, 3L, 100L, 100L)
你这里的errors.shape (103041, )是这么打印的吗？按我上面的这个示例的话，errors.shape[0]应该是-1

您好，我把batch_size设为固定了，然后把所有输入都压为一维之后的值是103041,因为topk的k好像不支持输入为-1，我的打印方式是errors.shape，我把batch_size固定取消之后，压平后的维度是-1，K设为1,报了Expected input_dims[input_dims.size() - 1] >= k这样的错

压平使用的什么操作，reshape还是flatten？

我使用的是reshape的操作

能贴下代码吗，我用你的代码运行下看问题在哪

def lovasz_softmax(logit,labels):
    """
    多分类的lovasz_softmax
    :param logit:[batch,class_nums,width,height] 1,9,321,321
    :param labels:[batch,width,height] 1,321,321
    :return:loss
    """
    labels = fluid.layers.elementwise_min(
        labels,
        fluid.layers.assign(np.array(
            [num_classes - 1], dtype=np.int32)))
    logit = fluid.layers.transpose(logit, [0, 2, 3, 1])
    logit = fluid.layers.reshape(logit, [-1, num_classes])
    logit = fluid.layers.softmax(logit)
    labels = fluid.layers.reshape(labels, [-1])
    labels = fluid.layers.cast(labels, 'int64')
    C = 9
    losses = []
    for c in range(C):
        # 得到一个类别label的list
        fg = (labels == c).astype('float32')
        # 得到一个类别所有预测结果的分数
        crop_logit = fluid.layers.crop(logit,offsets=[0,c],shape=[103041,1])
        crop_logit.stop_gradient =True
        crop_logit = fluid.layers.reshape(crop_logit,[-1])
        # label的list减去infer的分数
        errors = fluid.layers.abs(fg - crop_logit)
        # print errors.shape
        # # 按从大到小排序
        errors_sorted, perm = fluid.layers.topk(errors,k=errors.shape[0])
        # 把index注入到fg中
        fg_sorted = fluid.layers.gather(fg,perm)
        #  计算梯度
        gts = fluid.layers.reduce_sum(fg_sorted)
        cumsum = fluid.layers.cumsum(fg_sorted.astype("float32"))
        gts = fluid.layers.zeros(cumsum.shape, dtype='float32') + gts
        intersection = gts - cumsum
        union = gts + fluid.layers.cumsum((1 - fg_sorted).astype("float32"))
        jaccard = 1. - intersection / union
        jaccard_0 = fluid.layers.crop(jaccard, offsets=[0], shape=[1])
        jaccard_1 = fluid.layers.crop(jaccard, offsets=[1], shape=[103041])
        jaccard_2 = fluid.layers.crop(jaccard, offsets=[0], shape=[103040])
        jaccard = fluid.layers.concat([jaccard_0, jaccard_1 - jaccard_2], 0)
        losses.append(fluid.layers.elementwise_mul(errors_sorted,jaccard))
    losses = fluid.layers.stack(losses)
    return losses

您好，我想要实现的是分割中使用的lovasz_softmax(多分类)，参考是https://github.com/bermanmaxim/LovaszSoftmax/blob/master/tensorflow/lovasz_losses_tf.py
函数在fluid.program_guard中使用的，刚刚接触这个框架，麻烦帮忙调试下,谢谢

其他部分的代码也贴下吧，数据可以改成用随机数

# -- coding:utf-8 --

import paddle.fluid as fluid

def lovasz_softmax(logit,labels):
    """
    多分类的lovasz_softmax
    :param logit:[batch,class_nums,width,height]
    :param labels:[batch,width,height]
    :return:loss
    """
    logit = fluid.layers.transpose(logit, [0, 2, 3, 1])
    logit = fluid.layers.reshape(logit, [-1, 9])
    logit = fluid.layers.softmax(logit)
    labels = fluid.layers.reshape(labels, [-1])
    labels = fluid.layers.cast(labels, 'int64')
    C = 9
    losses = []
    for c in range(C):
        # 得到一个类别label的list
        fg = (labels == c).astype('float32')
        # 得到一个类别所有预测结果的分数
        crop_logit = fluid.layers.crop(logit,offsets=[0,c],shape=[103041,1])
        crop_logit.stop_gradient =True
        crop_logit = fluid.layers.reshape(crop_logit,[-1])
        # label的list减去infer的分数
        errors = fluid.layers.abs(fg - crop_logit)
        # # 按从大到小排序
        errors_sorted, perm = fluid.layers.topk(errors,k=errors.shape[0])
        # 把顺序注入到fg中
        fg_sorted = fluid.layers.gather(fg,perm)
        # # 计算梯度
        gts = fluid.layers.reduce_sum(fg_sorted)
        cumsum = fluid.layers.cumsum(fg_sorted.astype("float32"))
        gts = fluid.layers.zeros(cumsum.shape, dtype='float32') + gts
        intersection = gts - cumsum
        union = gts + fluid.layers.cumsum((1 - fg_sorted).astype("float32"))
        jaccard = 1. - intersection / union
        jaccard_0 = fluid.layers.crop(jaccard, offsets=[0], shape=[1])
        jaccard_1 = fluid.layers.crop(jaccard, offsets=[1], shape=[103041])
        jaccard_2 = fluid.layers.crop(jaccard, offsets=[0], shape=[103040])
        jaccard = fluid.layers.concat([jaccard_0, jaccard_1 - jaccard_2], 0)
        losses.append(fluid.layers.elementwise_mul(errors_sorted,jaccard))
    losses = fluid.layers.stack(losses)
    return losses

sp = fluid.Program()
tp = fluid.Program()

with fluid.program_guard(tp, sp):
    logit = fluid.layers.uniform_random(shape=[1,9,321,321],min=0.0,max=1.0)
    logit = fluid.layers.cast(logit,dtype='float32')
    label = fluid.layers.uniform_random(shape=[1,321,321],min=0.0,max=1.0)
    lovasz_loss = lovasz_softmax(logit,label)

place = fluid.CPUPlace()
exe = fluid.Executor(place)
exe.run(sp)

exe.run(tp)

全部代码就是这样的，麻烦您看看是什么问题

我试了下，在Paddle fluid1.2 版本上确实会报错，原因在于[/paddle/paddle/fluid/framework/ddim.cc:235]
PADDLE_ENFORCE(begin < end, "Begin index must be less than end index in ddim slice.");
对ddim做slice的时候要求end > begin，而这里begin和end都等于0。

在新版本fluid 1.3里面已经修复了这个问题，建议使用1.3版。

OK，好的

您好，现在官网好像还没有1.3的版本，1.3的版本要怎么安装呢？