PyTorch：了解RuntimeError：梯度运算所需的变量之一已通过就地操作进行了修改

问题描述

这是一个问题设置：

将rgb转换为hsv的函数：

def rbg2hsv(img_rgb):
    batch_size,channel,height,width = img_rgb.size()

    r,g,b = img_rgb[:,:,:],img_rgb[:,1,2,:]

    t = torch.min(img_rgb,dim=1,keepdim=False)[0]
    v = torch.max(img_rgb,keepdim=False)[0]

    s = (v - t) / (v + 1e-6)
    s[v == 0] = 0

    # v==r
    hr = 60 * (g - b) / (v - t + 1e-6)
    # v==g
    hg = 120 + 60 * (b - r) / (v - t + 1e-6)
    # v==b
    hb = 240 + 60 * (r - g) / (v - t + 1e-6)

    h = torch.zeros(batch_size,width,requires_grad=False)
    if torch.cuda.is_available():
        h = h.cuda()
    h = h.flatten()

    hr = hr.flatten()
    hg = hg.flatten()
    hb = hb.flatten()

    h[(v == b).flatten()] = hb[(v == b).flatten()]
    h[(v == g).flatten()] = hg[(v == g).flatten()]
    h[(v == r).flatten()] = hr[(v == r).flatten()]

    h[h < 0] += 360

    h = torch.reshape(h,(batch_size,width))

    img_hsv = torch.stack([h,s,v])
    img_hsv = img_hsv.permute(1,3)

    return img_hsv

这是重现问题的最小代码：

img_rgb = torch.rand(2,3,32,32)
img_rgb.requires_grad = True
instance_norm = nn.InstanceNorm2d(1,affine=False)

gamma = 2
beta = 1

img_hsv = rbg2hsv(img_rgb)

# 1: OK
#img_hsv[:,2:3,:] = img_hsv[:,:] * gamma + beta

# 2: FAIL
img_hsv[:,:] = instance_norm(img_hsv[:,:])

# 3: OK
#img_hsv[:,:].clone())

img_hsv.mean().backward() # Fails on this line

print('img_rgb.grad.size()',img_rgb.grad.size())

变体2: FAIL失败，并显示错误：RuntimeError: one of the variables needed for gradient computation has been modified by an inplace operation: [torch.FloatTensor [1,32]],which is output 0 of ViewBackward,is at version 1; expected version 0 instead. Hint: enable anomaly detection to find the operation that failed to compute its gradient,with torch.autograd.set_detect_anomaly(True).

如果我将rbg2hsv更改为虚拟add_fn，则所有三种变体都可以工作：

def add_fn(x):
    x = x + 1
    return x

img_rgb = torch.rand(2,affine=False)

gamma = 2
beta = 1

img_hsv = add_fn(img_rgb)

# 1: OK
#img_hsv[:,:] * gamma + beta

# 2: OK
img_hsv[:,:].clone())

img_hsv.mean().backward()

print('img_rgb.grad.size()',img_rgb.grad.size())

所以我的问题是：

1. 为什么变量1没有失败（似乎它也是张量的原位修改）？
2. 为什么将rbg2hsv更改为add_fn才能使所有变体有效？ （即rbg2hsv有问题吗？）

这里有一个类似的问题https://discuss.pytorch.org/t/encounter-the-runtimeerror-one-of-the-variables-needed-for-gradient-computation-has-been-modified-by-an-inplace-operation/836，但是当我们需要使用.clone()以及何时不使用时（即某些操作是否已就位），我仍然不理解规则？

解决方法

暂无找到可以解决该程序问题的有效方法，小编努力寻找整理中！

如果你已经找到好的解决方法，欢迎将解决方案带上本链接一起发送给小编。

小编邮箱:dio#foxmail.com (将#修改为@）