对构造矩阵中的单个变量进行自动分级失败

问题描述

我想创建一个4x4矩阵来变形我的4 x N张量图像并将其优化为损失函数。损失函数可以工作，但是即使我使用纯矢量演算来确保它应该是可微的，但我初始化变形的方式还是有些错误。

def affine_transform(
        points1,points2,stepSize=0.01,iterations=2000):

    def zero_grads(elements):
        for elem in elements:
            elem.grad.data.zero_()

    # Constructing the affine transformation matrix

    ## translation in x and y which should be optimised
    tx = tc.zeros((1,1),requires_grad=True).float()
    ty = tc.zeros((1,requires_grad=True).float()

    # Construct vectors from the single variables size 4x1
    v1 = tc.from_numpy(np.array([[0,1]]).T).float()
    vx = v1.mm(tx)
    v2 = tc.from_numpy(np.array([[0,1,0]]).T).float()
    vy = v2.mm(ty)

    ## use outer product to increase dimentionality to 4x4
    u = tc.from_numpy(np.array([[1,0]]).T).float()
    T = u.mm(vx.T) + u.mm(vy.T).float()

    I = tc.from_numpy(np.eye(4)).float()
    R = I + T

    # Add ones to data
    n = points1.shape[1]
    one = np.ones(n)
    p1 = tc.from_numpy(np.vstack((points1,one))).float()
    p2 = tc.from_numpy(np.vstack((points2,one))).float()

    # Setup optimizer to tx and ty only
    optimizer = optim.Adam([tx,ty],stepSize)

    count = 0
    first = True
    while True:
        # Constructing the affine transformation matrix
        R = I + T  # + P

        # Calculate transition image (works as expected)
        transImg = bilinear_interpolate_torch(
                p1,tc.mm(R,p1))

        # Calculate loss
        loss = tc_ssd(p2,transImg)

        if first:
            loss.backward(retain_graph=True)
        else:
            loss.backward()
            
        # take a step
        optimizer.step()

        # Zero the gradients
        zero_grads([tx,ty])

        if count >= iterations and iterations != -1:
            print("Max iterations reached!")
            break

        count += 1

    R = I + T
    H3 = bilinear_interpolate_torch(p1.float(),p1.float()))

    return (H3,R)

它引发错误

[W python_anomaly_mode.cpp:60] Warning: Error detected in MmBackward. Traceback of forward call that caused the error:
  File "<stdin>",line 1,in <module>
  File "/home/fuzie/DropBox/KU/7.aar/MIA/Handins/hw4/code/affine_im.py",line 314,in <module>
    b1mesh.T,b2mesh.T)
  File "/home/fuzie/DropBox/KU/7.aar/MIA/Handins/hw4/code/affine_im.py",line 225,in affine_transform
    vy = v2.mm(ty)
 (function print_stack)
Traceback (most recent call last):
  File "<stdin>",line 257,in affine_transform
    # Stepping
  File "/home/fuzie/.pyenv/versions/mia/lib/python3.8/site-packages/torch/tensor.py",line 185,in backward
    torch.autograd.backward(self,gradient,retain_graph,create_graph)
  File "/home/fuzie/.pyenv/versions/mia/lib/python3.8/site-packages/torch/autograd/__init__.py",line 125,in backward
    Variable._execution_engine.run_backward(
RuntimeError: one of the variables needed for gradient computation has been modified by an inplace operation: [torch.FloatTensor [1,1]] is at version 1; expected version 0 instead. Hint: the backtrace further above shows the operation that Failed to compute its gradient. The variable in question was changed in there or anywhere later. Good luck!

但是我不明白为什么不允许这样构造矩阵T。

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@）

autograd python torch torch