pytorch之tensor操作报错汇总

1. torch.split()

t = torch.ones((2,5))

list_of_tensors = torch.split(t,[2,1,1],dim=1)

print("t:{}".format(t))
for idx,t in enumerate(list_of_tensors):
    print("di {} ge tensor:{},shape is {}".format(idx+1,t,t.shape))

RuntimeError: split_with_sizes expects split_sizes to sum exactly to 5 (input tensor's size at dimension 1), but got split_sizes=[2, 1, 1]

是说list_of_tensors = torch.split(t,[2,1,1],dim=1)中，输入tensor t 的第一维的维度是5，但是split_size=[2, 1, 1]的和是4不是5，改为和为5才不会报错。

2. torch.index_select()

t_select = torch.index_select(t,dim=1,index=idx)

t = torch.randint(0,9,(3,3))
idx = torch.tensor([0,2],dtype=torch.long)
t_select = torch.index_select(t,dim=1,index=idx)
print("t:\n{}\nt_select:\n{}".format(t,t_select))

输出:

t:
tensor([[4, 6, 6],
        [1, 5, 6],
        [0, 0, 8]])
t_select:
tensor([[4, 6],
        [1, 6],
        [0, 8]])

t_select = torch.index_select(t,dim=0,index=idx)

t = torch.randint(0,9,(3,3))
idx = torch.tensor([0,2],dtype=torch.long)
t_select = torch.index_select(t,dim=0,index=idx)
print("t:\n{}\nt_select:\n{}".format(t,t_select))

输出:

t:
tensor([[0, 6, 0],
        [4, 1, 0],
        [2, 2, 6]])
t_select:
tensor([[0, 6, 0],
        [2, 2, 6]])

idx = torch.tensor([0,2],dtype=torch.float)

t = torch.randint(0,9,(3,3))
idx = torch.tensor([0,2],dtype=torch.float)
t_select = torch.index_select(t,dim=0,index=idx)
print("t:\n{}\nt_select:\n{}".format(t,t_select))

报错:

RuntimeError: Expected object of scalar type Long but got scalar type Float for argument #3 'index'

t_select = torch.index_select(t,dim=0,index=idx)中的index应该是long类型，设置为float类型会报错，idx = torch.tensor([0,2],dtype=torch.long)改为long类型就不会报错了。

3. backward()

import torch
torch.manual_seed(10)

flag = True
# flag = False
if flag:
    w = torch.tensor([1.],requires_grad=True)
    x = torch.tensor([2.],requires_grad=True)

    a = torch.add(w,x)
    a.retain_grad()
    b = torch.add(w,1)
    y = torch.mul(a,b)

    y.backward()
    # print(w.grad)
    y.backward()

RuntimeError: Trying to backward through the graph a second time, but the buffers have already been freed. Specify retain_graph=True when calling backward the first time.

进行了两次反向传播，retain_graph默认为False，在第一次反向传播之后计算图会被释放，如果想继续使用计算图，要把retain_graph设置为True。

    y.backward(retain_graph=True)
    # print(w.grad)
    y.backward()