RuntimeError: The size of tensor a (3) must match the size of tensor b (256) at non-singleton dimens_runtimeerror: the size of tensor a (262144) must m-CSDN博客

RuntimeError: The size of tensor a (3) must match the size of tensor b (256) at non-singleton dimension 2

1. 问题位置

# visuals_test是有序字典 test_fake_TList是张量列表，有三个张量
test_fake_TList = visuals_test['fake_Ts'] 
save_image(test_fake_TList_cat, os.path.join(output_dir1, f'test_imgs_epoch{i}+{total_test_step}.png'), nrow=4)

2. 报错：

Traceback (most recent call last):
  File "d:/SSIR-Gzm/ssir/inference1.py", line 280, in <module>
    save_image(test_fake_TList,output_dir1,"1.png")
  File "D:\Anaconda\envs\pytorch\lib\site-packages\torch\autograd\grad_mode.py", line 28, in decorate_context
    return func(*args, **kwargs)
  File "D:\Anaconda\envs\pytorch\lib\site-packages\torchvision\utils.py", line 132, in save_image
    grid = make_grid(tensor, **kwargs)
  File "D:\Anaconda\envs\pytorch\lib\site-packages\torch\autograd\grad_mode.py", line 28, in decorate_context
    return func(*args, **kwargs)
  File "D:\Anaconda\envs\pytorch\lib\site-packages\torchvision\utils.py", line 108, in make_grid
    ).copy_(tensor[k])
RuntimeError: The size of tensor a (3) must match the size of tensor b (256) at non-singleton dimension 2

3. 打印出来test_fake_TList

torch.Size([4, 3, 256, 256]), torch.Size([4, 3, 256, 256]), torch.Size([4, 3, 256, 256])]
torch.float32, torch.float32, torch.float32]
[device(type='cuda', index=0), device(type='cuda', index=0), device(type='cuda', index=0)]
[(tensor(0., device='cuda:0'), tensor(1., device='cuda:0')), (tensor(0., device='cuda:0'), tensor(1., device='cuda:0')), (tensor(0., device='cuda:0'), tensor(1., device='cuda:0'))]

4. 疑惑：三个张量的shape都一样，为什么还会报张量不兼容呢？

报错原因是调用save_image(test_fake_TList,output_dir1,“1.png”)时，test_fake_TList是一个列表，而不是一个单一的Tensor。
torchvision.utils.save_image和make_grid期望输入是一个形状类似于(N, C, H, W)的Tensor，而不是[Tensor, Tensor, …]的列表。

5. 解决方法

你需要将你的test_fake_TList（形如[tensor1, tensor2, tensor3]，每个tensor shape为(4, 3, 256, 256)）拼接成一个大的tensor。

import torch
from torchvision.utils import save_image

# 假设 test_fake_TList = [tensor1, tensor2, tensor3]

# 1. 拼接为一个 (N, C, H, W) 的tensor，N=4*3=12
test_fake_T = torch.cat(test_fake_TList, dim=0)  # (12, 3, 256, 256)

# 2. 保存图片
save_image(test_fake_T, os.path.join(output_dir1, "1.png"))