那些年踩过的深度学习坑一 维度问题

Posted 2022-03-02 cv君

报错： ValueError: axes don't match array

ValueError:轴与数组不匹配 问题，下面是报错：

报错行代码经常为：image = np.transpose(image, (2, 0, 1))

这行代码的作用是调换整图像的shape顺序

[width, height, channels]

（2， 0， 1）即为结果的顺序

[channels, height, width]

  0%|          | 0/1 [00:00<?, ?it/s]Epoch: 0
  0%|          | 0/1 [00:00<?, ?it/s]
Traceback (most recent call last):
  File "D:/cver/Others/Auto_Classification/Auto_Classification/train.py", line 174, in <module>
    run(images_list, val_list)
  File "D:/cver/Others/Auto_Classification/Auto_Classification/train.py", line 147, in run
    train_one(train_dataloader, model, optimizer, loss_fn, None)
  File "D:/cver/Others/Auto_Classification/Auto_Classification/train.py", line 76, in train_one
    for bi, data in enumerate(tk):
  File "D:\\Users\\11145999\\AppData\\Roaming\\Python\\Python37\\site-packages\\tqdm\\std.py", line 1180, in __iter__
    for obj in iterable:
  File "D:\\Users\\11145999\\AppData\\Roaming\\Python\\Python37\\site-packages\\torch\\utils\\data\\dataloader.py", line 521, in __next__
    data = self._next_data()
  File "D:\\Users\\11145999\\AppData\\Roaming\\Python\\Python37\\site-packages\\torch\\utils\\data\\dataloader.py", line 561, in _next_data
    data = self._dataset_fetcher.fetch(index)  # may raise StopIteration
  File "D:\\Users\\11145999\\AppData\\Roaming\\Python\\Python37\\site-packages\\torch\\utils\\data\\_utils\\fetch.py", line 49, in fetch
    data = [self.dataset[idx] for idx in possibly_batched_index]
  File "D:\\Users\\11145999\\AppData\\Roaming\\Python\\Python37\\site-packages\\torch\\utils\\data\\_utils\\fetch.py", line 49, in <listcomp>
    data = [self.dataset[idx] for idx in possibly_batched_index]
  File "D:/cver/Others/Auto_Classification/Auto_Classification/train.py", line 46, in __getitem__
    "image": self.normalize(image),
  File "D:/cver/Others/Auto_Classification/Auto_Classification/train.py", line 26, in normalize
    image = np.transpose(image, (2, 0, 1)) # [3,224,224]
  File "<__array_function__ internals>", line 6, in transpose
  File "D:\\Users\\11145999\\AppData\\Roaming\\Python\\Python37\\site-packages\\numpy\\core\\fromnumeric.py", line 660, in transpose
    return _wrapfunc(a, 'transpose', axes)
  File "D:\\Users\\11145999\\AppData\\Roaming\\Python\\Python37\\site-packages\\numpy\\core\\fromnumeric.py", line 57, in _wrapfunc
    return bound(*args, **kwds)
ValueError: axes don't match array

但此时发生了一些问题，如果图像混杂进了灰度图等图片，那么这行代码就会报错（如上）

解决方法

from skimage.color import gray2rgb
image = Image.open(image_name)
if image.mode == 'L':      # 判断是否为灰度图
    image = gray2rgb(image)  # 如果是，就转称彩色 （通道数变成三）

仔细看可以发发现，opencv 的坑，所以大家在一步步排查问题的时候，比如 当一个训练任务你原本能跑，换了数据后不能跑了，有可能就是混入了灰度图，你第一次看到ValueError: axes don't match array ，百度会说是维度问题，然后大家根据正常debug，用opencv读的图片，发现所有图片都是三通道，没问题，不过实际是opencv 的一个玩笑，大家看我上面的图片就知道了。本期opencv 坑 就到这，下期再见。