Runtimeerror: output with shape [1, 256, 256] doesn't match the broadcast shape [3, 256, 256] [error]

resolvent

stay dataloader Before dataset In the function , Add .convert(‘RGB’),eg：

image = Image.open(image_file).convert('RGB')

To solve the process

In the process of paper reproduction, try to apply the unused data set to the source code , If there is a report error . After checking, it was found that the error occurred in dataloader in , Consider the problem of inputting pictures , Judge that the newly added data set is a grayscale image through the following code （1. The grayscale image has only one channel 2. Directly view the pixel value ）, So infer dataloader Only three channels are accepted RGB chart , Convert picture to RGB Just figure .

Print picture pixel values

img_path = ' Write the address of the image you want to view here '
from PIL import Image
from torchvision import transforms
img = Image.open(img_path)
change = transforms.ToTensor()
img = change(img)
print(img)

Check the pixel value of the picture

import cv2
img_path = ' Write the address of the image you want to view here '
img_c = cv2.imread(img_path,cv2.IMREAD_UNCHANGED)
cv2.imshow('photo',img_c)
key = cv2.waitKey(0)

The solution of blogging is transforms.Compose（） Add the following program segments in , My program hasn't improved , If it's a grayscale ,RGB Transformation and repeat The converted three channel images are consistent , This may be another reason for the program .

transforms.Lambda(lambda x: x.repeat(3, 1, 1))