RuntimeError: Unsupported marker type 0x4f on 2nd epoch

vision
1

Greetings,
I get this error when I train a resnet18 model from Torchvision on a custom dataset. It only happens on the 2nd epoch after a few dozen iterations.

---------------------------------------------------------------------------
RuntimeError                              Traceback (most recent call last)
<ipython-input-63-e338034a4492> in <module>()
----> 1 model_ft = train_model(model_ft, criterion, optimizer_ft, exp_lr_scheduler,num_epochs=25)

8 frames
/usr/local/lib/python3.7/dist-packages/torchvision/io/image.py in decode_image(input, mode)
    200         output (Tensor[image_channels, image_height, image_width])
    201     """
--> 202     output = torch.ops.image.decode_image(input, mode.value)
    203     return output
    204 

RuntimeError: Unsupported marker type 0x4f

I use these transforms:

transform = nn.Sequential(
            T.Resize([256, ]), 
            T.CenterCrop(224),
            T.ConvertImageDtype(torch.float),
            T.Normalize([0.485, 0.456, 0.406], [0.229, 0.224, 0.225])
        )

Does anyone have an idea of what could be wrong?

2

The image decoding is failing and torchvision seems to reraise the error. Are you able to load the file with another library and if so, could you save it using a different file name, and try to load it again via torch.ops.image.decode_image?

3

I read the image and saved with PIL and it solved the problem. Thanks!

p = 'wikiart/Post_Impressionism/vincent-van-gogh_l-arlesienne-portrait-of-madame-ginoux-1890.jpg'
from PIL import Image, ImageFile
ImageFile.LOAD_TRUNCATED_IMAGES = True
with Image.open(p) as im:
    im.save(p)

Here is the image for context:

Unknown
Unknown900×1172 79.5 KB