KeyError: 'pixel_values'

[zeyunZhao]

Hello Authors,

This is a good work. I am trying to reproduce your work as your guidance, but it has this bug, could you tell me why? I am using UKB fMRI data to train the model with 1 batchsize in both training and inference.

Any reply will be appreciated!

[wenruifan]

It seems to be a bug in the code. I notice that authors use pixel_values from dataset (in BrainLM_Tutorial.ipynb).

with torch.no_grad():  # No gradient tracking for inference
    for batch in tqdm(train_ds, desc="Running inference"):
        pixel_values = batch["pixel_values"].unsqueeze(0)
        batch_size, channels, height, width = pixel_values.shape

However, when they build the dataset, I don't find any steps adding this feature.