[Question] Getting image paths along along with a batch?

[KnockerPulsar]

Hello, thanks for the awesome work.
I'm working on using ViTs to classify some video data. Frame by frame classification is fine and all, but the temporal aspect is also important. Instead of just slapping on an LSTM I wanted to experiment with something: classifying all frames in a video and then choosing the most frequently appearing classification as the final classification for the frame/video.

My main issue is getting the frame's path or really anything to identify it, I've tried forking this repo and editing the scripts on Colab but I hit a brick wall. I'm not really familiar with how the data loaders work but from what I understood I just needed to edit __getitem__ in the dataset class, I followed that and did that but it seemed to do nothing.
Following the code, I discovered that if no-prefetcher is False, the code uses PrefetchLoader which seems to work differently from PyTorch's built in loader. The brick wall I'm currently at is in __iter__ in the line:

for next_input, next_target in self.loader:

I tested a bit and it seems that __getitem__ doesn't seem to affect it. My main issue now is that for example if I make it return, let's say, the read image, the label, and the path and then update the for loop to receive an extra placeholder, It says that it expected 3 values but got 2. So clearly my modification didn't change anything. self.loader is a black box to me so I don't really know what happens in the for loop statement.

Thanks in advance for any help provided.

[rwightman]

@KnockerPulsar ordering for eval/test loaders is consistent (you don't need to include it in any iterable/getitem), you can do it completely out of band.

[Kieran31]

Hi, I know I can get the filenames in ImageDataset() class, and filenames' order is consistent with the order of images.
for example, eval_ds = ImageDataset('eval'), filenames = eval_ds.parser.filenames().
When I create a dataloader (PrefetchLoader), is there any method I can still get the filenames? Because I want to keep records of those mis-predicted data. I know I can do it without a dataloader, in other words, just iterate the eval_ds. But this can be inefficient.

[rwightman]

it's in the inference script, the datasets have a filenames() method that matches the iteration order during inference (assuming that no shuffling was enabled) https://github.com/rwightman/pytorch-image-models/blob/d7e1e7144a62f0368787c081eb999d27cb97d159/inference.py#L103-L123