Closed alexandrumeterez closed 1 year ago
Hi, thanks for the question. Theoretically, it's possible to do batch inference to speed up automatic evaluation. However, it's not easy to modify the evaluation code to achieve this goal. The main issue is that different images may require different numbers of clicks, so it will be very tricky to handle this in evaluation.
Hi @qinliuliuqin. Thanks for replying so fast.
I was thinking along the lines of same number of clicks per image. Basically the batch would have the shape (BS, C, H, W) for the images and (BS, N_clicks) for the clicks.
@alexandrumeterez Thanks for the feedback. Yes, that's doable. Our code is based on RITM, which assumes BS=1 for inference. You may need to modify the code accordingly.
Hello! First of all, great work! I have just one question, since I can't quite get it from the code. Is it possible at inference time, to give batches of images and clicks? i.e. the input to the model is for example a batch of 16 images, each with its own set of clicks (of same length)?
Thanks!