Gerard Walsh

This could be due to the fact the optical flow operations require two frames - thus the last frame in the sequence is used only for optical flow calculations and...

Same issue when training on a custom dataset with one class.

I was trying to use the pascal dataloader for training on a custom dataset with 1 class - have switched to coco dataset loader and have left `self.max_objs = 128`,...

Line in question [here](https://github.com/karpathy/llm.c/blob/cc5931bfdbd89c85ae8a0420888c327e26d5c688/train_gpt2.py#L25). Any idea on speed up/reduction in number of operations?

As far as I understand, train.lst is simply a text file where each line is the path for a particular training image.

Which implementations did you use to generate the noisy labels? It seems quite hard to find implementations of the 4 methods the authors chose to generate groundtruth.

Thanks, that would be great to have your implementations. @JiangAndy I used OpenCV's saliency detection [module](https://docs.opencv.org/master/d8/d65/group__saliency.html).