Dataloader

[Hickey8]

Hi! I didn't quite understand the function of self._pos_start in RandomSampler and Sequenatial Sampler in dataloader.py. From my perspective, it can work in the same way without it, which means this two sampler needn‘t be overwritten?

[SHYuanBest]

Thanks for your interest! The self._pos_start determines the position from which the dataloader starts selecting elements; it cannot be omitted.

[Hickey8]

Thanks for your interest! The self._pos_start determines the position from which the dataloader starts selecting elements; it cannot be omitted.感谢您的关注！self._pos_start 确定数据加载器开始选择元素的位置;它不能省略。
I’ve been reviewing the logic related to self._pos_start, which is initialized to 0, and I’m currently unable to identify any scenario where it might exceed n. From my understanding, it seems to have no practical effect on data retrieval, but I’d really appreciate it if you could clarify clearly.
Additionally, while preparing to pair the video, bbox, and tracks_mask data to avoid runtime No such file or directory errors during training, I found that there are 2063 empty annotation folders under the tracks_masks_data directory. Could you please let me know whether this issue might have occurred during data upload on your side, or if it’s more likely related to my download? I've checked downloading twice.

[SHYuanBest]

When using resume, sampler_pos_start.pkl will be loaded, which will change self._pos_start. Detail is here :`https://github.com/PKU-YuanGroup/ConsisID/blob/main/train.py#L601. For the second question, this looks like a corrupted file on huggingface, we will fix this in the future.

[Hickey8]

When using resume, sampler_pos_start.pkl will be loaded, which will change self._pos_start. Detail is here :https://github.com/PKU-YuanGroup/ConsisID/blob/main/train.py#L601. For the second question, this looks like a corrupted file on huggingface, we will fix this in the future.使用 resume 时，会加载 sampler_pos_start.pkl，这会改变 self._pos_start`。详情在这里 ：'https://github.com/PKU-YuanGroup/ConsisID/blob/main/train.py#L601。第二个问题，这看起来像是 huggingface 上的文件损坏了，我们将来会修复这个问题。

Got it! Thanks a lot.