Question About thr Sampler Indices

[JoyHuYY1412]

Hi thank you for your code. I am reading your semi_sampler.py and find it a little bit confusing.

https://github.com/microsoft/SoftTeacher/blob/bef9a256e5c920723280146fc66b82629b3ee9d4/ssod/datasets/samplers/semi_sampler.py#L170

Why do you need to concatenate all the indices? I think the last indice already contains all the images in the dataset. And after concatenation, the indices become a list of #max_iteration arrays. I really appreciate your reply.

[MendelXu]

The last indice only contains one group of images in the dataset. However, each dataset usually contain two groups( w<h and w>h).

[JoyHuYY1412]

The last indice only contains one group of images in the dataset. However, each dataset usually contain two groups( w<h and w>h).

Thank you for your help!

I am not so familiar with mmdet, and I understand your code as: If the ratio is 1 and batch size is 2, the code works like: After creating the data loader that reading separately from labeled data and unlabeled data, one batch sampled from the dataset is contained of (<augmented labeled img, label>, <weak augmented unlabeled img, meaningless label>, <strong augmented unlabeled img, meaningless label> ) Although I am not so sure what happened in https://github.com/microsoft/SoftTeacher/blob/bef9a256e5c920723280146fc66b82629b3ee9d4/ssod/apis/train.py#L206 but I think the batched data is sent to the model and trained as in SoftTeacher.py.

Could you please point out my mistakes and give me some advice?

[MendelXu]

Yes, it works just like your description.

[JoyHuYY1412]

Yes, it works just like your description.

Thanks!

[JoyHuYY1412]

Yes, it works just like your description.

Hi sorry to bother you again, but I have a few more questions and your help is appreciated.

After using ordinary augmentation for labeled data, and using MultiBranch augmentation to generate the weak and strong augmented unlabeled images in

https://github.com/microsoft/SoftTeacher/blob/bef9a256e5c920723280146fc66b82629b3ee9d4/ssod/datasets/pipelines/rand_aug.py#L953-L965

, I think the labeled data format is a dict of {"img_metas":XXX, "img":XXX, "gt_bboxes":XXX, "gt_labels":XXX} , but the format of unlabeled data is a list of two augmented version like [ {"img_metas":XXX, "img":XXX, "gt_bboxes":XXX, "gt_labels":XXX} , {"img_metas":XXX, "img":XXX, "gt_bboxes":XXX, "gt_labels":XXX} ].

My questions are:

1. Am I right about the data format? Did I omit some important functions? Because I found MultiBrach is also noted in this function https://github.com/microsoft/SoftTeacher/blob/bef9a256e5c920723280146fc66b82629b3ee9d4/tools/misc/browse_dataset.py#L50-L73 , but I am not so clear what this function is for.
2. If I am right about the data format, how do you collate these two formats into a batch? Could you please point it out in the code?

I really appreciate your reply cause I am really lost...

[MendelXu]

See here. https://github.com/microsoft/SoftTeacher/blob/bef9a256e5c920723280146fc66b82629b3ee9d4/ssod/datasets/builder.py#L161 It will flatten the nested list to a one dimension list like [a,[b,c]]->[a,b,c].

[JoyHuYY1412]

See here.

https://github.com/microsoft/SoftTeacher/blob/bef9a256e5c920723280146fc66b82629b3ee9d4/ssod/datasets/builder.py#L161

It will flatten the nested list to a one dimension list like [a,[b,c]]->[a,b,c].

Thank you so much!