[Reimplementation] Label Assignment Distillation

[alaa-shubbak]

Prerequisite

• [X] I have searched Issues and Discussions but cannot get the expected help.
• [X] I have read the FAQ documentation but cannot get the expected help.
• [X] The bug has not been fixed in the latest version (master) or latest version (3.x).

💬 Describe the reimplementation questions

Hello , I am trying to run a distillation model of lad (Label Assignment Distillation). I want to use the same distillation structure but with different model rather tan the paa model. this model have similar backbone as well as similar neck , but with different head.

The dataset i used is the COCO dataset.

I got this error message:

ValueError: Fitting the mixture model failed because some components have ill-defined empirical covariance (for instance caused by singleton or collapsed samples). Try to decrease the number of components, or increase reg_covar.

I am not sure what is exactly the issue , although i used same structure with much smaller dataset, and it works.

Environment

here is my environment , i create it using python environment tools , not with conda.

Expected results

I think i can train the model of distillation structure with different types of detectors .

is there any document that helps in understand how to use distillation structure (teacher and strudent models ) and change the config according to different models ?

I will be so thankful if so.

Additional information

No response

[alaa-shubbak]

hello I think i solved the problem. the model start running and training. But i have another constrain.

the loss of both bbox and global loss become negative values. here is it

as well as the mAP and the accuracy from the beginning of the training until epoch 13 still at same value 0.267

[ZwwWayne]

Could you provide the config you use?