penghui-yang/l2d

Official Implementation of ICCV'23 paper "Multi-Label Knowledge Distillation"

l2d's Issues

Could you release the model weights of the pre-trained teacher model?

@penghui-yang
Could you release the model weights of the pre-trained teacher model on the multi-label classification dataset? Especially on COCO and VOC.

coco 类别名称

dataloader.py 中的 class_name 好像和appendix 的anno 索引对应不上。

Pretrained teacher models

Could the pretrained teacher models be downloaded?

questions about the ML_decoder

Hi, awesome work. I have a few questions about the ML_decoder.

I noticed that during the initialization, the pre-trained teacher network is loaded from state_dict, and then the fc is discarded and replaced by a randomly initialized ML_decoder, in the following lines.

L2D/models/tools.py

Line 19 in d305c3d

def add_ml_decoder_head(model, num_classes=-1, num_of_groups=-1, decoder_embedding=768):

Does that mean the ML_decoder part of the teacher network is not pre-trained, since the teacher network is fixed, will it produce random results for distillation?