Qing Lian

Hi @kaixinbear , Could you try to replace the code in [train.py ](https://github.com/open-mmlab/mmdetection3d/blob/master/tools/train.py#L171) from ``` # init distributed env first, since logger depends on the dist info. if args.launcher ==...

@kouxichao Have you set the model to eval mode when generating the pseudo label code?

@kouxichao Have you also set the model to train model --> model.train() after generating pseudo bounding box in every iteration, which may solve the problem.

I can get similar results from the released code in the validation dataset. Thanks for kindly sharing the code.

I see that. Thank you for your update. : )

Should I use the `loss_weights_file` during training? I found that I only can get 7% mIoU on source only setting. It must something wrong. I am debugging. : (

Hi Junjie, Could we also record the results of bevdet4d_stt_tiny_cbgs. The following are my reproduced results ran on 8 GPU. I am not sure if the performance gap is due...

@bkj Hello, Did you change any parameters on the scripts to get this result. I only could get almost 0.88 acc on cifar using micro structure.

I have fixed the code style issues. Regarding to "[Question] line 286, 294: will this cause problem for inference of encoder-decoder models?", the pad_token_id argument is added to the kwargs,...

I have updated the demo with multi-modality conversation function: The example script is: ``` export model="Salesforce/blip2-opt-2.7b" deepspeed examples/inference.py \ --deepspeed configs/ds_config_multimodal.json \ --model_name_or_path ${model} \ --arch_type vision_encoder_decoder \ --input_text "please...