search

cure-lab / MagicDrive

[ICLR24] Official implementation of the paper “MagicDrive: Street View Generation with Diverse 3D Geometry Control”
https://gaoruiyuan.com/magicdrive/
GNU Affero General Public License v3.0
664 stars 40 forks source link

Question about initialize class_tokens #70

Closed ailearnerkyh closed 3 months ago

ailearnerkyh commented 3 months ago

When I run the command 

python tools/test.py resume_from_checkpoint=${RUN_LOG_DIR} task_id=${ANY} \
    runner.validation_times=4 runner.pipeline_param.init_noise=rand_all \
    ++dataset.data.val.ann_file=${ROOT}/../data/nuscenes_mmdet3d-12Hz/nuscenes_interp_12Hz_infos_val.pkl

The warning [ContinuousBBoxWithTextEmbedding] Your class_tokens is not trainable and used without initialization. Please check your training code! ] occurs. Besides, when I train the code, I can't find your class_tokens initialized with random in the INFO, nor can I find Initializing your class_tokens with text_encoder. Is this because I use the checkpoint SDv1.5mv-rawbox_2023-09-07_18-39_224x400, so the class_tokens have no need to be initialized? Looking forward to your reply.

flymin commented 3 months ago

If you load a pretrained model with this param, this may occur but it is fine. Otherwise, if you see the warning, please check your code.