omegaconf.errors.ConfigAttributeError: ListConfig does not support attribute access

[HoWingYip]

When I try to run inference using APE-Ti, I get the following error:

Traceback (most recent call last):
  File "/home/jupyter/TIL-2024/vlm/train/APE/demo/demo_lazy.py", line 134, in <module>
    demo = VisualizationDemo(cfg, args=args)
  File "/home/jupyter/TIL-2024/vlm/train/APE/demo/predictor_lazy.py", line 138, in __init__
    self.metadata = MetadataCatalog.get(
  File "/home/jupyter/.local/lib/python3.9/site-packages/omegaconf/listconfig.py", line 176, in __getattr__
    self._format_and_raise(
  File "/home/jupyter/.local/lib/python3.9/site-packages/omegaconf/base.py", line 190, in _format_and_raise
    format_and_raise(
  File "/home/jupyter/.local/lib/python3.9/site-packages/omegaconf/_utils.py", line 821, in format_and_raise
    _raise(ex, cause)
  File "/home/jupyter/.local/lib/python3.9/site-packages/omegaconf/_utils.py", line 719, in _raise
    raise ex.with_traceback(sys.exc_info()[2])  # set end OC_CAUSE=1 for full backtrace
omegaconf.errors.ConfigAttributeError: ListConfig does not support attribute access
    full_key: dataloader.train[dataset]
    object_type=list

Error occurs on line 138 of demo/predictor_lazy.py when accessing cfg.dataloader.train.dataset.names.
cfg.dataloader.train is a ListConfig of 9 datasets.
When inferring on APE-D, cfg.dataloader.train appears to be a single dataset, and I don't get this error.

The command I run for APE-Ti is:

python3 demo/demo_lazy.py \
--config-file configs/LVISCOCOCOCOSTUFF_O365_OID_VGR_SA1B_REFCOCO_GQA_PhraseCut_Flickr30k/ape_deta/ape_deta_vitt_eva02_vlf_lsj1024_cp_16x4_1080k_mdl.py \
--input demo/examples/DesertStorm.jpeg \
--output demo/examples_output/DesertStorm.jpeg/ \
--confidence-threshold 0.1 \
--text-prompt 'white fighter jet' \
--with-box \
--opts \
train.init_checkpoint=configs/LVISCOCOCOCOSTUFF_O365_OID_VGR_SA1B_REFCOCO_GQA_PhraseCut_Flickr30k/ape_deta/ape_deta_vitt_eva02_vlf_lsj1024_cp_16x4_1080k_mdl_20240203_230000/model_final.pth \
model.model_vision.select_box_nums_for_evaluation=500 \
model.model_vision.text_feature_bank_reset=True

For APE-D, I run:

python3 demo/demo_lazy.py \
--config-file configs/LVISCOCOCOCOSTUFF_O365_OID_VGR_SA1B_REFCOCO_GQA_PhraseCut_Flickr30k/ape_deta/ape_deta_vitl_eva02_clip_vlf_lsj1024_cp_16x4_1080k.py \
--input demo/examples/DesertStorm.jpeg \
--output demo/examples_output/DesertStorm.jpeg/ \
--confidence-threshold 0.1 \
--text-prompt 'white fighter jet' \
--with-box \
--opts \
train.init_checkpoint=configs/LVISCOCOCOCOSTUFF_O365_OID_VGR_SA1B_REFCOCO_GQA_PhraseCut_Flickr30k/ape_deta/ape_deta_vitl_eva02_clip_vlf_lsj1024_cp_16x4_1080k_mdl_20230829_162438/model_final.pth \
model.model_vision.select_box_nums_for_evaluation=500 \
model.model_vision.text_feature_bank_reset=True

Any help is appreciated. Thanks.

[shenyunhang]

During inference, do not use the config file that ends with _mdl.py, which is only used during training.

Try to replace

configs/LVISCOCOCOCOSTUFF_O365_OID_VGR_SA1B_REFCOCO_GQA_PhraseCut_Flickr30k/ape_deta/ape_deta_vitt_eva02_vlf_lsj1024_cp_16x4_1080k_mdl.py 

with

configs/LVISCOCOCOCOSTUFF_O365_OID_VGR_SA1B_REFCOCO_GQA_PhraseCut_Flickr30k/ape_deta/ape_deta_vitt_eva02_vlf_lsj1024_cp_16x4_1080k.py 

.