search

Ivan-Tang-3D / ViewRefer3D

(ICCV2023) Official implementation of 'ViewRefer: Grasp the Multi-view Knowledge for 3D Visual Grounding with GPT and Prototype Guidance'
56 stars 5 forks source link

Question about your checkpoint #7

Open ericlenny opened 8 months ago

ericlenny commented 8 months ago

I download the checkpoint you mentioned by https://drive.google.com/drive/folders/1YqD7OklOl2rdXyG5aubLtEj6Jqth54Jc?usp=sharing. When I try to use your trained model to run the test.sh, I met some problems:

Traceback (most recent call last): File "/home/Harddisk/cc/ViewRefer3D/referit3d/scripts/train_referit3d.py", line 200, in loaded_epoch = load_state_dicts(args.resume_path, map_location=device, model=model) File "/home/Harddisk/cc/ViewRefer3D/referit3d/models/utils.py", line 61, in load_state_dicts value.load_state_dict(checkpoint[key]) File "/home/anaconda3/envs/viewrefer/lib/python3.7/site-packages/torch/nn/modules/module.py", line 1052, in load_state_dict self.class.name, "\n\t".join(error_msgs))) RuntimeError: Error(s) in loading state_dict for ReferIt3DNet_transformer: Missing key(s) in state_dict: "language_encoder.embeddings.position_ids".

How can I solve this problem？Thanks for your great work.

Ivan-Tang-3D commented 8 months ago

Have U downloaded the BERT ckpt, It seems related with language Encoder(BERT).

ericlenny commented 8 months ago

Have U downloaded the BERT ckpt, It seems related with language Encoder(BERT).

Yeah I used link https://huggingface.co/bert-base-uncased/tree/main to download it

Ivan-Tang-3D commented 8 months ago

This parameter 'language_encoder.embeddings.position_ids' is merely related with the BERT ckpt. So u could try to load the Bert CKPT to check whether this parameter exists or even not necessary to exist

ericlenny commented 8 months ago

This parameter 'language_encoder.embeddings.position_ids' is merely related with the BERT ckpt. So u could try to load the Bert CKPT to check whether this parameter exists or even not necessary to exist

I tried to train my own model yesterday and it worked very fine. The training process hasn't finished yet, but I think maybe not the Bert CKPT problem