Decoded structures appear severely distorted after encode then decode

[clxhsa]

Hi, thanks for open-sourcing this project!

I’m seeing a consistent issue during decoding: the reconstructed structures look severely distorted / collapsed, rather than resembling the input backbone.

To rule out preprocessing differences on my side, I also tested with the preprocessed CAMEO2024 dataset (.h5 format). But the decoded structures still look incorrect. I’ll attach screenshots for reference.

[mahdip72]

@clxhsa Hey,
Thanks for your interest in this project.
Hmmm, seems strange.
Will take a look at this as soon as I find some free time.
Meanwhile, have you tested the pip install version to see if it works? It makes the encoding and decoding much easier without having an H5 file.

[mahdip72]

@clxhsa
I tried to reproduce the same problem on my side, but I could not get the collapsed structure when I used a full saved model run directory. Because of that, I think the problem may come from the setup or from the model files being used, not from the decoder itself. In the repo, the scripts expect trained_model_dir to be a complete folder with config_vqvae.yaml, config_gcpnet_encoder.yaml, config_geometric_decoder.yaml, and checkpoints/best_valid.pth. In your screenshot, I saw gcp_vqvae/weight, so I am not fully sure if this folder contains the full correct model bundle, or if some config/checkpoint files are mixed from different sources.

Could you share a little more information so I can check this better? Please paste your full inference_encode_config.yaml and inference_decode_config.yaml, and also the output of ls -R gcp_vqvae/weight so I can see what files are inside that folder. It would also help a lot if you share the first few lines of your encoded CSV with head -n 5 of indices_csv_path file you set in the decoder config, and the full terminal output from both python inference_encode.py and python inference_decode.py.

Also please confirm if you used the repo scripts directly or the gcp-vqvae package, whether you downloaded the model from Hugging Face or from the zip checkpoint link, and whether the model was large or lite.

[clxhsa]

@mahdip72

inference_encode_config.yaml
inference_decode_config.yaml

encode.log
decode.log

Thanks for your reply!
I tried the inference_encode.py and inference_decode.py in the repo again, the full output is attached. The output structure is still wrong. I wrote another script with the gcp-vqvae package, but it still doesn't work:

run_cameo_gcp_vqvae_roundtrip.py

Here is the content of gcp_vqvae/weight:

The model used is the large version, downloaded from the zip checkpoint link.

[mahdip72]

@clxhsa
Thanks for sharing the files and logs. I checked them on my side. One thing I noticed is that your decode.log shows decoding finished from the CSV, but in the current master branch the raw output of inference_encode.py and the input expected by inference_decode.py do not fully match. Because of that, I think your local setup is not exactly the same as the current repo state. I also tested the public large model on my side, and it gave a normal reconstruction, so at this moment I do not think the published model itself is broken.

It is still possible that the problem is related to dependencies or environment differences, so I want to check that too. Also, in your package script, it still uses use_hf=False and points to the local gcp_vqvae/weight folder, so it is still testing the same local model bundle, not the Hugging Face version. Could you please share a little more information so I can narrow this down better? Please paste the output of git rev-parse HEAD, the first few lines of the actual vq_indices.csv file you used for decoding, and the versions of the main packages in your environment. For example, run and paste the output of:

git rev-parse HEAD

head -n 5 inference_encode_results/2026-03-31__08-11-03/vq_indices.csv

python --version
python -c "import torch; print('torch', torch.__version__)"
python -c "import torch_geometric; print('torch_geometric', torch_geometric.__version__)"
python -c "import pandas; print('pandas', pandas.__version__)"
python -c "import accelerate; print('accelerate', accelerate.__version__)"
python -c "import Bio; print('biopython', Bio.__version__)"
python -c "import graphein; print('graphein', graphein.__version__)"