Default config.json Issue

[MXXXZ]

Hi!

Due to network environment issues, I am unable to directly download the model weights through the installation script. Therefore, I used the download script mentioned in Issue #39 for direct URL downloading. (That is: ["original"]="https://zenodo.org/record/7733060/files/original.zip", ["original_no_seq"]="https://zenodo.org/record/7733060/files/original_no_seq.zip", ["small_gpu"]="https://zenodo.org/record/7733060/files/small_gpu.zip")

When running Build, with using --model-bundle-path path_to_download_weights/original for execution, I encountered an error:

AttributeError: 'Args' object has no attribute 'box_size'

After attempting to add '"box_size": 64, 'to the original/config.json and running it again, I encountered another error:

AttributeError: 'Args' object has no attribute 'voxel_size'

As a result, I'm wondering if the issues are due to the default config.json in the weights I downloaded not being updated, or because I need to specify additional parameters myself?

There is my original config.json in the original package:

{
  "standardize_mrc_args":
    {
      "target_voxel_size": 1.5,
      "crop_z": 0,
      "bfactor_to_apply": 0,
      "auto_mask": false
    },
  "ca_infer_args":
    {
      "model_checkpoint": "chkpt.torch",
      "bfactor": 0,
      "batch_size": 4,
      "stride": 16,
      "dont_mask_input": true,
      "threshold": 0.05,
      "save_real_coordinates": false,
      "save_cryo_em_grid": false,
      "do_nucleotides": false,
      "save_backbone_trace": false,
      "save_ca_grid": false,
      "crop": 6
    },
  "gnn_infer_args":
    {
      "num_rounds": 3,
      "crop_length": 200,
      "repeat_per_residue": 3,
      "esm_model": "esm1b_t33_650M_UR50S",
      "aggressive_pruning": false,
      "seq_attention_batch_size": 200
    }
}

Thank you for your help!

Mency

[kvr2007]

I was experiencing exactly the same issue (our HPC does is protected from the outside), but could figure it out :)

Download the files from Zenodo (the list is taken from model_angelo/utils/torch_utils.py):

        "original": "https://zenodo.org/record/7733060/files/original.zip",
        "original_no_seq": "https://zenodo.org/record/7733060/files/original_no_seq.zip",
        "small_gpu": "https://zenodo.org/record/7733060/files/small_gpu.zip",
        "nucleotides": "https://zenodo.org/record/7942241/files/nucleotides.zip",
        "nucleotides_no_seq": "https://zenodo.org/record/7942241/files/nucleotides_no_seq.zip",

Create a folder to store all models and save it to the environment variable TORCH_HOME:

export TORCH_HOME=/path/to/folder/

Create additional folders:

mkdir -p $TORCH_HOME/hub/checkpoints/model_angelo_v1.0

Unzip the archives to $TORCH_HOME/hub/checkpoints/model_angelo_v1.0, then you need to create success.txt files in each model's folder:

find $TORCH_HOME/hub/checkpoints/model_angelo_v1.0 -name config.json -execdir bash -c "echo Successfully downloaded model > success.txt" \;

Lastly, you need to download two other models:

https://dl.fbaipublicfiles.com/fair-esm/models/esm1b_t33_650M_UR50S.pt
https://dl.fbaipublicfiles.com/fair-esm/regression/esm1b_t33_650M_UR50S-contact-regression.pt

And save them to $TORCH_HOME/hub/checkpoints/