[Request] Environment setup challenges & Request for Dockerfile for reproducible inference

[vegetableno1]

Hi MeshCoder team,

First of all, thank you for this impressive work! I am very excited to try out the inference_shape2code_point_input.py script with the Llama-3.2-1B model.

However, I have encountered a series of environment compatibility issues while trying to set up the dependencies following the current documentation. It seems that the version constraints between torch, peft, transformers, and pytorch3d are quite strict and somewhat conflicting in a standard setup.

Here is a summary of the issues I faced:

1. Version Incompatibility: Conflicts between peft and transformers (e.g., ModuleNotFoundError: No module named 'transformers.modeling_layers') and issues with LossKwargs imports when using Llama 3.2.

2. Binary Mismatches: torch-scatter and pytorch3d required manual recompilation or specific wheels to match the PyTorch version, which caused undefined symbol errors.

3. Missing System Libraries: The inference script (relying on OpenCV and Blender) failed multiple times due to missing Linux libraries in a standard container environment (e.g., libGL.so.1, libXi.so.6, libxkbcommon.so.0, etc.).

Suggestion:

To make it easier for the community to reproduce your results and avoid "dependency hell," would it be possible to provide an official Dockerfile?

A Docker image containing the correct versions of PyTorch, PyTorch3D, pre-compiled binaries, and all necessary system-level libraries would be incredibly helpful.

Thank you for your time and for open-sourcing this project!