Manual Installation
If some packages fail to install using the standard methods due to network issues or other reasons, this guide provides detailed instructions for manual installation.
Setting Up Conda Environment and CUDA 12.1
# Create a new conda environment
conda create -n fastsam3d python=3.11
# Activate the environment
conda activate fastsam3d
# Install dependencies from YAML file
conda env update -f environments/default.yml
# Configure PyPI indexes for NVIDIA and PyTorch CUDA packages
export PIP_EXTRA_INDEX_URL="https://pypi.ngc.nvidia.com https://download.pytorch.org/whl/cu121"
Installing requirements.txt and dev.txt
# Install the package in development mode with dev dependencies
pip install -e '.[dev]'
Installing PyTorch for Python 3.11 and CUDA 12.1
# Install PyTorch with CUDA 12.1 support
pip install torch torchvision torchaudio --index-url https://download.pytorch.org/whl/cu121
Manually Installing utils3d
Download utils3d from GitHub to your local machine:
# Clone the utils3d repository
git clone https://github.com/EasternJournalist/utils3d.git
# Navigate to the repository
cd utils3d
# Checkout the specific commit
git checkout 3913c65d81e05e47b9f367250cf8c0f7462a0900
Transfer the utils3d directory to the server at MoGe/utils3d and install manually:
# Navigate to utils3d directory
cd utils3d
# Install the package (NOT in editable mode)
pip install .
Important: Do NOT use
pip install . -e(editable mode)
Manually Installing MoGe
MoGe depends on utils3d and is not available on PyPI, so it must be downloaded from Git.
# 1. Comment out the utils3d Git dependency in MoGe/requirements.txt:
# Original line: git+https://github.com/EasternJournalist/utils3d.git@3913c65d81e05e47b9f367250cf8c0f7462a0900
# Change to: # git+https://github.com/EasternJournalist/utils3d.git@3913c65d81e05e47b9f367250cf8c0f7462a0900
# 2. In pyproject.toml, change:
# "EasternJournalist/utils3d.git@3913c65d81e05e47b9f367250cf8c0f7462a0900"
# To:
# "utils3d"
# (since we've already installed utils3d)
# Navigate to MoGe directory and install
cd MoGe
pip install .
Manually Installing gsplat
First, install the required build tools to avoid no-build-isolation errors:
pip install hatchling hatch-requirements-txt
# Navigate to gsplat directory and compile
cd gsplat
MAX_JOBS=4 pip install . --no-build-isolation -v
Successful compilation output: Successfully installed gsplat-1.5.3 jaxtyping-0.3.3 wadler-lindig-0.1.7
Manually Installing pytorch3d
First, check if these dependencies are already installed:
pip show fvcore
pip show iopath
# Navigate to pytorch3d directory and compile
cd pytorch3d
export FORCE_CUDA=1
export TORCH_CUDA_ARCH_LIST="8.0"
MAX_JOBS=4 pip install . --no-build-isolation -v
Successfully installed: pytorch3d-0.7.8
Installing inference.txt Dependencies
Install these manually (avoid pip install -e '.[inference]' as it reinstalls requirements.txt):
# Configure Kaolin download link
export PIP_FIND_LINKS="https://nvidia-kaolin.s3.us-east-2.amazonaws.com/torch-2.5.1_cu121.html"
# Install inference dependencies
pip install kaolin==0.17.0
conda install -c conda-forge "numpy>=2,<2.3.0"
pip install seaborn==0.13.2
pip install gradio==5.49.0
Compiling flash-attention
Installing [p3d] would reinstall requirements.txt, so we install flash-attention manually.
Note: This compilation is time-consuming (approximately 1 hour).
Default installation method:
pip install -e '.[p3d]'
Alternative (recommended):
MAX_JOBS=4 pip install flash_attn==2.8.3 --no-build-isolation -i https://pypi.tuna.tsinghua.edu.cn/simple
Reinstalling requirement.txt (if some packages failed)
If some packages from requirements.txt were not installed successfully:
pip install -r requirement.txt