Skip to content

Latest commit

 

History

History
75 lines (52 loc) · 3.31 KB

FAQ.md

File metadata and controls

75 lines (52 loc) · 3.31 KB

Q1: [Error] AttributeError: module 'distutils' has no attribute 'version'

Fix:

pip install setuptools==59.5.0

Ref: pytorch #69894

Q2: How to setup on Ampere GPUs (e.g., A100, A800)?

As informed here, please pick a compatible gcc version and cuda version. Note that, you should use cuda 11+ for Ampere GPUs.

We tested with python=3.8, gcc-0.9.3, g++-0.9.3, and cuda-11.3. The final environment file can be find requirements/a800.yaml, where mmdet3d from third_party/bevfusion is not included since we install it with develop mode.

Otherwise, here is a step-by-step guide to setup the environment above.

# python3.8, cuda113
pip install https://download.pytorch.org/whl/cu113/torch-1.10.2%2Bcu113-cp38-cp38-linux_x86_64.whl
pip install https://download.pytorch.org/whl/cu113/torchvision-0.11.3%2Bcu113-cp38-cp38-linux_x86_64.whl
pip install https://download.openmmlab.com/mmcv/dist/cu113/torch1.10.0/mmcv_full-1.4.5-cp38-cp38-manylinux1_x86_64.whl

# now, you need to comment torch, torchvision, and mmcv_full in 
# `requirements/dev.txt`, and run
pip install -r requirements/dev.txt

cd third_party/diffusers
pip install .

cd third_party/bevfusion
python setup.py develop

Now, you should be able to run our demo.

Q3: [Error] nvcc fatal : Unsupported gpu architecture 'compute_80'

This may appear when you install bevfusion (mmdet3d) on cuda10.2. The latest version of bevfusion supports Ampere GPUs by hard-coding compile parameters, leading to error when compiled with cuda10.2. One can get rid of this error by comment these lines in third_party/bevfusion/setup.py (L19).

"-D__CUDA_NO_HALF2_OPERATORS__",
"-gencode=arch=compute_70,code=sm_70",
"-gencode=arch=compute_75,code=sm_75",
# "-gencode=arch=compute_80,code=sm_80",
# "-gencode=arch=compute_86,code=sm_86",

Q4: video generation

The authors do not have plan to integrate video generation with main branch codebase. For those who need code for video generation, please find this in video branch.

Q5: AttributeError: 'BasicTransformerBlock' object has no attribute '_args'

#7. This can occur when you install the official version of diffusers rather than from third_party/diffusers. To install the local version, please make sure you use the following commands:

# if you have install, remove first
pip uninstall diffusers

# install our version
cd third_party/diffusers
pip install -vvv .

Q6: model loading path vs accelerate checkpoints

#25. There are two formats for model saving.

  • checkpoints-xxxx: saved by accelerate, default by runner.checkpointing_steps: 5000. This is used to resume training process.
  • weight-Exx-Sxxxxx: saved by diffusers, default is on the end of each epoch/model training. This is used to load for testing and evaluation.

If you want to change the frequency for saving diffusers format for testing, please invoke this line manual in your code.