/deepspeed_mmdetection3d

Deepspeed integration with mmdetection3d

Primary LanguagePythonMIT LicenseMIT

DeepSpeed Integration for MMDetection3d

Setup

  • Init and update MMDetection3d submodule
git submodule init
git submodule update
git clone https://github.com/microsoft/DeepSpeed.git
cd DeepSpeed
DS_BUILD_OPS=1 ./install.sh
cd ..
  • Third, install NVIDIA Apex like:
git clone https://github.com/NVIDIA/apex
cd apex
pip install -v --disable-pip-version-check --no-cache-dir --global-option="--cpp_ext" --global-option="--cuda_ext" ./
cd ..
  • Symlink the Lyft dataset like
mkdir -p ./data/
# ln -s path/to/lyft/data /path/to/deepspeed_mmdetection3d/data/
ln -s /home/hd10/data/lyft /home/hd10/deepspeed_mmdetection3d/data/

Running

  • Run FP16 training like
nohup deepspeed --num_nodes 1 --num_gpus 1 ds/train.py --work-dir ./outputs_fp16/ --gpus 1 configs/ssn/ssn_lyft_base.py --deepspeed_config configs/ds/ds_config_fp16.json &
  • Run AMP training like
nohup deepspeed --num_nodes 1 --num_gpus 1 ds/train.py --work-dir ./outputs_amp/ --gpus 1 configs/ssn/ssn_lyft_base.py --deepspeed_config configs/ds/ds_config_amp.json &

For more options look at the DeepSpeed and AMP documentation.