Skip to content

Usage.md

Latest commit

 

History

History
83 lines (61 loc) · 2.72 KB

Usage.md

File metadata and controls

83 lines (61 loc) · 2.72 KB
Raw

Usage Guide

We have tested HERMES on NVIDIA H20 servers and recommend using GPUs with at least 80GB memory for smooth operation.
We plan to release DeepSpeed deployment support in the future to further reduce GPU memory usage.

If you are using NVIDIA H20 96GB, you may remove some checkpoint usage in hermes.py and hermes_future_render_head.py for improved efficiency.

1. Training

Multi-GPU / Multi-Node Training

torchrun --nproc_per_node $PROC_PER_NODE \
        --master_addr $MASTER_ADDR \
        --master_port $MASTER_PORT \
        --nnodes $NODE_COUNT \
        --node_rank $NODE_RANK \
        extra_tools/train.py \
        projects/configs/hermes/stage3.py \
        --work-dir ./work-dir \
        --launcher pytorch \
        --no-validate

Single-GPU Debugging

python extra_tools/train.py projects/configs/hermes/stage3.py --work-dir ./work-dir --no-validate

You can modify different config files to train different stages as needed.

Checkpoint Cleanup (Optional)

Since we use mmcv for training pipeline management, the saved checkpoints include the frozen LLM pretrained parameters, which can be large.
You can use our provided script to remove unnecessary parts and save storage space:

python extra_tools/ckpt_convertor.py ./path/to/your_custom_trained.pth --delete_optimizer --save_path path/to/your_custom_trained_cleaned.pth

Both --delete_optimizer and --save_path are optional.

2. Inference (Testing)

torchrun --nproc_per_node $PROC_PER_NODE \
        --master_addr $MASTER_ADDR \
        --master_port $MASTER_PORT \
        --nnodes $NODE_COUNT \
        --node_rank $NODE_RANK \
        extra_tools/test.py \
        projects/configs/hermes/stage3.py \
        ckpt/hermes_final.pth \
        --launcher pytorch
  • The results for each scenario will be saved under outputs/stage3/hermes_eval/results.
  • As nuScenes contains a large number of scenes, we recommend using multi-GPU/multi-node inference.

3. Evaluation

We evaluate the results by reading the saved JSON files and calculating Chamfer Distance and text metrics.

Install necessary libraries:

pip install pycocoevalcap nltk openai pyfiglet
python -c "import nltk; nltk.download('punkt'); nltk.download('wordnet'); nltk.download('punkt_tab')"

Run evaluation

python extra_tools/eval_hermes_results.py ./outputs/stage3/hermes_eval/results

Due to the nature of LLMs, inference results may vary slightly each time.

For more details and custom configuration, please refer to the config files.