[Pinned] Installation & FAQ — read this first

[nankingjing]

PoseMamba — Installation & FAQ

Official troubleshooting thread · AAAI 2025 · Paper · README

Before opening a new issue, please check this thread. Most past issues are covered here.

---

1. Environment setup (recommended)

conda create -n posemamba python=3.8.5
conda activate posemamba
pip install torch==1.13.1+cu117 torchvision==0.14.1+cu117 torchaudio==0.13.1 --extra-index-url https://download.pytorch.org/whl/cu117
pip install -r requirements.txt
cd kernels/selective_scan && pip install -e . && cd ../..

You do NOT need to clone the official Mamba repo. Use kernels/selective_scan included in this repository.

---

2. FAQ

Q1: selective_scan_cuda_core is not defined

CUDA extension not built correctly. Rebuild:

cd kernels/selective_scan
pip uninstall selective-scan -y 2>/dev/null || true
pip install -e .
cd ../..

Ensure PyTorch/CUDA versions match (tested: torch 1.13.1 + cu117). On CUDA 12.x, try editing setup.py:

MODES = ["core", "oflex"]

---

Q2: kernels/selective_scan not found

The folder is in the repo root. Re-clone the full repository — do not use incomplete zip downloads.

ls kernels/selective_scan/setup.py

---

Q3: Dataset not found (h36m_sh_conf_cam_source_final.pkl)

1. Download preprocessed H36M data from README links → data/motion3d/
2. Run slicing:

cd tools && python convert_h36m.py && cd ..

3. Verify:

ls data/motion3d/MB3D_f243s81/h36m_sh_conf_cam_source_final.pkl

---

Q4: Evaluation command / checkpoint format

There is no --eval-only flag. Use --evaluate:

CUDA_VISIBLE_DEVICES=0 python train.py \
  --config checkpoint/pose3d/PoseMamba_B/config.yaml \
  --evaluate checkpoint/pose3d/PoseMamba_B/best_epoch.bin \
  --checkpoint eval/checkpoint

Source	Format	Notes
Training output	best_epoch.bin	dict with model_pos key
Google Drive release	.pth.tr	load with torch.load() — do not unzip

If config.yaml is missing in checkpoint dir (pretrained-only eval):

mkdir -p checkpoint/pose3d/PoseMamba_L
cp configs/pose3d/PoseMamba_train_h36m_L.yaml checkpoint/pose3d/PoseMamba_L/config.yaml

See eval.sh for S/B/L examples.

---

Q5: _pickle.UnpicklingError when loading weights

Usually PyTorch version mismatch. Use README-tested torch 1.13.1, or:

torch.load('weights.pth', map_location='cpu')

---

Q6: Multi-GPU / distributed training

Not officially supported. Use single GPU:

CUDA_VISIBLE_DEVICES=0 python train.py --config configs/pose3d/PoseMamba_train_h36m_S.yaml --checkpoint checkpoint/pose3d/PoseMamba_train_h36m_S

nn.DataParallel with custom CUDA ops often causes device mismatch errors.

---

Q7: Results gap vs paper (~1–2 mm)

Checklist:

• Correct model variant (S / B / L) and matching config
• Flip test enabled (flip: True in config)
• Same metric protocol (detected 2D vs GT 2D)
• kernels/selective_scan built on current machine
• Full 120 epochs if training from scratch

Reopen an issue with: config yaml, eval command, and MPJPE log if gap persists.

---

Q8: Training time per epoch

With default batch_size=4, clip_len=243: ~10–15 min/epoch on 3090/A100 is normal. Speed up with larger batch (if VRAM allows), num_workers=4–8, and SSD for data.

---

Q9: MACs computation

We use thop (see lib/model/PoseMamba.py). Paper MACs = thop FLOPs ÷ 2, consistent with MotionBERT / MotionAGFormer reporting.

---

Q10: COCO / custom 2D training

PoseMamba is 2D-to-3D lifting (H36M / 3DHP). For wild video, vis.py uses an external 2D detector + PoseMamba 3D lifting. Native COCO 2D-only training is out of current scope.

---

3. Quick links

• Demo: python vis.py --video sample_video.mp4 --gpu 0
• Configs: configs/pose3d/ (not configs/h36m)
• Pretrained weights: Google Drive
• Author: Google Scholar

---

4. Still stuck?

Open a new issue with:

1. OS + GPU
2. python -V / torch.__version__ / nvcc --version
3. Full command and traceback
4. Config yaml (or checkpoint path)

We'll respond as soon as possible. Thanks for using PoseMamba!