MAMP self-supervised pre-training on 90k iMiGUE 46-joint pose skeletons, followed by full fine-tuning (train+val) and cRT classifier re-training on the iMiGUE-32 gesture classification task.

This model achieves 70.188% Top-1 on the official test set, and contributes to our final 74.66% ensemble (1st on Kaggle).

📦 1. Installation

Download the entire bundle (code + pretrain NPZ + finetune PKL + pretrain ckpt + finetune ckpt + cRT classifier ckpt) from HuggingFace:

hf download <account>/MiGA26_Track1_Skeleton_pose90k_FT --repo-type dataset --local-dir ./ --token <TBD>

Then install Python deps. The verified environment is Python 3.8 with CUDA 11.1, torch==1.8.1+cu111, torchvision==0.9.1+cu111, and timm==0.3.2; install the matching CUDA-enabled PyTorch build for your machine before or while installing the remaining packages.

pip install -r requirements.txt

We follow MAMP's environment. See the upstream environment file at https://github.com/maoyunyao/MAMP for the full conda spec.

📂 2. Data Preparation

After hf download, the bundle layout is:

Skeleton_pose90k_ft/
├── data/
│   ├── pretrain.npz         # 19G, 90k iMiGUE 46-joint pose (root1) for MAMP pre-training
│   └── finetune.pkl         # 625M, iMiGUE-32 skeleton dataset for fine-tuning
├── weights/
│   ├── checkpoint-399.pth   # 119M, MAMP pre-trained encoder (400 epochs)
│   ├── checkpoint-99.pth    # 350M, train+val fine-tuned classifier (100 epochs)
│   └── official_crt/
│       └── final_model_checkpoint.pth  # cRT classifier used for the 70.188% submission
├── config/
│   ├── pretrain.yaml
│   ├── finetune.yaml
│   └── test.yaml
├── feeder/
├── model/
├── model_mamp/
├── util/
├── tools/
├── scripts/
│   ├── pretrain.sh
│   ├── finetune.sh
│   ├── crt.sh
│   └── test.sh
├── main_pretrain.py
├── main_finetune.py
├── engine_pretrain.py
├── engine_finetune.py
├── test_list.json
├── requirements.txt
└── README.md

The pre-built data/pretrain.npz is the 70k-from-iMiGUE-tarballs + 20k-random-subsampled-from-120k merged pose archive in root-translation-normalized 46-joint format. The pre-built data/finetune.pkl contains the official iMiGUE-32 train/val/test splits in the same skeleton format. Both are ready to consume without further processing.

Note. Reproducing the 70.188% submission only needs data/finetune.pkl plus the bundled weights/checkpoint-99.pth and weights/official_crt/final_model_checkpoint.pth (see §3.5). data/pretrain.npz (~19 GB) is only required if you want to redo the self-supervised pre-training stage (§3.2) yourself.

🏋️‍♂️ 3. Training & Testing

3.1 Pre-trained Weights

Stage	Source data	Epochs	File	Size	Test Top-1
Pre-train	90k iMiGUE pose (root1)	400	`weights/checkpoint-399.pth`	119M	n/a
Fine-tune	iMiGUE-32 train+val	100	`weights/checkpoint-99.pth`	350M	intermediate
cRT	frozen fine-tuned features + class-aware sampler	10	`weights/official_crt/final_model_checkpoint.pth`	117M	70.188%

All listed weights ship inside the HuggingFace bundle, so you only need to run pre-training, fine-tuning, or cRT if you want to reproduce the training process yourself.

The three stages below are independent and use the bundled ./weights/ as the default input source. Each stage can be re-run on its own without re-running the previous one.

3.2 Pre-training (optional, reproduce)

Trains MAMP encoder-decoder (8+5 layers, mask ratio 0.9, motion-aware tau 0.80) for 400 epochs on 4 GPUs.

Input: ./data/pretrain.npz
Output: ./output_dir/pretrain/checkpoint-399.pth

bash scripts/pretrain.sh

Override CUDA_VISIBLE_DEVICES, NPROC_PER_NODE, MASTER_PORT, OUT_DIR, or CONFIG via environment variables.

3.3 Fine-tuning (optional, reproduce)

Fine-tunes the pre-trained encoder on the iMiGUE-32 train+val split for 100 epochs.

Input: ./data/finetune.pkl and pre-trained ckpt (default ./weights/checkpoint-399.pth)
Output: ./output_dir/finetune/checkpoint-99.pth

bash scripts/finetune.sh
# or use a self-trained pretrain ckpt:
PRETRAIN_CKPT=./output_dir/pretrain/checkpoint-399.pth bash scripts/finetune.sh

3.4 cRT classifier re-training (optional, reproduce)

Re-trains only the classifier on frozen checkpoint-99.pth features with the official classifier-balancing cRT recipe.

Input: ./data/finetune.pkl and fine-tuned ckpt (default ./weights/checkpoint-99.pth)
Output: ./output_dir/crt/final_model_checkpoint.pth

bash scripts/crt.sh
# or use a self-trained fine-tune ckpt:
FINETUNE_CKPT=./output_dir/finetune/checkpoint-99.pth bash scripts/crt.sh

3.5 Test (submission CSV)

Runs cRT inference on the iMiGUE-32 test split and writes the submission CSV in official test-list order.

Input: ./data/finetune.pkl, ./test_list.json, fine-tuned ckpt (default ./weights/checkpoint-99.pth), cRT ckpt (default ./weights/official_crt/final_model_checkpoint.pth)
Output: ./output_dir/test/test_pred_official_crt_testlist_order.csv

bash scripts/test.sh
# or use a self-trained finetune ckpt:
FINETUNE_CKPT=./output_dir/finetune/checkpoint-99.pth \
CRT_CKPT=./output_dir/crt/final_model_checkpoint.pth \
bash scripts/test.sh

Running bash scripts/test.sh directly on the bundled ./weights/checkpoint-99.pth and ./weights/official_crt/final_model_checkpoint.pth reproduces our 70.188% cRT submission.

📊 4. Results

Model	Modality	Test Top-1
This model (Skeleton MAMP, 90k pose pretrain)	Skeleton	70.188%
Skeleton MAMP, MA-52 pretrain (see `../Skeleton_pose_ma52_ft`)	Skeleton	59.688%
RGB official (see `../RGB_Official_Code`)	RGB	69.224%
... and 3 PoseConv3D / VideoSwinT baselines	mixed	-
Ensemble (6 models, see `../ensemble`)	all	74.66% (1st on Kaggle)

🙏 5. Acknowledgement

This codebase is built on top of MAMP (Motion-Aware Masked autoencoder for skeleton-based action recognition) and classifier-balancing for cRT. We thank the authors for releasing their code.

📧 6. Contact

For questions about this skeleton branch, please open an issue at the project repository or contact the MiGA26 Track 1 team.

Downloads last month: 4

Total file size:

20.8 GB