MAMP self-supervised pre-training on MA-52 49-joint DWPose skeletons, followed by full fine-tuning (train+val) and cRT classifier re-training on the iMiGUE-32 gesture classification task (36 keypoints zero-padded to 49).

This model achieves 59.688% Top-1 on the official test set (finetune-only historical baseline) and contributes to our final 74.66% ensemble (1st on Kaggle).

📦 1. Installation

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

hf download <虚位以待>/MiGA26_Track1_Skeleton_pose_ma52_FT --repo-type dataset --local-dir ./ --token <虚位以待>

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_pose_ma52_ft/
├── data/
│   └── finetune.pkl         # iMiGUE-32 skeleton dataset (36 keypoints) for fine-tuning
├── weights/
│   ├── checkpoint-399.pth   # 119M, MAMP pre-trained encoder on MA-52 (400 epochs)
│   ├── checkpoint-99.pth    # 368M, train+val fine-tuned classifier (100 epochs) -> 59.688%
│   └── official_crt/        # cRT classifier (not bundled, see §3.4)
├── config/
│   ├── pretrain.yaml
│   ├── finetune.yaml
│   └── test.yaml
├── feeder/
├── model/
├── model_mamp/
├── util/
├── tools/
├── scripts/
│   ├── pretrain.sh
│   ├── finetune.sh
│   ├── crt.sh
│   └── test.sh
├── third_party/classifier-balancing/  # official cRT code (BSD)
├── main_pretrain.py
├── main_finetune.py
├── engine_pretrain.py
├── engine_finetune.py
├── test_list.json
├── requirements.txt
└── README.md

The pre-built data/finetune.pkl contains the official iMiGUE-32 train/val/test splits in 36-joint DWPose format. The feeder zero-pads to 49 joints before feeding the model. This file is ready to consume without further processing.

Note. The MA-52 hybrid49 pre-training NPZ used to produce checkpoint-399.pth is not bundled. The pretrain stage (§3.2) is therefore not reproducible from this bundle out of the box until you place your own MA-52 49-joint DWPose NPZ at ./data/ma52.npz and update config/pretrain.yaml. You can also skip the pre-training stage and start from the bundled weights/checkpoint-399.pth.

Note. The cRT classifier checkpoint at weights/official_crt/final_model_checkpoint.pth is not bundled either. The historical 59.688% submission used finetune-only inference (no cRT). The cRT pipeline code is included for parity with the pose90k bundle; run scripts/crt.sh (§3.4) to train your own cRT classifier on top of the bundled weights/checkpoint-99.pth, then run scripts/test.sh (§3.5) to produce a cRT-augmented submission CSV.

🏋️‍♂️ 3. Training & Testing

3.1 Pre-trained Weights

Stage	Source data	Epochs	File	Size	Test Top-1
Pre-train	MA-52 49-joint DWPose	400	`weights/checkpoint-399.pth`	119M	n/a
Fine-tune	iMiGUE-32 train+val	100	`weights/checkpoint-99.pth`	368M	59.688%
cRT	frozen fine-tuned features + class-aware sampler	10	`weights/official_crt/final_model_checkpoint.pth`	(not bundled, see §3.4)	TBD

The pre-train and fine-tune weights ship inside the HuggingFace bundle; the cRT checkpoint is not bundled and must be generated locally with §3.4 if you want to use the cRT-aware test pipeline in §3.5.

The four stages below are independent. 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/ma52.npz (not bundled, see §2 note)
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

Note. The historical 59.688% submission did not use cRT. cRT was added in this bundle for parity with the pose90k pipeline; running it here will produce a cRT-augmented number that may differ from the finetune-only baseline.

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

Note that scripts/test.sh requires a cRT classifier checkpoint; if you only have the bundled weights/checkpoint-99.pth, run §3.4 first.

📊 4. Results

Model	Modality	Test Top-1
This model (Skeleton MAMP, MA-52 pretrain), finetune-only baseline	Skeleton	59.688%
Skeleton MAMP, 90k pose pretrain (see `../Skeleton_pose90k_ft`)	Skeleton	70.188%
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:

6.3 GB