Input: (B, L, T, J, C)
Short-term encoder: PoseFormerFactorized
- Temporal attention (per joint)
- Spatial attention (per frame)
Long-term encoder: Transformer over segment-level features

Image stream (optional)

Backbone: ResNet18 / ResNet50
Temporal pooling per segment

Fusion

concat (default): feature concatenation + MLP
xattn: shallow cross-attention (pose tokens ↔ image token)

Output

Activity classification logits
Optional intermediate embeddings / tokens

🏗️ Model Components

Module	Description
`PoseFormerFactorized`	Short-term pose transformer
`LongTermTemporalBlock`	Long-range temporal modeling
`ImageEncoder`	CNN-based RGB feature extractor
`MMFusionConcatLN`	Concatenation-based multimodal fusion
`MMFusionCrossAttnShallow`	Cross-attention multimodal fusion
`SSLHeads`	Contrastive, reconstruction, temporal order heads

📥 Input Format

Pose Input

(B, L, T, J, C)

B: batch size
L: number of temporal segments
T: frames per segment
J: number of joints (e.g. 17)
C: joint channels (2D or 3D)

Image Input (optional)

(B, L, T, 3, H, W)

3: RGB channels
H: image height
W: image width

🚀 Usage

1️⃣ Load Model Code

python

from model_har_final import (
    PoseFormerFactorized,
    MultiScaleTemporalModel
)

2️⃣ Build Model

python

pose_backbone = PoseFormerFactorized(
    joints=17,
    in_ch=3,
    dim=128,
    layers=4,
    num_classes=6,
    return_tokens=True
)

model = MultiScaleTemporalModel(
    short_seq_model=pose_backbone,
    num_classes=6,
    enable_long_term=True,
    multimodal=True,
    fusion_mode="concat"  # or "xattn"
)

3️⃣ Load Weights

python

import torch

ckpt = torch.load("best_stage3_dual_sched.pth", map_location="cpu")
model.load_state_dict(ckpt)
model.eval()

ℹ️ This model is saved using state_dict, not pickle-serialized objects, for maximum compatibility.

4️⃣ Inference

python

with torch.no_grad():
    logits = model(pose_seq, img_seq)
    preds = logits.argmax(dim=-1)

🧪 Training Strategy

The model is trained in three stages:

Stage 1 — Pose SSL + Weak Supervision

Masked joint modeling (MJM)
Contrastive learning (InfoNCE)
Temporal order prediction
Optional labeled supervision

Stage 2 — Pose-only SSL Refinement

Contrastive + reconstruction losses
Temporal attention disabled for stability

Stage 3 — Multimodal Supervised Fine-tuning

Label smoothing CE
Semantic prototype distillation
Metric learning (triplet)
Optional knowledge distillation

📊 Outputs

Output	Shape
Logits	`(B, num_classes)`
Pose embedding	`(B, D)`
Pose tokens (optional)	`(B, T, J, D)`

📎 Limitations

Not compatible with AutoModel.from_pretrained
Requires custom code to instantiate architecture
Input pose format must match training configuration

📜 License

Apache License 2.0

📌 Citation

If you use this work, please cite:

bibtex

@misc{kim2025multiscalehar,
  title        = {Multi-Scale Multimodal Pose Transformer for Human Activity Recognition},
  author       = {Minjae Kim},
  year         = {2025},
  howpublished = {\url{https://huggingface.co/m97j/har-safety-model}}
}

m97j/uni-har

Multi-Scale Multimodal Pose HAR

🧠 Model Overview

Architecture Summary

🏗️ Model Components

📥 Input Format

Pose Input

Image Input (optional)

🚀 Usage

1️⃣ Load Model Code

2️⃣ Build Model

3️⃣ Load Weights

4️⃣ Inference

🧪 Training Strategy

Stage 1 — Pose SSL + Weak Supervision

Stage 2 — Pose-only SSL Refinement

Stage 3 — Multimodal Supervised Fine-tuning

📊 Outputs

📎 Limitations

📜 License

📌 Citation

m97j/
uni-har