File size: 3,524 Bytes
74a5c57 1dd383d 74a5c57 1dd383d 74a5c57 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 | ---
license: cc-by-nc-sa-4.0
---
# **DriftMoE** – A Mixture of Experts Approach to Handle Concept Drifts
Model weights for paper [DriftMoE](https://www.arxiv.org/abs/2507.18464)This repository hosts *weights only* so you can plug the model straight into your Python pipeline. These weights correspond to one training run on the LED_g stream.
Full training code & utilities live in a separate GitHub repo: **[https://github.com/miguel-ceadar/drift-moe](https://github.com/miguel-ceadar/drift-moe)**
---
## 📂 Files
We have two folders, one for the MoE-Task variant and other for the MoE-Data variant, both share this file structure:
```
router.pth # PyTorch state_dict for the gating MLP
expert_0.pkl # CapyMOA HoeffdingTree for each expert(pickled)
expert_1.pkl
…
expert_{N‑1}.pkl
```
---
## ⚡ Quick Start (CPU or GPU)
### 1 · Install runtime deps
You need to install Java and have a working java Runtime Environment to run Capymoa:
[https://openjdk.org/install/](https://openjdk.org/install/)
```bash
python -m pip install torch capymoa numpy river
# git clone training repo – needed so we can recreate RouterMLP & Expert wrappers
git clone https://github.com/miguel-ceadar/drift-moe drift_moe
```
### 2 · Load the router & experts
```python
import torch, pickle, numpy as np
from capymoa.misc import load_model
from drift_moe.driftmoe.moe_model import RouterMLP, Expert
INPUT_DIM = 24 # ↩ must match dimensions of LED_g stream
NUM_CLASSES = 10 # ↩ idem
N_EXPERTS = 12 # ↩ number of expert_*.pkl files (12 for MoE_Data and 10 for MoE_Task)
DEVICE = 'cpu' # or 'cuda'
# 2‑a) Router
router = RouterMLP(input_dim=INPUT_DIM, hidden_dim=256, output_dim=N_EXPERTS)
router.load_state_dict(torch.load('path/to/router.pth', map_location=DEVICE))
router = router.to(DEVICE).eval()
# 2‑b) Experts (pickled CapyMOA trees)
experts = []
for i in range(N_EXPERTS):
with open(f'path/to/expert_{i}.pkl', 'rb') as f:
ex = load_model(f) # HoeffdingTree object
experts.append(ex)
```
*Reference: the official CapyMOA [save & load notebook](https://capymoa.org/notebooks/save_and_load_model.html).*
### 3 · Single‑sample inference helper
```python
def predict_one(instance) -> int:
"""Route a single feature vector through driftMoE and return class index."""
x_vec = instance.x
x_t = torch.tensor(x_vec, dtype=torch.float32).unsqueeze(0).to(DEVICE)
with torch.no_grad():
logits = router(x_t) # shape [1, N_EXPERTS]
eid = int(torch.argmax(logits, 1).item())
return experts[eid].predict(instance)
```
---
## 🚰 Streaming usage
```python
from capymoa.stream.generator import LEDGenerator
stream = LEDGenerator()
while stream.has_more_instances():
inst = stream.next_instance()
y_hat = predict_one(inst)
print(y_hat)
print(inst.y_index)
```
The experts are frozen; only the router runs every forward pass.
## ✏️ Citation
```bibtex
@misc{aspis2025driftmoemixtureexpertsapproach,
title={DriftMoE: A Mixture of Experts Approach to Handle Concept Drifts},
author={Miguel Aspis and Sebastián A. Cajas Ordónez and Andrés L. Suárez-Cetrulo and Ricardo Simón Carbajo},
year={2025},
eprint={2507.18464},
archivePrefix={arXiv},
primaryClass={stat.ML},
url={https://arxiv.org/abs/2507.18464},
}
```
*Questions or issues?* Open an issue on the GitHub repo and we’ll be happy to help.
|