ODNus/timesfm-omnion-lora

TimesFM 2.5 + OMNION LoRA Adapter

OMNION 해양 환경 모니터링 데이터로 LoRA 파인튜닝한 TimesFM 2.5 어댑터.

Model Description

• Base Model: google/timesfm-2.5-200m-transformers
• Architecture: TimesFM (Time Series Foundation Model) by Google Research — 200M params
• Fine-tuning Method: LoRA (Low-Rank Adaptation) — Rank 8, Alpha 32
• LoRA Size: ~6.6 MB (adapter weights only)
• Domains: Temperature, Dissolved Oxygen (DO), pH, Salinity

Training Data

14 OMNION ocean monitoring devices deployed across South Korean coastal waters (Jeollanam-do region):

Device ID	Location
jcia01~07	Shinan, Jindo, Wando, Goheung (Jeollanam-do)
mecs01~02	Aquafarm monitoring stations
v2test02~05	Test/development units
kunsan01	Gunsan (West coast)

• Data duration: 30 days of 30-minute interval readings
• Parameters: temperature (°C), DO (mg/L), pH, salinity (g/Kg)
• Total samples: sliding window augmentation (context=128, horizon=32, stride=16)

Hyperparameters

• LoRA Rank (r): 8
• LoRA Alpha: 32
• LoRA Dropout: 0.1
• Target Modules: q_proj, v_proj, k_proj, o_proj
• Epochs: 3
• Learning Rate: 5e-5
• Batch Size: 4
• Optimizer: AdamW
• Precision: bfloat16 (MPS) / float32 (CPU)

Usage

from transformers import TimesFm2_5ModelForPrediction
from peft import PeftModel

# Load base model
base = TimesFm2_5ModelForPrediction.from_pretrained(
    "google/timesfm-2.5-200m-transformers"
)

# Load LoRA adapter
model = PeftModel.from_pretrained(base, "ODNus/timesfm-omnion-lora/lora-omnion")

# Inference
import torch
import numpy as np

# Prepare input (context of 128 time steps)
context = np.random.randn(128).astype(np.float32)
inputs = torch.tensor(context).unsqueeze(0)

with torch.no_grad():
    output = model(past_values=inputs)

predictions = output.mean_predictions[0].numpy()
print(f"Forecast shape: {predictions.shape}")

Inference Pipeline (in OMNION System)

Supabase Data → TimesFM 2.5 → LoRA Adapter → Bias Correction → Forecast Result

The LoRA adapter is automatically loaded by the ODNMON MCP server at startup. If present, it's applied silently; if missing, the base model is used as fallback.

Bias Correction

A per-device, per-parameter bias correction layer is applied post-inference:

• Analyzes recent 7-day prediction residuals
• Computes systematic offset for each device×parameter pair
• Cached across restarts for consistency

WebUI

The OMNION WebUI at omnion.odnus.com provides an interactive prediction interface where users can:

1. Select a device and parameter
2. Choose forecast horizon (6h–48h)
3. View actual vs predicted values with confidence intervals
4. Explore model architecture documentation

Contact Us

1. odn_promo@odn.us
2. https://odn.us

License

Apache 2.0