You can deploy this model and start earning money today!

ZenbiteXYZ/
sentistream-encoder-v4

onnx

sentistream_encoder

Views

No views yet

SentiStream Encoder V4

ML-Stack: PyTorch

Inference: ONNX

Tracking: W&B

A specialized dimensionality reduction engine designed for real-time sentiment analysis in streaming environments. It compresses high-dimensional BGE-small-en-v1.5 embeddings (384D) into a structured 5D manifold optimized for incremental clustering via DBSTREAM.

🧠 Architecture & Methodology

The project solves the "Curse of Dimensionality" in online clustering by implementing a two-stage training pipeline:

Supervised Warm-up: The encoder is pre-trained to classify sentiment polarity (binary), ensuring that the resulting latent space is semantically separated by sentiment.
Parametric UMAP (PUMAP): A neural-network-based UMAP implementation that learns a global transformation function, allowing new, unseen data to be projected into the same manifold without retraining.

📊 Performance Metrics

Input Dimensions: 384 (Compatible with BAAI/bge-small-en-v1.5)
Latent Dimensions: 5
Pre-training Accuracy: 99.92% (Amazon Reviews Polarity Subset)
Inference Latency: < 1ms per batch (on CPU via ONNX Runtime)

📂 Repository Structure

model.onnx: Production-ready weights for high-speed inference.
pytorch_model.bin: Original PyTorch state dictionary.
scaler_config.json: Mean and Scale parameters for StandardScaler (mandatory for inference).
config.json: Model hyperparameters and architectural metadata.

🚀 Quick Start

1. Requirements

pip install onnxruntime numpy

2. Inference Example (ONNX)

Python

import onnxruntime as ort
import numpy as np
import json

# Load scaler parameters
with open("scaler_config.json", "r") as f:
    config = json.load(f)
    mean, scale = np.array(config['mean']), np.array(config['scale'])

# Load ONNX session
session = ort.InferenceSession("model.onnx")

def encode(embeddings):
    # Standardize input
    x = (embeddings - mean) / scale
    # Run model
    inputs = {session.get_inputs()[0].name: x.astype(np.float32)}
    return session.run(None, inputs)[0]

🛠 Training Details

Optimizer: Adam (lr=3.57e-3)

Loss: Hybrid Cross-Entropy + UMAP Loss

Manifold Hyperparams: n_neighbors=26, min_dist=0.00366

Environment: ArchLinux, Python 3.12, CUDA 12.x

⚖ License

MIT

ZenbiteXYZ/sentistream-encoder-v4