README.md

---
base_model: bert-base-cased
datasets:
- ma2za/many_emotions
license: apache-2.0
tags:
- onnx
- emotion-detection
- BaseLM:bert-base-cased
---

# BERT-Based Emotion Detection on ma2za/many_emotions

This repository hosts a fine-tuned emotion detection model built on [BERT-base-cased](https://huggingface.co/bert-base-cased). The model is trained on the [ma2za/many_emotions](https://huggingface.co/datasets/ma2za/many_emotions) dataset to classify text into one of seven emotion categories: anger, fear, joy, love, sadness, surprise, and neutral. The model is available in both PyTorch and ONNX formats for efficient deployment.

## Model Details

### Model Description

- **Developed by:** *Your Name or Organization*
- **Model Type:** Sequence Classification (Emotion Detection)
- **Base Model:** bert-base-cased
- **Dataset:** ma2za/many_emotions
- **Export Format:** ONNX (for deployment)
- **License:** Apache-2.0
- **Tags:** onnx, emotion-detection, BERT, sequence-classification

This model was fine-tuned on the ma2za/many_emotions dataset, where the text is classified into emotion categories based on the content. For quick experimentation, a subset of the training data was used; however, the full model has been trained with the complete dataset and is now publicly available.

## Training Details

### Dataset Details
- **Dataset ID:** ma2za/many_emotions
- **Text Column:** `text`
- **Label Column:** `label`

### Training Hyperparameters
- **Epochs:** 1 (for quick test; adjust to your needs)
- **Per Device Batch Size:** 96
- **Learning Rate:** 1e-5
- **Weight Decay:** 0.01
- **Optimizer:** AdamW
- **Training Duration:** The full training run on the complete dataset (approximately 2.44 million training examples) was completed in about 3 hours and 40 minutes.



## ONNX Export

The model has been exported to the ONNX format using opset version 14, ensuring support for modern operators such as `scaled_dot_product_attention`. This enables flexible deployment scenarios across different platforms using ONNX Runtime.

## How to Load the Model

Instead of loading the model from a local directory, you can load it directly from the Hugging Face Hub using the repository name `iimran/EmotionDetection`.

### Loading with Transformers (PyTorch)

```python
import os
import numpy as np
import onnxruntime as ort
from transformers import AutoTokenizer, AutoConfig
from huggingface_hub import hf_hub_download

# Specify the repository details.
repo_id = "iimran/EmotionDetection"
filename = "model.onnx"

# Download the ONNX model file from the Hub.
onnx_model_path = hf_hub_download(repo_id=repo_id, filename=filename)
print("Model downloaded to:", onnx_model_path)

# Load the tokenizer and configuration from the repository.
tokenizer = AutoTokenizer.from_pretrained(repo_id)
config = AutoConfig.from_pretrained(repo_id)

# Check whether the configuration contains an id2label mapping.
if hasattr(config, "id2label") and config.id2label and len(config.id2label) > 0:
    id2label = config.id2label
else:
    # Default mapping for ma2za/many_emotions if not present in the config.
    id2label = {
        0: "anger",
        1: "fear",
        2: "joy",
        3: "love",
        4: "sadness",
        5: "surprise",
        6: "neutral"
    }
print("id2label mapping:", id2label)

# Create an ONNX Runtime inference session using the local model file.
session = ort.InferenceSession(onnx_model_path)

def onnx_infer(text):
    """
    Perform inference on the input text using the exported ONNX model.
    Returns the predicted emotion label.
    """
    # Tokenize the input text with a fixed maximum sequence length matching the model export.
    inputs = tokenizer(
        text,
        return_tensors="np",
        truncation=True,
        padding="max_length",
        max_length=256
    )
    
    # Prepare the model inputs.
    ort_inputs = {
        "input_ids": inputs["input_ids"],
        "attention_mask": inputs["attention_mask"]
    }
    
    # Run the model.
    outputs = session.run(None, ort_inputs)
    logits = outputs[0]
    
    # Get the predicted class id.
    predicted_class_id = int(np.argmax(logits, axis=-1)[0])
    
    # Map the predicted class id to its emotion label.
    predicted_label = id2label.get(str(predicted_class_id), id2label.get(predicted_class_id, str(predicted_class_id)))
    
    print("Predicted Emotion ID:", predicted_class_id)
    print("Predicted Emotion:", predicted_label)
    return predicted_label

# Test the inference function.
onnx_infer("That rude customer made me furious.")
```
## Evaluation
The model is primarily evaluated using the accuracy metric during training. For deployment, further evaluation on unseen data is recommended to ensure robustness in production settings.