|
|
--- |
|
|
language: en |
|
|
license: mit |
|
|
tags: |
|
|
- text-classification |
|
|
- multi-label |
|
|
- emotion-detection |
|
|
- huggingface |
|
|
--- |
|
|
|
|
|
# emo-detector |
|
|
|
|
|
**emo-detector** is a multi-label emotion detection model for text. It predicts one or more emotions from the following labels: |
|
|
|
|
|
- anger |
|
|
- fear |
|
|
- joy |
|
|
- sadness |
|
|
- surprise |
|
|
|
|
|
## Model Details |
|
|
|
|
|
- **Architecture:** Pretrained DeBERTa + custom FFNN classifier |
|
|
- **Task:** Multi-label text classification |
|
|
- **Tokenizer:** DeBERTa tokenizer (`microsoft/deberta-v3-base`) |
|
|
- **Output:** Probabilities → Thresholded to 0/1 |
|
|
|
|
|
## Custom Model Class |
|
|
|
|
|
This model uses a custom architecture defined inside the `emo_detector/` module: |
|
|
|
|
|
- `emo_detector/configuration_bert_ffnn.py` → `BertFFNNConfig` |
|
|
- `emo_detector/modeling_bert_ffnn.py` → `BERT_FFNN` |
|
|
|
|
|
To load or fine-tune this model, you must download the full repository (including the `emo_detector/` folder). |
|
|
The recommended way is to use `snapshot_download()` from Hugging Face Hub. |
|
|
|
|
|
## Installation |
|
|
```bash |
|
|
pip install torch transformers huggingface_hub |
|
|
``` |
|
|
|
|
|
## Usage |
|
|
|
|
|
```python |
|
|
import sys |
|
|
import torch |
|
|
from transformers import AutoTokenizer |
|
|
from huggingface_hub import snapshot_download |
|
|
|
|
|
# Download entire repository |
|
|
repo_dir = snapshot_download("NeuralNest05/emo-detector") |
|
|
sys.path.append(repo_dir) |
|
|
|
|
|
# Import custom architecture + config |
|
|
from emo_detector.configuration_bert_ffnn import BertFFNNConfig |
|
|
from emo_detector.modeling_bert_ffnn import BERT_FFNN |
|
|
|
|
|
DEVICE = "cuda" if torch.cuda.is_available() else "cpu" |
|
|
|
|
|
# Load tokenizer |
|
|
tokenizer = AutoTokenizer.from_pretrained("NeuralNest05/emo-detector") |
|
|
|
|
|
# Load model config and architecture |
|
|
config = BertFFNNConfig.from_pretrained("NeuralNest05/emo-detector") |
|
|
model = BERT_FFNN(config) |
|
|
|
|
|
# Load weights |
|
|
model.load_state_dict(torch.load(f"{repo_dir}/pytorch_model.bin", map_location=DEVICE)) |
|
|
model.to(DEVICE) |
|
|
model.eval() |
|
|
|
|
|
# Example prediction |
|
|
texts = ["I am very happy today!", "This is scary..."] |
|
|
encodings = tokenizer(texts, truncation=True, padding=True, return_tensors="pt").to(DEVICE) |
|
|
|
|
|
with torch.no_grad(): |
|
|
logits = model(**encodings) |
|
|
probs = torch.sigmoid(logits) |
|
|
threshold = 0.5 |
|
|
preds = (probs > threshold).int() |
|
|
|
|
|
print(preds) |
|
|
``` |
|
|
|
|
|
## Output Format |
|
|
Each prediction corresponds to the five emotion labels in this order: |
|
|
``` |
|
|
["anger", "fear", "joy", "sadness", "surprise"] |
|
|
``` |
|
|
Output is a multi-hot vector, e.g.: |
|
|
``` |
|
|
[0, 0, 1, 0, 0] → joy |
|
|
``` |
|
|
|
|
|
## License |
|
|
MIT License |
|
|
|
|
|
## Acknowledgements |
|
|
|
|
|
- Microsoft DeBERTa-v3 |
|
|
- Hugging Face Transformers |
|
|
- PyTorch |
