Adya662/bert-tiny-amd

1#!/usr/bin/env python32"""3Simple script to upload model files to Hugging Face Hub4"""5 6import os7import torch8from transformers import AutoTokenizer9from huggingface_hub import HfApi10import json11from pathlib import Path12 13# Configuration14REPO_ID = "Adya662/bert-tiny-amd"15MODEL_PATH = "best_enhanced_progressive_amd.pth"16BASE_MODEL = "prajjwal1/bert-tiny"17 18def create_model_config():19    """Create model configuration"""20    config = {21        "model_type": "bert",22        "architectures": ["BertForSequenceClassification"],23        "attention_proxy_dtype": "float32",24        "attention_dropout": 0.1,25        "classifier_dropout": None,26        "hidden_act": "gelu",27        "hidden_dropout_prob": 0.1,28        "hidden_size": 128,29        "initializer_range": 0.02,30        "intermediate_size": 512,31        "layer_norm_eps": 1e-12,32        "max_position_embeddings": 512,33        "model_type": "bert",34        "num_attention_heads": 2,35        "num_hidden_layers": 2,36        "num_labels": 1,37        "pad_token_id": 0,38        "position_embedding_type": "absolute",39        "problem_type": "single_label_classification",40        "torch_dtype": "float32",41        "transformers_version": "4.21.0",42        "type_vocab_size": 2,43        "use_cache": True,44        "vocab_size": 3052245    }46    return config47 48def create_training_metadata():49    """Create training metadata"""50    metadata = {51        "model_name": "bert-tiny-amd",52        "base_model": "prajjwal1/bert-tiny",53        "task": "text-classification",54        "dataset": "ElevateNow call center transcripts",55        "language": "en",56        "license": "mit",57        "pipeline_tag": "text-classification",58        "tags": [59            "text-classification",60            "answering-machine-detection",61            "bert-tiny",62            "binary-classification",63            "call-center",64            "voice-processing"65        ],66        "performance": {67            "validation_accuracy": 0.9394,68            "precision": 0.9275,69            "recall": 0.8727,70            "f1_score": 0.899371        },72        "training_details": {73            "total_samples": 3548,74            "training_samples": 2838,75            "validation_samples": 710,76            "epochs": 15,77            "batch_size": 32,78            "learning_rate": 3e-5,79            "device": "mps"80        }81    }82    return metadata83 84def upload_files():85    """Upload files to Hugging Face Hub"""86    87    print("🚀 Starting file upload to Hugging Face Hub...")88    89    # Initialize HF API90    api = HfApi()91    92    # Create model configuration93    config = create_model_config()94    95    # Save config96    with open("config.json", "w") as f:97        json.dump(config, f, indent=2)98    99    # Create training metadata100    metadata = create_training_metadata()101    102    # Save training metadata103    with open("training_metadata.json", "w") as f:104        json.dump(metadata, f, indent=2)105    106    # Load and save tokenizer from base model107    print("📥 Loading tokenizer...")108    tokenizer = AutoTokenizer.from_pretrained(BASE_MODEL)109    tokenizer.save_pretrained(".")110    111    # Copy model weights112    if os.path.exists(MODEL_PATH):113        print("📥 Copying model weights...")114        import shutil115        shutil.copy2(MODEL_PATH, "pytorch_model.bin")116        print("✅ Model weights copied successfully")117    else:118        print(f"❌ Model file {MODEL_PATH} not found!")119        return False120    121    # Create README.md122    readme_content = """---123license: mit124tags:125- text-classification126- answering-machine-detection127- bert-tiny128- binary-classification129- call-center130- voice-processing131pipeline_tag: text-classification132---133 134# BERT-Tiny AMD Classifier135 136A lightweight BERT-Tiny model fine-tuned for Answering Machine Detection (AMD) in call center environments.137 138## Model Description139 140This model is based on `prajjwal1/bert-tiny` and fine-tuned to classify phone call transcripts as either human or machine (answering machine/voicemail) responses. It's designed for real-time call center applications where quick and accurate detection of answering machines is crucial.141 142## Model Architecture143 144- **Base Model**: `prajjwal1/bert-tiny` (2 layers, 128 hidden size, 2 attention heads)145- **Total Parameters**: ~4.4M (lightweight and efficient)146- **Input**: User transcript text (max 128 tokens)147- **Output**: Single logit with sigmoid activation for binary classification148- **Loss Function**: BCEWithLogitsLoss with positive weight for class imbalance149 150## Performance151 152- **Validation Accuracy**: 93.94%153- **Precision**: 92.75%154- **Recall**: 87.27%155- **F1-Score**: 89.93%156- **Training Device**: MPS (Apple Silicon GPU)157- **Best Epoch**: 15 (with early stopping)158 159## Training Data160 161- **Total Samples**: 3,548 phone call transcripts162- **Training Set**: 2,838 samples163- **Validation Set**: 710 samples164- **Class Distribution**: 30.8% machine calls, 69.2% human calls165- **Source**: ElevateNow call center data166 167## Usage168 169### Basic Inference170 171```python172from transformers import AutoTokenizer, AutoModelForSequenceClassification173import torch174 175# Load model and tokenizer176model = AutoModelForSequenceClassification.from_pretrained("Adya662/bert-tiny-amd")177tokenizer = AutoTokenizer.from_pretrained("Adya662/bert-tiny-amd")178 179# Prepare input180text = "Hello, this is John speaking"181inputs = tokenizer(text, return_tensors="pt", max_length=128, truncation=True, padding=True)182 183# Make prediction184with torch.no_grad():185    outputs = model(**inputs)186    logits = outputs.logits.squeeze(-1)187    probability = torch.sigmoid(logits).item()188    is_machine = probability >= 0.5189 190print(f"Prediction: {'Machine' if is_machine else 'Human'}")191print(f"Confidence: {probability:.4f}")192```193 194## Training Details195 196- **Optimizer**: AdamW with weight decay (0.01)197- **Learning Rate**: 3e-5 with linear scheduling198- **Batch Size**: 32199- **Epochs**: 15 (with early stopping)200- **Early Stopping**: Patience of 3 epochs201- **Class Imbalance**: Handled with positive weight202 203## Limitations204 205- Trained on English phone call transcripts206- May not generalize well to other languages or domains207- Performance may vary with different transcription quality208- Designed for short utterances (max 128 tokens)209 210## License211 212MIT License - see LICENSE file for details.213"""214    215    with open("README.md", "w") as f:216        f.write(readme_content)217    218    # Upload to Hub219    print("⬆️ Uploading to Hugging Face Hub...")220    try:221        api.upload_folder(222            folder_path=".",223            repo_id=REPO_ID,224            repo_type="model",225            commit_message="Upload trained BERT-Tiny AMD model"226        )227        print("✅ Model uploaded successfully!")228        print(f"🔗 Model available at: https://huggingface.co/{REPO_ID}")229        return True230    except Exception as e:231        print(f"❌ Upload failed: {e}")232        return False233 234if __name__ == "__main__":235    success = upload_files()236    if success:237        print("\n🎉 Model deployment completed successfully!")238    else:239        print("\n💥 Model deployment failed!")