Merge branch 'main' of https://huggingface.co/datasets/rocket-wave/hf-video-scoring into main

SPEED_OPTIMIZATIONS.md

@@ -1,110 +0,0 @@
-# NFL Play Detection Speed Optimizations
-
-## 🚀 Current Performance (Medium Model - Optimal Quality)
-
-### Speed Test Results:
-- **Average per clip**: 16.67s 
-- **Throughput**: 3.6 clips/minute
-- **Video classification**: 2.22s
-- **Audio transcription**: 13.09s
-- **Play state analysis**: <0.001s
-
-### Processing Time Estimates:
-- **10 clips**: ~2.8 minutes
-- **50 clips**: ~13.9 minutes  
-- **100 clips**: ~27.8 minutes
-- **1 hour of footage (1800 clips)**: ~8.3 hours
-
-## 🎯 Optimizations Implemented
-
-### 1. **Model Selection**
-- **Whisper-Medium**: 3GB model for optimal NFL broadcast transcription
-- **Trade-off**: Slower than base model but essential for complex broadcast audio
-- **Recommendation**: Medium model is required for accurate NFL commentary transcription
-
-### 2. **Speed-Optimized Parameters**
-```python
-# Audio transcription optimizations
-generate_kwargs={
-    "language": "en",           # Skip language detection
-    "task": "transcribe",       # No translation
-    "temperature": 0.0,         # Deterministic (fastest)
-    "do_sample": False,         # Greedy decoding
-    "num_beams": 1,             # Single beam search
-}
-```
-
-### 3. **Video Processing Optimizations**
-- Disabled debug logging for production speed
-- Optimized preprocessing pipeline
-- Efficient tensor operations
-
-### 4. **Enhanced Play Detection**
-- **Play State Classification**: Categorizes clips as `play_active`, `play_action`, `non_play`, or `unknown`
-- **Sequence Analysis**: Detects play boundaries across multiple clips
-- **NFL-Specific Logic**: Custom weights for football-related actions
-
-## 📊 Play Detection Accuracy Improvements
-
-### New Play State Categories:
-1. **`play_active`**: Active football plays (passing, kicking)
-2. **`play_action`**: Football-related actions (catching, throwing) 
-3. **`non_play`**: Non-game activities (applauding, marching)
-4. **`unknown`**: Low confidence or unclear state
-
-### Sequence Boundary Detection:
-- **Play Start**: Transition from `non_play` → `play_active/action`
-- **Play End**: Transition from `play_active/action` → `non_play`
-- **Confidence Thresholds**: Configurable confidence levels
-
-## 🔧 Further Optimization Options
-
-### For Even Faster Processing:
-1. **Disable Audio Transcription**: ~7x faster (video only, ~2.2s per clip)
-2. **Use Whisper-Base**: ~4x faster but much lower accuracy for broadcasts
-3. **GPU Acceleration**: Use CUDA if available
-4. **Model Quantization**: INT8 quantized models
-5. **Batch Processing**: Process multiple clips simultaneously
-
-### For Better Accuracy:
-1. **Whisper-Large**: Highest accuracy but even slower
-2. **Custom Fine-tuning**: Train on NFL-specific data
-3. **Ensemble Methods**: Combine multiple models
-4. **Temporal Context**: Use longer sequences
-
-## 🎬 Usage Examples
-
-### Speed Testing:
-```bash
-python speed_test.py
-```
-
-### Fast Batch Processing:
-```bash
-python run_all_clips.py
-```
-
-### Single Clip with Play Analysis:
-```bash
-python inference.py data/segment_001.mov
-```
-
-## 📈 Performance Tuning Tips
-
-1. **For Real-time**: Consider video-only processing (~2.2s per clip)
-2. **For Batch Processing**: Current 16.67s/clip provides excellent transcription quality
-3. **For Large Datasets**: Consider parallel processing or overnight batch jobs
-4. **Memory Usage**: ~6GB RAM with Whisper-Medium
-
-## 🏈 Play Detection Features
-
-### Intelligent Football Analysis:
-- Recognizes 80+ NFL terms (teams, positions, plays)
-- Corrects common speech recognition errors
-- Provides confidence scores for decisions
-- Generates comprehensive play analysis reports
-
-### Output Files:
-- `classification.json`: Video classification results
-- `transcripts.json`: Audio transcription results  
-- `play_analysis.json`: Play state and boundary analysis