tensorflow.mdc

---
description: "TensorFlow: Keras, model training, production deployment"
globs: ["*.py", "requirements.txt", "pyproject.toml"]
alwaysApply: true
---

# TensorFlow Rules

## Model Building
- Use tf.keras.Model or Sequential for model architecture
- Define layers in __init__, forward pass in call() method
- Use Input layer to define input shape explicitly
- Implement custom layers by inheriting tf.keras.layers.Layer
- Use functional API for complex model architectures

## Data Pipeline
- Use tf.data.Dataset for efficient data loading
- Apply transformations with map(), filter(), and batch()
- Use tf.data.AUTOTUNE for optimal performance
- Implement proper data augmentation in pipeline
- Use tf.data.experimental.AUTOTUNE for num_parallel_calls

## Training Configuration
- Configure mixed precision with policy = tf.keras.mixed_precision.Policy('mixed_float16')
- Use appropriate optimizers (Adam, AdamW, SGD) with learning rate schedules
- Implement callbacks for checkpointing, early stopping, and monitoring
- Use tf.keras.utils.plot_model for architecture visualization
- Configure proper loss functions and metrics

## Memory & Performance
- Use tf.function decorator for graph compilation
- Avoid Python loops in graph mode operations
- Use tf.GradientTape for custom training loops
- Implement gradient clipping for training stability
- Use tf.distribute.Strategy for multi-GPU training

## Model Persistence
- Save models with model.save() for complete model persistence
- Use SavedModel format for production deployment
- Export models to TensorFlow Lite for mobile/edge deployment
- Implement versioning strategy for model management
- Use TensorFlow Serving for production inference

## Best Practices
- Set random seeds for reproducibility: tf.random.set_seed()
- Use tf.keras.backend.clear_session() to reset state
- Implement proper input validation and preprocessing
- Use TensorBoard for training visualization and debugging
- Configure GPU memory growth to avoid OOM errors
- Use tf.debugging assertions for runtime validation