tensorflow

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