tensorflow（tensorflow 开发）

tf.function Retracing

• New input shape/dtype causes retrace — expensive, prints warning
• Use input_signature for fixed shapes — @tf.function(input_signature=[tf.TensorSpec(...)])
• Python values retrace — pass as tensors, not Python ints/floats
• Avoid Python side effects in tf.function — only runs once during tracing

GPU Memory

• TensorFlow grabs all GPU memory by default — set memory_growth=True before any ops
• tf.config.experimental.set_memory_growth(gpu, True) — must be called before GPU init
• OOM with large models — reduce batch size or use gradient checkpointing
• CUDA_VISIBLE_DEVICES="" to force CPU — for testing without GPU

Data Pipeline

• tf.data.Dataset without .prefetch() — CPU/GPU idle time between batches
• .cache() after expensive ops — but before random augmentation
• .batch() before .map() for vectorized ops — faster than per-element
• num_parallel_calls=tf.data.AUTOTUNE — parallel preprocessing
• Dataset iteration in eager mode is slow — use in tf.function or model.fit

Shape Issues

• First dimension is batch — None for variable batch size in Input layer
• model.build(input_shape) if not using Input layer — or first call errors
• Reshape errors unclear — tf.debugging.assert_shapes() for debugging
• Broadcasting silently succeeds — may hide shape bugs

Gradient Tape

• Variables watched by default — tensors need tape.watch(tensor)
• persistent=True for multiple gradients — otherwise tape consumed after first use
• tape.gradient returns None if no path — check for disconnected graph
• @tf.custom_gradient for custom backward — not all ops have gradients

Training Gotchas

• model.trainable = False after compile does nothing — set before compile
• BatchNorm behaves differently in training vs inference — training=True/False matters
• model.fit shuffles by default — shuffle=False for time series
• validation_split takes from end — shuffle data first if order matters

Saving Models

• model.save() saves everything — architecture, weights, optimizer state
• model.save_weights() only weights — need model code to restore
• SavedModel format for serving — tf.saved_model.save(model, path)
• H5 format limited — doesn't save custom objects well, use SavedModel

Common Mistakes

• Mixing Keras and raw tf ops incorrectly — use layers.Lambda to wrap tf ops in Sequential
• tf.print vs Python print — Python print only runs at trace time in tf.function
• NumPy ops in graph — use tf ops, numpy executes eagerly only
• Loss returns scalar per sample — Keras averages, custom loops may need tf.reduce_mean