TensorFlowTensorFlow常见错误

tf.function Retracing

• - New input shape/dtype causes retrace — expensive, prints warning
• Use input_signature for fixed shapes — INLINECODE1
• 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
• INLINECODE3 — must be called before GPU init
• OOM with large models — reduce batch size or use gradient checkpointing
• INLINECODE4 to force CPU — for testing without GPU

Data Pipeline

• - tf.data.Dataset without .prefetch() — CPU/GPU idle time between batches
• INLINECODE7 after expensive ops — but before random augmentation
• INLINECODE8 before .map() for vectorized ops — faster than per-element
• INLINECODE10 — 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
• INLINECODE12 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 INLINECODE14
• INLINECODE15 for multiple gradients — otherwise tape consumed after first use
• INLINECODE16 returns None if no path — check for disconnected graph
• INLINECODE17 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
• INLINECODE20 shuffles by default — shuffle=False for time series
• INLINECODE22 takes from end — shuffle data first if order matters

Saving Models

• - model.save() saves everything — architecture, weights, optimizer state
• INLINECODE24 only weights — need model code to restore
• SavedModel format for serving — INLINECODE25
• 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
• INLINECODE27 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 INLINECODE28