TensorFlow

Avoid common TensorFlow mistakes — tf.function retracing, GPU memory, data pipeline bottlenecks, and gradient traps.

Security Scan

This is an instruction-only TensorFlow tips skill that is internally consistent with its description and requests no credentials or installs, so it appears to be what it claims.

Purpose & Capability

The name/description match the SKILL.md content. The only required binary is python3, which is reasonable for TensorFlow guidance; there are no unrelated env vars, binaries, or config paths requested.

Instruction Scope

SKILL.md contains best-practice notes and code snippets (tf.function, GPU memory settings, tf.data, gradients, saving). It does not instruct reading user files, secrets, or sending data to external endpoints. It mentions CUDA_VISIBLE_DEVICES as a common testing env var but does not attempt to read or exfiltrate secrets.

Install Mechanism

No install spec or code files are present; this is instruction-only, so nothing is downloaded or written to disk by the skill itself (lowest install risk).

Credentials

The skill declares no required environment variables or credentials. It references a common runtime env var (CUDA_VISIBLE_DEVICES) in examples but does not require or request any secrets or unrelated credentials.

Persistence & Privilege

always is false and the skill is user-invocable. It does not request persistent presence or system-wide configuration changes. disable-model-invocation is false (normal platform default).

Assessment

This skill is instruction-only (no downloads or code) and provides TensorFlow debugging tips — low risk. Note source/homepage are not provided, so you can't verify the author; that matters more for code that would be installed. Before following commands: ensure you run them in a controlled environment where you have permission (e.g., your development VM or container), verify TensorFlow version compatibility, and cross-check the snippets with official TensorFlow docs. Do not provide any credentials to this skill (none are requested). If you want even lower risk, manually copy the relevant advice rather than allowing an agent to act autonomously on your environment.

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