TensorFlow

What This Skill Does

The TensorFlow skill for OpenClaw acts as an expert-level debugger and architect for TensorFlow-based machine learning projects. It provides deep technical insights into the most common pitfalls of the TensorFlow ecosystem, including computational graph optimization, GPU memory management, pipeline efficiency, and gradient tape nuances. This skill helps developers resolve performance bottlenecks and runtime errors before they reach production.

Installation

To integrate this skill into your environment, run the following command in your terminal: clawhub install openclaw/skills/skills/ivangdavila/tensorflow

Use Cases

• Graph Optimization: Troubleshooting tf.function retracing to ensure your model runs at maximum compiled speed rather than falling back to eager execution.
• Performance Tuning: Identifying data pipeline stalls by implementing tf.data.AUTOTUNE and optimizing prefetching patterns.
• Memory Management: Resolving Out-Of-Memory (OOM) errors by configuring GPU memory growth correctly and applying gradient checkpointing for large architectures.
• Gradient Debugging: Tracing disconnected gradient flows and implementing custom backward passes using tf.custom_gradient.
• Production Readiness: Navigating the differences between SavedModel formats and H5 to ensure model portability and serving integrity.

Example Prompts

1. "I'm getting OOM errors when training my Transformer model on a single GPU. How do I configure TensorFlow to handle memory growth and optimize my batch size?"
2. "My model is triggering constant retracing warnings in tf.function. Can you review my code and show me how to use input_signature to fix this?"
3. "My training loop seems slow, and I suspect the data pipeline is starving the GPU. How can I optimize my tf.data.Dataset using prefetch and parallel mapping?"

Tips & Limitations

• Trace Awareness: Always remember that tf.function executes Python side effects only once during the tracing phase. Avoid using non-TensorFlow operations (like standard Python logging or print statements) inside decorated functions.
• Hardware Configuration: Ensure set_memory_growth is invoked at the very start of your program, as it cannot be modified after the GPU has been initialized.
• Broadcasting: While broadcasting is powerful, be cautious of silent shape errors; utilize tf.debugging.assert_shapes() liberally in your custom training loops to catch dimension mismatches early.
• Versioning: This skill is optimized for TensorFlow 2.x paradigms. While it addresses foundational concepts, always verify specific API changes against the latest stable documentation.