30 Golden Rules of Deep Learning Performance

Is your GPU starving for data? Learn 30 rules to eliminate bottlenecks and slash your deep learning training times.

#1about 5 minutes

The high cost of waiting for deep learning models to train

Long training times are a major bottleneck for developers, wasting both time and hardware resources.

#2about 2 minutes

Fine-tune your existing hardware instead of buying more GPUs

Instead of simply buying more expensive hardware, you can achieve significant performance gains by optimizing your existing setup.

#3about 3 minutes

Using transfer learning to accelerate model development

Transfer learning provides a powerful baseline by fine-tuning pre-trained models for specific tasks, drastically reducing training time.

#4about 4 minutes

Diagnose GPU starvation using profiling tools

Use tools like the TensorBoard Profiler and nvidia-smi to identify when your GPU is idle and waiting for data from the CPU.

#5about 3 minutes

Prepare your data efficiently before training begins

Optimize data preparation by serializing data into moderately sized files, pre-computing transformations, and leveraging TensorFlow Datasets for high-performance pipelines.

#6about 5 minutes

Construct a high-performance input pipeline with tf.data

Use the tf.data API to build an efficient data reading pipeline by implementing prefetching, parallelization, caching, and autotuning.

#7about 3 minutes

Move data augmentation from the CPU to the GPU

Avoid CPU bottlenecks by performing data augmentation directly on the GPU using either TensorFlow's built-in functions or the NVIDIA DALI library.

#8about 5 minutes

Key optimizations for the model training loop

Speed up the training loop by enabling mixed-precision training, maximizing the batch size, and using multiples of eight to leverage specialized hardware like Tensor Cores.

#9about 2 minutes

Automatically find the optimal learning rate for faster convergence

Use a learning rate finder library to systematically identify the optimal learning rate, preventing slow convergence or overshooting the solution.

#10about 2 minutes

Compile Python code into a graph with the tf.function decorator

Gain a significant performance boost by using the @tf.function decorator to compile eager-mode TensorFlow code into an optimized computation graph.

#11about 2 minutes

Use progressive sizing and curriculum learning strategies

Accelerate training by starting with smaller image resolutions and simpler tasks, then progressively increasing complexity as the model learns.

#12about 3 minutes

Optimize your environment and scale up your hardware

Install hardware-specific binaries and leverage distributed training strategies to scale your jobs across multiple GPUs on-premise or in the cloud.

#13about 3 minutes

Learn from cost-effective and high-speed training benchmarks

Analyze benchmarks like DawnBench and MLPerf to adopt strategies for training models faster and more cost-effectively by leveraging optimized cloud resources.

#14about 3 minutes

Select efficient model architectures for fast inference

For production deployment, choose lightweight yet accurate model architectures like MobileNet, EfficientDet, or DistilBERT to ensure fast inference on end-user devices.

#15about 2 minutes

Shrink model size and improve speed with quantization

Use model quantization to convert 32-bit weights to 8-bit integers, significantly reducing the model's size and memory footprint for faster inference.

From learning to earning

Jobs that call for the skills explored in this talk.