October 2018 Summaries

The benchmarking code used by Lambda Labs has been open-sourced, allowing anyone to reproduce the benchmarks or run their own. The results are published on this official page for the Lambda Community Benchmarks, and contributors can email their results to `[email protected]` or tweet @LambdaAPI. The code includes a template to paste in and run, which outputs a table with specifications such as CPU, GPU, Tensorflow version, NVIDIA driver, CUDA, cuDNN, and Python versions. There are also tables showing the benchmark results for different models (resnet-50, resnet-152, inception-v3, vgg-vd-19, alexnet, and ssd-300) with input sizes, parameter memory, feature memory, and FLOPS values. The benchmarking is performed on various configurations of GPU, including NVIDIA GeForce GTX 1080 Ti and RTX 2080 Ti, and the results are compared across different models and configurations.

The NVIDIA RTX 2080 Ti is the best GPU for deep learning, offering 37% faster FP32 performance and 62% faster FP16 performance compared to the GTX 1080 Ti. The V100 GPU comes in second, with 80% of the RTX 2080 Ti's FP32 performance and 82% of its FP16 performance, at a significantly lower cost. However, the V100 is not ideal for FP64 compute or extremely large models that require more memory. The RTX 2080 Ti is like a Porsche 911, fast but expensive, while the V100 is like a Bugatti Veyron, incredibly fast but also very expensive. The study used synthetic data and trained multiple models on each GPU to evaluate their performance, with benchmarks conducted using a Lambda Vector workstation with swapped GPUs. The results show that the RTX 2080 Ti is the best choice for deep learning, offering the best balance of performance and cost.

The problem of image segmentation is a challenging task that involves assigning each pixel in an image a class label, taking into account the context and interactions between adjacent pixels. The state-of-the-art model for this task is conditional random fields (CRFs), which uses unary and pairwise terms to penalize misclassifications and incoherent labeling. However, convolutional neural networks (CNNs) have recently shown promising results in image segmentation due to their Markovian nature, allowing them to learn important features and local interactions between these features. The U-Net architecture is a variant of the fully convolutional network that uses deconvolution to up-sample the output, which allows for more accurate object localization. Deconvolution can be used to avoid the checkerboard artifact that occurs when using filter sizes not divisible by the stride. The U-Net architecture has been widely studied and has gained success in applications beyond image segmentation, including image-to-image translation tasks.