14.9K ·GitHub
What is this?
A collection of NVIDIA-maintained, state-of-the-art deep learning reference implementations across computer vision, NLP, recommenders, speech, and more, tuned for reproducible accuracy and performance on NVIDIA Tensor Core GPUs.
A collection of NVIDIA-maintained, state-of-the-art deep learning reference implementations across computer vision, NLP, recommenders, speech, and more, tuned for reproducible accuracy and performance on NVIDIA Tensor Core GPUs.Is this one library or a bunch of separate model projects?It is an index of many independent model implementations (ResNet, BERT, DLRM, Tacotron 2, etc.), each living in its own folder with its own README, spread across PyTorch, TensorFlow, TensorFlow2, MXNet and PaddlePaddle.
It is an index of many independent model implementations (ResNet, BERT, DLRM, Tacotron 2, etc.), each living in its own folder with its own README, spread across PyTorch, TensorFlow, TensorFlow2, MXNet and PaddlePaddle.How do I actually run one of these examples?Open the README for the specific model you want and follow its instructions, or pull NVIDIA's monthly-updated Docker container from the NGC registry which bundles the examples with the full optimized software stack.
Open the README for the specific model you want and follow its instructions, or pull NVIDIA's monthly-updated Docker container from the NGC registry which bundles the examples with the full optimized software stack.Do I need a specific NVIDIA GPU to use these?Yes, the examples are built and tuned for NVIDIA Volta, Turing and Ampere GPUs to take advantage of Tensor Cores, AMP and TF32.
Yes, the examples are built and tuned for NVIDIA Volta, Turing and Ampere GPUs to take advantage of Tensor Cores, AMP and TF32.Are all models equally production-ready or supported?No, support level varies per model and feature: some have full TensorRT/Triton 'Supported' status while others only have an 'Example', and multi-node training is limited to a handful of models; each model README states its own support tier.
No, support level varies per model and feature: some have full TensorRT/Triton 'Supported' status while others only have an 'Example', and multi-node training is limited to a handful of models; each model README states its own support tier.What is the NGC container mentioned in the README?It is a monthly-updated Docker image from NVIDIA GPU Cloud that bundles these examples together with NVIDIA's upstream framework contributions and optimized libraries like cuDNN, NCCL and cuBLAS, all quality-checked for performance.
It is a monthly-updated Docker image from NVIDIA GPU Cloud that bundles these examples together with NVIDIA's upstream framework contributions and optimized libraries like cuDNN, NCCL and cuBLAS, all quality-checked for performance.What do AMP and TF32 mean for these models?AMP (Automatic Mixed Precision) and TF32 are precision modes on Volta/Turing/Ampere GPUs that speed up training (TF32 can give up to 10x over FP32 on Volta) with little to no code change and are enabled on most listed models.
AMP (Automatic Mixed Precision) and TF32 are precision modes on Volta/Turing/Ampere GPUs that speed up training (TF32 can give up to 10x over FP32 on Volta) with little to no code change and are enabled on most listed models.What problem domains do these examples cover?Computer vision (classification, detection, segmentation), NLP (BERT, GNMT, ELECTRA), recommender systems (DLRM, NCF, Wide&Deep), speech-to-text (Jasper, QuartzNet), text-to-speech (Tacotron 2, HiFi-GAN), graph neural networks, and time-series forecasting (Temporal Fusion Transformer).
Computer vision (classification, detection, segmentation), NLP (BERT, GNMT, ELECTRA), recommender systems (DLRM, NCF, Wide&Deep), speech-to-text (Jasper, QuartzNet), text-to-speech (Tacotron 2, HiFi-GAN), graph neural networks, and time-series forecasting (Temporal Fusion Transformer).Is every model in the repo kept actively up to date?Not necessarily; each model README specifies its own support tier, ranging from ongoing updates and improvements to a one-time 'point-in-time' release meant mainly for thought leadership.
Not necessarily; each model README specifies its own support tier, ranging from ongoing updates and improvements to a one-time 'point-in-time' release meant mainly for thought leadership.What are similar projects?DeepLearning, transformers.js, insightface, tensor2tensor, Transformers-Tutorials, FireRedAudio, ultralytics, dgl
DeepLearning, transformers.js, insightface, tensor2tensor, Transformers-Tutorials, FireRedAudio, ultralytics, dglWhich topics is it in?paddlepaddle, tensorflow2, mxnet, graph-neural-networks, bert, forecasting, image-classification, speech-synthesis, object-detection, onnx, tensorflow, speech-recognition
paddlepaddle, tensorflow2, mxnet, graph-neural-networks, bert, forecasting, image-classification, speech-synthesis, object-detection, onnx, tensorflow, speech-recognition