This week at the annual Computer Vision and Pattern Recognition (CVPR) conference, NVIDIA released Feature Map Explorer for visualizing 4D image-based feature maps, new computer vision models on NGC, and CUDA-X libraries such as cuDNN, NCCL and DALI to support the new NVIDIA Ampere architecture. NVIDIA also showcased Misty, an interactive 3D conversational AI application that uses vision and speech to naturally interact with users.
Learn more about the new features of the NVIDIA A100 GPU that make it ideal for computer vision (CV) workloads, on the NVIDIA Developer Blog, Improving Computer Vision with NVIDIA A100 GPUs.
Visit the NVIDIA CVPR page for more information on NVIDIA’s accepted research, workshops, tutorials, and featured demos.
Below are a few of the top releases:
NVIDIA Feature Map Explorer
NVIDIA FME (Feature Map Explorer) enables the visualization of 4D image-based feature maps. It is available for download on developer.nvidia.com for Windows and Linux. When feature maps are analyzed, FME produces a range of data, from low-level channel visualizations to detailed numerical information about the entire tensor and each channel slice. DNN developers can now peer into the neural “black box” to see what a model has learned and where it might benefit from changes in design or training protocol.
See today how FME can provide a rich set of data processing options to help highlight points of interest from a feature map, and make it easier to check for opportunities to improve speed by dropping to lower precision numerical formats.
NVIDIA Riva is an accelerated SDK for multimodal conversational AI services that delivers real-time performance on GPUs. This week at CVPR, NVIDIA introduced Misty, a 3D animated, intelligent, interactive chatbot, brought to life in Omniverse. Misty connects the Riva multimodal conversational AI technology to Omniverse’s state-of-the-art AI computer graphics technology.
The NVIDIA CUDA Deep Neural Network library (cuDNN) is a GPU-accelerated library of primitives for deep neural networks. This version of cuDNN includes:
- Tuned for peak performance on NVIDIA A100 GPUs including new TensorFloat-32, FP16, and FP32
- Redesigned low-level API provides direct access to cuDNN kernels for greater control and performance tuning
- Backward compatibility layer maintains support for cuDNN 7.x letting developers manage their transition to the new cuDNN 8 API
- New optimizations for computer vision, speech, and language understanding networks
- Fuse operators to accelerate convolutional neural networks with a new API
Deep Learning Frameworks and Models in NGC 20.03
NGC provides containers, models, and scripts with the latest performance enhancements. All of our containers are updated monthly with the latest CUDA libraries. This month’s updates include:
- Added UNet_Medical and UNet_Industrial Tensorflow models that speeds up training upto 2.5X
- Accelerate model training of MaskRCNN by 1.4X and ResNet-50 by 3.4X with a single line of code
- Visit GitHub for more available and upcoming models in NGC
Refer to each package’s release notes in documentation for additional information.