Developer Blog

AI / Deep Learning | Autonomous Machines | IVA/IoT |

Developer Blog: Training with Custom Pretrained Models Using the NVIDIA Transfer Learning Toolkit

Supervised training of deep neural networks is now a common method of creating AI applications. To achieve accurate AI for your application, you generally need a very large dataset especially if you create from scratch. Gathering and preparing a large dataset and labeling all the images is expensive, time-consuming, and often requires domain expertise.

To enable faster and accurate AI training, NVIDIA just released highly accurate, purpose-built, pretrained models with the NVIDIA Transfer Learning Toolkit (TLT) 2.0. You can use these custom models as the starting point to train with a smaller dataset and reduce training time significantly. These purpose-built AI models can either be used as-is, if the classes of objects match your requirements and the accuracy on your dataset is adequate, or easily adapted to similar domains or use cases.

The TLT is a Python-based AI toolkit for creating highly optimized and accurate AI apps using transfer learning and pretrained models. The TLT makes AI accessible to everyone: data scientists, researchers, new system developers, and software engineers who are just getting started with AI. Along with creating accurate AI models, the TLT is also capable of optimizing models for inference to achieve the highest throughput for deployment.

This post walks you through the workflow, from downloading the TLT Docker container and AI models from NVIDIA NGC, to training and validating with your own dataset and then exporting the trained model for deployment on the edge using NVIDIA DeepStream SDK and NVIDIA TensorRT. Alternatively, these models can be exported and converted to a TensorRT engine for deployment. We describe each of the models, four detection and two classification models.

In addition to the purpose-built models, TLT 2.0 supports training on some of the most popular object detection architectures, such as YOLOv3, FasterRCNN, SSD/DSSD, and RetinaNet, as well as popular classification networks such as ResNet, DarkNet, and MobileNet.

Read the full blog, Training with Custom Pretrained Models Using the NVIDIA Transfer Learning Toolkit, on the NVIDIA Developer Blog.