NVIDIA Transfer Learning Toolkit (TLT) is a Python package to enable NVIDIA customers the ability to fine-tune pretrained models with customer’s own data and export them for TensorRT based inference through an edge device.
Features included in this release:
Python3 pip package as a unified Command Line Interface (CLI)
TLT Launcher now supports interactive sessions for debugging
Support for docker hosted from different registries
Support for overridable configs via environment variables
Jupyter notebook examples showing how to use the pretrained models effectively.
Pretrained models for several public architectures and reference applications serving computer vision related object classification, detection and segmentation use cases.
Support for YOLOv3, YOLOv4, FasterRCNN, SSD, RetinaNet and DSSD object detection models.
Support for MaskRCNN Instance segmentation model
Support for pruning MaskRCNN models
Support for UNet Semantic segmentation model
Support for pruning a UNet Semantic Segmentation model
Support for serializing a template DeepStream config and labels file
Support for training highly accurate purpose-built models:
Facial Landmarks Estimation
Instructions for running TLT in the cloud with AWS or GCP
Quantization aware training for accurate INT8 models
Support for Automatic Mixed Precision (AMP) training
Offline augmentation tool for object detection datasets
Model adaptation and retraining that is easy to use in heterogeneous multiple GPU environments.
Reference Application for deploying the following models to NVIDIA Triton:
Converter utility to generate device specific optimized TensorRT engines.
TLT uses the CUDA multi-process service which helps in optimizing GPU utilization during multiple GPU training.
TLT Conversational AI:
Pretrained models for several public architectures and reference applications serving conversational AI related speech to text and natural language processing use cases, namely
Speech to Text
Speech to Text with Citrinet
Punctuation and capitalization
Intent and slot classification
Components included in this release:
TLT Launcher pip package
TLT - TF docker
TLT - Pytorch Docker
Jupyter notebook with sample workflows
Getting Started Guide containing usage and installation instructions
tlt-converter for x86 + discrete GPU platforms
tlt-converter for Jetson (ARM64) available here.
Pre-trained weights trained on Open Image dataset available on NGC
Unpruned and Pruned models for Purpose-built models - Pruned models can be deployed out-of-box with DeepStream and unpruned models can be used for re-training.
Trainable and out-of-box Deployable models for:
Ubuntu 18.04 LTS
Docker API > 1.40
Docker-ce > 19.03
Python > 3.6.9
NVIDIA GPU driver v455.xx or above
DeepStream 5.1 for inference and deployment is recommended.
The following system configuration is recommended to achieve reasonable training performance with the TLT and supported models provided:
32 GB system RAM
32 GB of GPU RAM
8 core CPU
1 NVIDIA GPU
100 GB of SSD space
TLT is supported on A100, V100 and RTX 30x0 GPUs.
The SSD, DSSD, YOLOv3, YOLOv4 RetinaNet, FasterRCNN and MaskRCNN integration with DeepStream requires custom plugins from the TensorRT Open Source Software (OSS) library. DeepStream 5.1 doesn’t natively support custom plugins from TensorRT OSS. Instructions to build TensorRT OSS and custom parsing code to run with Deepstream have been provided here.
Transfer Learning is not supported on pruned models across all applications.
When training with multiple GPUs, you might need to scale down the batch_size and/or scale up the learning rate to get the same accuracy seen in single GPU training.
When training DetectNet_v2 for object detection use-cases with more than 10 classes, you may need to either update the cost_weight parameter in the cost_function_config, or balance the number of samples per class in the dataset for better training.
When training a DetectNet_v2 network for datasets with less than 20,000 images, please use smaller batch-sizes (1, 2 or 4) to get better accuracy.
The infer subtask of DetectNet_v2 doesn’t output confidence and generates 0. as value. You may ignore these values and only consider the bbox and class labels as valid outputs.
When using MaskRCNN, please make sure GPU 0 is free.
ResNet101 pre-trained weights from NGC is not supported on YOLOv3, YOLOv4, SSD, DSSD and RetinaNet.
UNet is currently not supported for inference in DS 5.0. DeepStream 5.1 supports it.
New backbones added to UNet training
Vanilla UNet Dynamic
UNet supports multi-source data loading
When generating int8 engine with
tlt-converter, please use
-sif there is TensorRT error message saying weights are outside of fp16 range.
.etltfiles are not directly compatible with DeepStream. Users are required to use the
tlt-converterto generate a TensorRT engine when integrating with DeepStream.
TLT Conversational AI
When running convAI models on a cloud VM, users should have root access to the VM.
ngc config set, the NGC CLI may not prompt the user to configure the team and org. In this case, users may run into an error when downloading models saying
Missing org - If apikey is set, org is also required.
Please maintain a back-up of your existing NGC API key from the ngc config at
~/.ngc/configand clear the ngc config by running the following command.
ngc config clear
Loading pretrained weights when retraining a pruned model