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.
Key Features
Features included in this release:
Pretrained models for several public architectures and reference applications serving computer vision related object classification and detection Intelligent Video Analytics (IVA) use cases.
Support for YOLOv3, FasterRCNN, SSD, RetinaNet and DSSD object detection models.
Support for MaskRCNN Instance segmentation model
Support to train highly accurate purpose-built models - PeopleNet, TrafficCamNet, DashCamNet, FaceDetectIR, VehicleTypeNet, VehicleMakeNet
Quantization aware training for accurate INT8 models
Support for Automatic Mixed Precision (AMP) training
Offline augmentation tool
Jupyter notebook examples showing how to use the pretrained models effectively.
Model adaptation and retraining that is easy to use in heterogeneous multiple GPU environments.
Pruning API that compresses the size of the model during training.
Model export API for integrating the model directly into the DeepStream environment.
Converter utility to generate a device specific optimized TensorRT engine.
TLT uses the CUDA multi-process service which helps in optimizing GPU utilization during multiple GPU training.
Contents
Components included in this release:
TLT docker
Jupyter notebook with sample workflows
Getting Started Guide containing usage and installation instructions
tlt-converter (for x86 it’s part of the package)
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.
Software Requirements
Ubuntu 18.04 LTS
NVIDIA GPU driver v410.xx or above
DeepStream 5.0 for inference and deployment is recommended.
Hardware Requirements
Minimum
4 GB system RAM
4 GB of GPU RAM
Single core CPU
1 NVIDIA GPU
50 GB of HDD space
Recommended
32 GB system RAM
32 GB of GPU RAM
8 core CPU
4 GPUs - V100
100 GB of SSD space
Known Issues
SSD, DSSD, YOLOv3, RetinaNet, FasterRCNN and MaskRCNN integration to DeepStream is a feature and requires custom plugins from the TensorRT Open Source Software (OSS) library. DeepStream 5.0 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 resuming a DetectNet_v2 training from checkpoint, please make sure to maintain the same number of GPUs and the same command line to restart the 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.
When using MaskRCNN, please make sure GPU 0 is free.
ResNet101 pre-trained weights from NGC is not supported on YOLOV3, SSD, DSSD and RetinaNet.
Resolved Issues
Loading pretrained weights when retraining a pruned model.
Incorrect image_extension in detectnet dataset_config returns unrelated error during tlt-evaluate.