If you have not visited the following landing pages, please do so before attempting to use this repository.
Develop Embedded Friendly Deep Neural Network Models in PyTorch 
This is an extension of the popular github repository pytorch/vision that implements torchvision - PyTorch based datasets, model architectures, and common image transformations for computer vision.
The scripts in this repository requires torchvision to be installed using this repository - the standard torchvision will not support all the features in this repository. Please install our torchvision extension using the instructions below.
These installation instructions were tested using Miniconda Python 3.7 on a Linux Machine with Ubuntu 18.04 OS.
Make sure that your Python version is indeed 3.7 or higher by typing:
python --version
Make sure gcc and g++ versions are greater then or equal to 7 - if not install the required versions. The installed versions can be checked by:
gcc --version
g++ --version
Clone this repository into your local folder
Execute the following shell script to install the dependencies:
./setup.sh
This installation script is derived from the instructions described in pytorch/vision to install/build from source.
We have three categories of models and scripts in this repository:
See documentation of our edgeailite extensions to torchvision
torchvision originally had only classification models and also did not have an implementation of Quantization Aware Training (QAT). So we went ahead added embedded friendly models and training scripts for tasks such as Semantic Segmentation, Depth Estimation, Multi-Task Estimation etc and also QAT. These models, transforms and utility functions are located in ./torchvision/edgeailite and the corresponding training scrpts are in ./references/edgeailite. The models trained using these scripts will carry a keyword "edgeailite". See the list of "edgeailite" models here: ./torchvision/edgeailite/xvision/models
Scripts are provided for training low complexity DNN models for tasks such as:
- Image Classification
- Semantic Segmentation
- Motion Segmentation
- Depth Estimation
- Multi-Task Estimation
- And more...
Tools and scripts for Quantization Aware Training (QAT) that is best suited for our devices are also provided.
The shell scripts run_edgeailite_....sh can be used to train, evaluate or export these "edgeailite" models.
Recently torchvision has started adding support for several tasks such as Object Detection, Semantic Segmentation etc. Some of these do not run on our platform due to the presence of unsupported layers. But we have a model surgery function that creates embedded friendly versions of these models. The models thus created carry a keyworkd "lite". See the list of "lite" models here: ./torchvision/models/model_lite.py, ./torchvision/models/detection/model_lite.py, ./torchvision/models/segmentation/model_lite.py
These "lite" models (1) provide more variety to our Model Zoo making it richer (2) extensible as torchvision adds more models in the future (3) stay close to the official version of these models.
The shell scripts run_torchvision_....sh can be used to train, evaluate or export these "lite" models.
Note: these models are still in developmet and some of them may not be fully supported in TIDL yet.
See the Original TorchVision documentation
While some models in this category work on our platform, several of them have unsupported layers and will either not work or will be slow. We recommend to use models from one of the above two categories for inference on our embedded SoC platforms.