Neural network model compiler for Arm Ethos-U NPUs
In order to be accelerated by the Ethos-U NPU the network operators must be quantised to either 8-bit (unsigned or signed) or 16-bit (signed).
The optimised model will contain TensorFlow Lite Custom operators for those parts of the model that can be accelerated by the Ethos-U NPU. Parts of the model that cannot be accelerated are left unchanged and will instead run on the Cortex-M series CPU using an appropriate kernel (such as the Arm optimised CMSIS-NN kernels).
After compilation the optimised model can only be run on an Ethos-U NPU embedded system.
The tool will also generate performance estimates (EXPERIMENTAL) for the compiled model.
The tool has limited functionality for compiling a TOSA neural network (EXPERIMENTAL).
- Vela 3.1.0 to current supports TensorFlow 2.5
- Vela 2.1.0 to 3.0.0 supports TensorFlow 2.4
- Vela 2.0.0 to 2.0.1 supports TensorFlow 2.3
- Vela 0.1.0 to 1.2.0 supports TensorFlow 2.1
Vela runs on the Linux and Microsoft Windows 10 operating systems, see note in Installation section below.
The following should be installed prior to the installation of Vela:
- Python >= 3.6
- GNU toolchain (GCC, Binutils and libraries)
- Pipenv virtual environment tool
Note: For installing on Microsoft Windows 10 you need to have a C99 capable toolchain installed. The recommended and tested toolchain is Microsoft Visual C++ 14.2 Build Tools, see https://wiki.python.org/moin/WindowsCompilers
Install Vela from PyPi using the following command:
pip3 install ethos-u-vela
First obtain the source code by either downloading the desired TGZ file from:
Or by cloning the git repository:
git clone https://review.mlplatform.org/ml/ethos-u/ethos-u-vela.git
Once you have the source code, Vela can be installed using the following command:
pip3 install .
Or, if you use
pipenv install .
Advanced Installation for Developers
If you plan to modify the Vela codebase then it is recommended to install Vela
as an editable package to avoid the need to re-install after every modification.
This is done by adding the
-e option to the above install commands like so:
pip3 install -e .
Or, if you use
pipenv install -e .
If you plan to contribute to the Vela project (highly encouraged!) then it is recommended to install Vela along with the pre-commit tools (see Vela Testing for more details).
Vela is run with an input
.tflite file passed on the command line. This file
contains the neural network to be compiled. The tool then outputs an optimised
version with a
_vela.tflite file prefix, along with the performance estimate
(EXPERIMENTAL) CSV files, all to the output directory. It also prints a
performance estimation summary back to the console, see
Vela Performance Estimation Summary.
If you use the
pipenv virtual environment tool then first start by spawning a
shell in the virtual environment:
After which running Vela is the same regardless of whether you are in a virtual environment or not.
- Compile the network
my_model.tflite. The optimised version will be output to
- Compile the network
/path/to/my_model.tfliteand specify the output to go in the directory
vela --output-dir ./results_dir /path/to/my_model.tflite
- Compile a network using a particular Ethos-U NPU. The following command selects an Ethos-U65 NPU accelerator configured with 512 MAC units.
vela --accelerator-config ethos-u65-512 my_model.tflite
- Compile a network while minimizing peak SRAM usage, therefore prioritising a lower SRAM usage over runtime performance.
vela --optimise Size my_model.tflite
- Compile a network to have maximum performance, i.e. the fastest inference time. This prioritises a higher runtime performance over a lower peak SRAM usage.
vela --optimise Performance my_model.tflite
- Compile a network while optimising for the fastest inference time possible, with an upper bound for the SRAM usage. The memory limit is set in bytes, i.e. run the following example if one requires a limit of 300KB.
vela --optimise Performance --arena-cache-size 300000 my_model.tflite
- Compile a network using a particular embedded system configuration defined in
Vela's configuration file. The following command selects the
My_Sys_Configsystem configuration along with the
My_Mem_Modememory mode from the
vela --config vela_cfg.ini --system-config My_Sys_Config --memory-mode My_Mem_Mode my_model.tflite
- To get a list of all available options (see CLI Options section below):
When running the Vela compiler it may report a number of warning messages to the console. These should all be thoroughly reviewed as they will indicate decisions that the compiler has made in order to create the optimised network.
Some example networks that contain quantised operators which can be compiled by Vela to run on the Ethos-U NPU can be found at: https://tfhub.dev/s?deployment-format=lite&q=quantized
Please see Vela External APIs.
Please see Vela Contributions.
Please see Vela Debug Database.
Please see Vela CLI Options. This includes a description of the system configuration file format.
Please see Vela Performance Estimation Summary.
Please see Vela Releases.
Please see Vela Security.
Please see Vela Supported Operators for the list of operators supported in this release.
Please see Vela Testing.
Additional useful information:
- Arm Products: Ethos-U55 NPU
- Arm Products: Ethos-U65 NPU
- Arm Developer: Ethos-U55 NPU
- Arm Developer: Ethos-U65 NPU
Vela is licensed under Apache License 2.0.
Release history Release notifications | RSS feed
Download the file for your platform. If you're not sure which to choose, learn more about installing packages.