Snips' tiny TensorFlow and ONNX inference engine.
This project used to be called tfdeploy, or Tensorflow-deploy-rust.
tract is a tensorflow- and ONNX- compatible inference library. It loads a
Tensorflow or ONNX frozen model from the regular protobuf format, and flows
data through it.
This is a semi-experimental support for real-time applications like voice processing. In many real time voice applications, processing must happen "as you go". One can not wait for the end of the incoming audio signal to start decoding.
While Kaldi has built its inference engine around this streaming constraint,
our approach to the same issue is a bit different. tract graph analyser and
optimiser will reason on "streamed" tensors, in order to generate an equivalent
stateful "pulsing" network that will propagate small time slices ("pulses") of
data. This makes optimisation efforts on pulsing and "finite" tensor modes
mutually benefit each other.
Obviously, this conversion only makes sense for a subset of operators, so not all networks can be converted to a pulse network: for instance, an aggregation (like a SoftMax) on the time dimension can only be given a value when the signal has been processed up to the end.
As of today (October 2019), tract passes successfully about 85% of ONNX backends
tests. All "real life" integration tests in Onnx test suite are passing:
bvlc_alexnet, densenet121, inception_v1, inception_v2, resnet50, shufflenet,
squeezenet, vgg19, zfnet512.
The following operators are implemented and tested.
Abs, Acos, Acosh, Add, And, ArgMax, ArgMin, Asin, Asinh, Atan, Atanh, AveragePool, BatchNormalization, Cast, CategoryMapper, Ceil, Clip, Compress, Concat, Constant, ConstantLike, ConstantOfShape, Conv, Cos, Cosh, DequantizeLinear, Div, Dropout, Elu, Equal, Erf, Exp, Expand, EyeLike, Flatten, Floor, GRU, Gather, Gemm, GlobalAveragePool, GlobalLpPool, GlobalMaxPool, Greater, HardSigmoid, Hardmax, Identity, IsNaN, LRN, LSTM, LeakyRelu, Less, Log, LogSoftmax, MatMul, Max, MaxPool, Mean, Min, Mul, Neg, Not, Or, PRelu, Pad, ParametricSoftplus, Pow, QuantizeLinear, RNN, Reciprocal, ReduceL1, ReduceL2, ReduceLogSum, ReduceLogSumExp, ReduceMax, ReduceMean, ReduceMin, ReduceProd, ReduceSum, ReduceSumSquare, Relu, Reshape, Rsqrt, ScaledTanh, Scan, Selu, Shape, Shrink, Sigmoid, Sign, Sin, Sinh, Size, Slice, Softmax, Softplus, Softsign, Split, Sqrt, Squeeze, Sub, Sum, Tan, Tanh, ThresholdedRelu, Tile, Transpose, Unsqueeze, Where, Xor
We test these operators against Onnx 1.4.1 (operator set 9) and Onnx 1.5.0 (operator set 10).
Even if tract is very far from supporting any arbitrary model, it can run
Google Inception v3 and Snips wake word models. Missing operators are easy
to add. The lack of easy to reuse test suite, and the wide diversity of
operators in Tensorflow make it difficult to target a full support.
The following operators are implemented and tested:
Abs, Add, AddN, AddV2, Assign, AvgPool, BatchToSpaceND, BiasAdd, BlockLSTM, Cast, Ceil, ConcatV2, Const, Conv2D, DepthwiseConv2dNative, Div, Enter, Equal, Exit, ExpandDims, FakeQuantWithMinMaxVars, Fill, FloorMod, FusedBatchNorm, GatherNd, GatherV2, Greater, GreaterEqual, Identity, Less, LessEqual, Log, LogicalAnd, LogicalOr, LoopCond, MatMul, Max, MaxPool, Maximum, Mean, Min, Minimum, Mul, Neg, NoOp, Pack, Pad, Placeholder, Pow, Prod, RandomUniform, RandomUniformInt, Range, RealDiv, Relu, Relu6, Reshape, Rsqrt, Shape, Sigmoid, Slice, Softmax, SpaceToBatchND, Squeeze, StridedSlice, Sub, Sum, Tanh, Tile, Transpose, VariableV2
TensorFlow-Lite is a TensorFlow subproject that also focuses on inference on smaller devices. It uses a precompiler to transform a TensorFlow network to its own format. It only supports a subset of operators from TensorFlow though, and is only optimised for devices with Arm Neon support.
Tract supports a wider subset of TensorFlow operators, and has been optimised for CPU of the previous generation (ARM VFP), also targetting devices in the Raspberry Pi Zero family.
Long story short, TensorFlow and Onnx formats are good for designing and training networks. They need to move fast to follow the research field, tend to integrate new features and operators greedily. They also exhibit a high level of expressibity to make facilitate network design.
On the other hand, only a subset of operators and network features actually reach production, so systems running production network do not have to deal with so many operators. Furthermore, some information required for training can be stripped from the network before going to production for prediction.
NNEF tries to bridge the gap between training frameworks and inference by proposing a format dedicated to production and prediction.
Tract NNEF support is partial, and alpha level:
- tract_nnef can load and execute networks NNEF networks
- tract command line can translate networks from TensorFlow or ONNX to NNEF
- tract supports most of the NNEF specification, the most notable exception being the ROI operators and deconvolution
- tract needs to extend NNEF with other operators (or extend some operators semantics) in order to support the subset of ONNX and TensorFlow that tract supports.
These models among others, are used to track tract performance evolution as part of the Continuous Integration jobs. See .travis/README.md and .travis/bundle-entrypoint.sh for more information.
https://github.com/ARM-software/ML-KWS-for-MCU
ARM demonstrated the capabilited of the Cortex-M family by providing
tutorials and pre-trained models for keyword spotting. While the exercise
is ultimately meant for micro-controllers, tract can run the intermediate
TensorFlow models.
For instance, on a Rasperry Pi Zero, the "CNN M" model runs in about 70 micro-seconds, and 11 micro-seconds on a Raspberry Pi 3.
https://arxiv.org/abs/1811.07684
Snips uses tract to run the wake word detectors. While earlier models were
class-based and did not require any special treatment, tract pulsing
capabilities made it possible to run WaveNet models efficiently enough for a
Raspberry Pi Zero.
| Device | Family | TensorFlow-lite | tract |
|---|---|---|---|
| Raspberry Pi Zero | Armv6 VFP | 113s | 39s |
| Raspberry Pi 2 | Armv7 NEON | 25s | 7s |
| Raspberry Pi 3 | aarch32 NEON | 5s | 5s |
Notes:
- while the Raspberry Pi 3 is an Armv8 device, this bench is running on Raspbian, an armv6 operating system, crippling the performance of both benches
- there exists other benches on the internet that show better performance results for TensorFlow (not -Lite) on the Pi 3. They use all four cores of the device. Both TensorFlow-Lite and tract here have been made to run on a single-core.
One important guiding cross-concern: this library must cross-compile as easily as practical to small-ish devices (think 20$ boards).
- nearly complete ONNX support, and wraps it as a backend
- integrate other TF models to use as example, test and benches
- consider acting as kaldi backend
Note: files in the tensorflow/protos directory are copied from the
TensorFlow project and are not
covered by the following licence statement.
Note: files in the onnx/protos directory are copied from the
ONNX project and are not
covered by the following licence statement.
All original work licensed under either of
- Apache License, Version 2.0 (LICENSE-APACHE or http://www.apache.org/licenses/LICENSE-2.0)
- MIT license (LICENSE-MIT or http://opensource.org/licenses/MIT) at your option.
Unless you explicitly state otherwise, any contribution intentionally submitted for inclusion in the work by you, as defined in the Apache-2.0 license, shall be dual licensed as above, without any additional terms or conditions.