Towards Machine Learning in Pharo with TensorFlow
- 1. Towards Machine Learning in Pharo with TensorFlow Dr. Serge Stinckwich le logotype : les différentes formes en couleur version verticale version compacte SergeStinckwich
- 2. UMMISCO, an International Joint Research Unit about Modelling and Simulation of Complex System UCA UY1 UGB UCAD
- 3. System Biology System EcologySystem Sociology Epidemic control Sustainable fishing effortCi:es Pollu:on 390 million dengue infections/year UMMISCO build and develop tools&methods to be applied to southern countries issues
- 4. What is TensorFlow ? • A general purpose numerical computing library • Open-source software • Developed originally by researchers from Google Brain in 2015 • Written in Python, C++ • Hardware independent: CPU (Eigen/ BLAS), GPU (CUDA/CuDNN), …
- 5. TensorFlow basics • Tensors • Data Flows • Runtime execution
- 6. What are Tensors ? • Tensors are multi-dimensional arrays • TensorFlow supports: float16, float32, float64, bfloat16, complex64, complex128, int8, uint8, uint16, uint32, uint64, int16, int32, int64, bool, string, qint8, quint8, qint16, quint16, qint32 Rank Math entity 0 Scalar 1 Vector 2 Matrix (2-Tensor) 3 3-Tensor n n-Tensor
- 8. Applications of TensorFlow • Originally: Quantum Physics • Mainly: AI, Machine/Deep Learning • but not only: Data analysis, image processing, big data, simulation, BioInformatics, computational neurosciences
- 9. Why Pharo should take care about TensorFlow? • Multi-platform library • Distributed execution (CPU, GPU, TPU) • Robust foundation for doing machine learning and deep learning frameworks (but not only) • TF already support: C++, Python, Java, R, Go, JavaScript, Scala, … • Will be more fun to build live TF code.
- 10. What is the meaning of flow in TensorFlow ? • nodes = operations: math functions, constants (initializing values), logging data, … • edges = data (tensors) flows between nodes • Dataflows are defined in Pharo, compiled in memory with TensorFlow and then executed on devices (CPU, GPU, …)
- 12. libTensorFlow C API • TF main functionalities are exposed through a C layer • We use UFFI (Unified Foreign Function Interface) to connect Pharo to TensorFlow • https://github.com/tensorflow/tensorflow/blob/r1.8/ tensorflow/c/c_api.h
- 13. How to use C API TensorFlowCAPI>>version "TF_Version returns a string describing version information of the TensorFlow library. TensorFlow using semantic versioning." "TF_CAPI_EXPORT extern const char* TF_Version();" ^ self ffiCall: #(String TF_Version #()) module: TensorFlowCAPI TensorFlowCAPI uniqueInstance version. => ‘1.10.1’
- 14. libTensorFlow-bindings- Pharo • Fork from CUIS TensorFlow bindings (FFI not exactly the same). Thank you Javier Burroni&Gerardo Richarte. • Part of PolyMath project: https://github.com/ PolyMathOrg/libtensorflow-pharo-bindings • Works on TF 1.10 on MacOS, Linux and Windows. • More than 100 green unit tests • Crash until recently ! Still finalization issues
- 15. Build Tensors • Float pi asTensor. • TF_Tensor fromFloats: #(1 2 3 4). • TF_Tensor fromFloats: #((1 2 3 4 5)(6 7 8 9 10)).
- 16. Ranks, shapes • pisTensor := TF_Tensor fromFloats: #(3.14 3.1415 3.141516). • pisTensor rank. => 1 • pisTensor shape. => #(3) • (TF_Tensor fromFloats: #((1 2 3 4 5)(6 7 8 9 10))) shape.
- 17. Graphs • A Graph contains a set of Operation objects, which represent units of computation; and Tensor objects, which represent the units of data that flow between operations. • Graph can be serializable (as protobuff) and can be exchanged between platforms. • TF_Graph create.
- 18. Operations • Operation is a node in a TF Graph that takes 0, 1 or n tensors as inputs and produces 0, 1 or n tensors as outputs. • Example of TF operations : • Arithmetic operators: Add, Multiply, Mod, etc … • Mathematic functions: Sin, Exp, Pow, etc … • Matrix math functions: Transpose, Inverse, Norm, … • Reduction dimensions&Segmentation
- 19. Runtime execution • After the data flow has been defined, the graph is executed within a session and inside a device (CPU, GPU, …). • Distributed execution: graph can be splitted on many devices • Portable: dataflow graph is a language-independant representation of the code (could be re-used with another language)
- 20. Demos
- 21. 3+4 in TensorFlow graph := TF_Graph create. c1 := graph const: 'c1' value: 3.0 asTensor. c2 := graph const: 'c2' value: 4.0 asTensor. sum := c1 + c2. session := TF_Session on: graph. result := session runOutput: (sum output: 0). result asNumbers
- 22. Multiply two matrices graph := TF_Graph create. t1 := TF_Tensor fromFloats: #((1 2)(3 4)). t2 := TF_Tensor fromFloats: #((5 6)(7 8)). c1 := graph const: 'c1' value: t1. c2 := graph const: 'c2' value: t2. mult := c1 * c2. session := TF_Session on: graph. result := session runOutput: (mult output: 0). result asNumbers
- 24. TensorBoard
- 25. Conclusion • First alpha version of the TF bindings available on github • TF dataflows are difficult to debug because when they are executed this is outside Pharo (TensorFlow board) • Still some random crashes from time to time (mostly due to finalization)
- 26. Perspectives • More complex visualisations of data flows with Roassal (ongoing work of a M2 student) • Build a DSL on top of TF bindings to ease ML&DP models building (like Keras) • Use GlamourousToolkit & Moldable inspectors • Use TF bindings for applications (PolyMath, DataFrame, Kendrick, etc …)
- 27. Don’t miss • Thursday, 14pm30 : CORMAS, a participatory and interdisciplinary modeling platform, P.Bommel, E.Delay, C. Le Page, H. Morales, N. Becu, B. Bonte, N.Papoulias, S.Stinckwich, CORMAS Team • Friday, 11am : PolyMath, O. Zaytsev, S.Stinckwich