![How does a computer see?
RGB [0,255]Black & White [0,1]
1: white
0: black
1 2 3
2 3 4
3 4 5
4 5 6
5 6 7
1 2
2 3
3 4
4 5
5 6
Grayscale [0,255]
255: white
[1,254]: gray
0: black](https://image.slidesharecdn.com/machinelearningwithtensorflow2-191123185443/85/Machine-Learning-with-TensorFlow-2-29-320.jpg)
Machine Learning with TensorFlow 2
- 3. Agenda • What is TensorFlow? • Get into Machine & Deep Learning • Hands on: Computer Vision in TensorFlow • 🎁 • What Else to Know about TensorFlow? • Question & Answers
- 5. Summary • Open-source framework to perform Machine and Deep Learning • A backend for large numeric computations • Basic concepts are tensors and graphs • For CPU, GPU, TPU* • C++ in the backend https://www.tensorflow.org/ * Tensor Processing Unit
- 6. Tensors 1 2 3 4 5 1 2 2 3 3 4 4 5 5 6 1 1 2 3 2 3 4 3 4 5 4 5 6 5 6 7 Multi-dimensional arrays
- 7. • All TensorFlow computations are represented as graphs • Nodes are operations • Edges are tensors • First: construction of graph • Second: Executed in a session Flow graph C A B Add MatMul …
- 8. It’s All about Usability!
- 9. TensorFlow 2.0 • Full integration of Keras library as high-level API • API Cleanup • Functions instead of sessions • Eager execution for a pythonic programming style • Upgrade script provided on TF website
- 10. Under the Hood Keras Tensorflow CUDA / cuDNN GPU BLAS, Eigen CPU If TensorFlow is executed on CPU it comes with a tensor operation library called Eigen NVIDIA CUDA Deep Neural Network Library is extremely efficient on a GPU
- 11. Under the Hood Keras Tensorflow CUDA / cuDNN GPU BLAS, Eigen CPU If TensorFlow is executed on CPU it comes with a tensor operation library called Eigen NVIDIA CUDA Deep Neural Network Library is extremely efficient on a GPU
- 12. Development Ecosystem tf.data, numpy Loading and pre-processing Build, train & reuse models DeployDevelopment Tools Tensor Board TF Hub https://keras.io/
- 13. Use Cases Text Speech Image Audio Content Summarization Sentiment Analysis Classification Recognition Speech to Text Text to Speech Clustering Object Detection Gesture Recognition Facial Modelling Translation Clustering Video Generation Text Generation Audio Generation
- 14. Use Cases Text Speech Image Audio Content Summarization Sentiment Analysis Classification Recognition Speech to Text Text to Speech Clustering Object Detection Gesture Recognition Facial Modelling Translation Clustering Video Generation Text Generation Audio Generation
- 16. Get into Machine & Deep Learning
- 17. Machine Learning Extract knowledge from data Build a generalised model Supervised Learning Approach Goal Types Unsupervised Learning Reinforcement Learning
- 18. Unsupervised Machine Learning • Algorithms draw conclusions on the data without human input • Find hidden patterns and structures in data that are yet unknown • e.g. Clustering Supervised vs. Unsupervised Supervised Machine Learning • “Teaching the model with knowledge” • Provide algorithm with labelled training data (e.g. correct classes) and features (input data) • Classification and Regression
- 20. Example for Classification % Leather Heel Gender Shoe Type 45 TRUE Female Pumps 81 FALSE Male Sneaker 71 TRUE Female Sneaker 50 FALSE Female Pumps Training 65 FALSE Female ? Prediction
- 21. Model Training and Evaluation • Split data into training and test set • Run for pre-defined iterations (= epochs) • Test set to evaluate model • Model learns from errors and can be improved • Goal: Find the minim error between actual and predicted values (e.g. Accuracy) Training Test Prediction Score
- 22. Over- and Underfitting Overfitting • The model learns too much from the training data • No generalisation of model • Bad performance on test data Underfitting • Training data does not contain full information of whole dataset • Model cannot learn patterns • Bad performance on training and test data
- 24. Deep Learning = Neural Networks Input Layer Hidden Layers Output Layer Input data in form of features / tensors Output can be a class or a numeric value Weights Weights Weights
- 25. Training of a Neural Network X1 X2 X3 SUM Activation function SUM Threshold w1,1 w2,1 w3,1 Input Layer Hidden Layers Output Layer
- 26. Activation Functions Activation functions are attached to each neuron (node) and decide whether it should be activated (fired) or not Step Function Sigmoid ReLu
- 27. Loss Function & Optimizer • Train weights and evaluate prediction with actual values of test set • Goal: Find model with minimal error • Loss function: error function used for validation • Uses gradient descent approach • Optimizers: Algorithms that perform a gradient descent approach • Algorithm in Neural Networks: Backpropagation
- 28. Hands on: Computer Vision in TensorFlow
- 29. How does a computer see? RGB [0,255]Black & White [0,1] 1: white 0: black 1 2 3 2 3 4 3 4 5 4 5 6 5 6 7 1 2 2 3 3 4 4 5 5 6 Grayscale [0,255] 255: white [1,254]: gray 0: black
- 30. Black & White picture
- 32. RGB Picture
- 33. “What are the most popular datasets for cross-sectional data?” http://www.lac.inpe.br/~rafael.santos/Docs/CAP394/WholeStory-Iris.html https://de.wikipedia.org/wiki/RMS_Titanic
- 35. Convolutional Neural Network Keep spatial information of pixels • Low level features: edges • Medium level features: circles, squares • High level features: faces, bodies
- 36. Convolutional Neural Networks 10x10x1 10x10x9 5x5x9 Conv Layer Pooling LayerInput Flatten Layer 100x1
- 37. Convolutional (Filter) Layer Filter, kernel size, padding, stride
- 39. 🎁
- 40. Too sexy?
- 41. What Else to Know about TensorFlow?
- 44. For desktop, mobile, web
- 45. To Production Development Production https://www.tensorflow.org/tfx/tutorials/serving/rest_simple Data http request Save and serve model TensorFlow Extended
- 46. Play with Neural Networks https://playground.tensorflow.org/
- 47. Wanna Learn more? • Tutorials and docu at: tensorflow.org & keras.io • Udemy German course: “Deep Learning, Neuronale Netze und TensorFlow in Python” by Jan Schaffranek • Many more on Coursera,… • Data Science Competitions: kaggle.com
- 48. Thank you!