Model Building with TensorFlow

A talk that introduces usage of TensorFlow library via building a simple model for image recognition. This presentation was for the IWD extended event at the Women TechMakers Nairobi group.

Transcript

1. Model Building with TensorFlow By: Rukayat Sadiq
2. Whoami… • Software Engineer at Andela, InVisionApp. • Postgraduate Student at Georgia Institute of Technology specialising in Machine Learning. • Problem Solver. • Polyglot.
3. Learning Objectives • Introducing TensorFlow • What makes TensorFlow better? • Applications of TensorFlow • Hands-on session with TensorFlow
4. TensorFlow is… • an open source library developed by Google for numerical computation • a library that allows creation of large scale neural networks with multiple layers • mainly used for classification, perception, understanding, discovery and prediction problems.
5. Why TensorFlow? • It uses data flow graphs, making it more efficient to compute all required operations at once. other libraries do single operations at a time. • It automatically calculates gradients needed to optimise graph variable for improved model performance.
6. Why TensorFlow? (cont’d) • Runs faster by leveraging CPUs, GPUs and more recently special chips called TPUs - TensorFlow Processing Units
7. Uses cases… • Classification Problems e.g image recognition • Voice/Sound Recognition e.g voice search, • Text Based Applications e.g sentimental analysis, threat detection • Time Series Analysis - trading predictions, recommenders systems Source: https://www.exastax.com/deep-learning/top-five-use-cases-of-tensorflow/
8. Notebook Link: https://github.com/Hvass-Labs/TensorFlow- Tutorials/blob/master/01_Simple_Linear_Model.ipynb Hands-on session: Model Building Disclaimer: I do not claim ownership for the tutorial used in this presentation. The tutorial is originally written by Magnus Pedersen. I have chosen not to re-invent the wheel since a basic tutorial exists to teach TensorFlow
9. A typical TF graph • Placeholder variables for graph inputs • Model variables: to be optimised for better model performance • Model: mathematical function calculating outputs based on inputs. • Cost measure: guides optimisation of variables • Optimisation method: changes the model variables
10. Placeholder Variables • Input images - x, float • true labels - y_true, float • true class value - y_true_cls, int
11. Model Variables • Weights • Biases
12. Model model = (x * weights ) + biases • Model Function • Normalized output of model using “softmax” function • Select index with highest class using “argmax” function
13. Cost Measure • Measure how well model performs with “cross-entropy”. • Calculate value with function “softmax_cross_entropy_with_logits” • Values range from 0 to positive values • 0 ==> perfect match • positive number ==> no perfect classification • Optimise model by changing weights and biases to minimise cross-entropy
14. Cost Measure (cont’d) • 0 ==> perfect match • positive number ==> imperfect classification with varying degree of inaccuracies • Optimise model by changing weights and biases to minimise cross-entropy
15. Optimisation • Optimise with Gradient Descent to obtain minimal cross-entropy. • Use “GradientDescentOptimizer” function from TensorFlow library.
16. Measuring Performance • Accuracy Acc = correct classification / total input images
17. Running TensorFlow • Create a session • Initialise all variables • Perform optimisation • Measure Performance session = tf.session()
18. Questions? Thank you for Listening
19. ruqoyyasadiq @ruqoyyasadiq Contact