Résumé du cours
This course will give you hands-on experience optimizing, deploying, and scaling a variety of production ML models. You’ll learn how to build scalable, accurate, and production-ready models for structured data, image data, time-series, and natural language text, along with recommendation systems.
A qui s'adresse cette formation
- Data Engineers and programmers interested in learning how to apply machine learning in practice
- Anyone interested in learning how to leverage machine learning in their enterprise
Certifications
Cette formation prépare à la/aux certifications:
Pré-requis
To get the most out of this course, participants should have:
- Knowledge of machine learning and TensorFlow to the level covered in Machine Learning on Google Cloud coursework
- Experience coding in Python
- Knowledge of basic statistics
- Knowledge of SQL and cloud computing (helpful)
Objectifs
This course teaches participants the following skills:
- Implement the various flavors of production ML systems—static, dynamic, and continuous training; static and dynamic inference; and batch and online processing
- Solve an ML problem by building an end-to-end pipeline, going from data exploration, preprocessing, feature engineering, model building, hyperparameter tuning, deployment, and serving
- Develop a range of image classification models from simple linear models to high-performing convolutional neural networks (CNNs) with batch normalization, augmentation, and transfer learning
- Forecast time-series values using CNNs, recurrent neural networks (RNNs), and LSTMs
- Apply ML to natural language text using CNNs, RNNs, LSTMs, reusable word embeddings, and encoder-decoder generative models
- Implement content-based, collaborative, hybrid, and neural recommendation models in TensorFlow
Outline: Advanced Machine Learning with TensorFlow on Google Cloud Platform (MLTF)
Module 1: Machine Learning on Google Cloud Platform
- Effective ML
- Fully Managed ML
Module 2: Explore the Data
- Exploring the Dataset
- BigQuery
- BigQuery and AI Platform Notebooks
Module 3: Creating the Dataset
- Creating a Dataset
Module 4: Build the Model
- Build the Model
Module 5: Operationalize the Model
- Operationalizing the Model
- Cloud AI Platform
- Train and Deploy with Cloud AI Platform
- BigQuery ML
- Deploying and Predicting with Cloud AI Platform
Module 6: Architecting Production ML Systems
- The Components of an ML System
- The Components of an ML System: Data Analysis and Validation
- The Components of an ML System: Data Transformation + Trainer
- The Components of an ML System: Tuner + Model Evaluation and Validation
- The Components of an ML System: Serving
- The Components of an ML System: Orchestration + Workflow
- The Components of an ML System: Integrated Frontend + Storage
- Training Design Decisions
- Serving Design Decisions
- Designing from Scratch
Module 7: Ingesting Data for Cloud-Based Analytics and ML
- Data On-Premises
- Large Datasets
- Data on Other Clouds
- Existing Databases
Module 8: Designing Adaptable ML Systems
- Adapting to Data
- Changing Distributions
- Right and Wrong Decisions
- System Failure
- Mitigating Training-Serving Skew Through Design
- Debugging a Production Model
Module 9: Designing High-Performance ML Systems
- Training
- Predictions
- Why Distributed Training?
- Distributed Training Architectures
- Faster Input Pipelines
- Native TensorFlow Operations
- TensorFlow Records
- Parallel Pipelines
- Data Parallelism with All Reduce
- Parameter Server Approach
- Inference
Module 10: Hybrid ML Systems
- Machine Learning on Hybrid Cloud
- KubeFlow
- Embedded Models
- TensorFlow Lite
- Optimizing for Mobile
Module 11: Welcome to Image Understanding with TensorFlow on GCP
- Images as Visual Data
- Structured vs. Unstructured Data
Module 12: Linear and DNN Models
- Linear Models
- DNN Models Review
- Review: What is Dropout?
Module 13: Convolutional Neural Networks (CNNs)
- Understanding Convolutions
- CNN Model Parameters
- Working with Pooling Layers
- Implementing CNNs with TensorFlow
Module 14: Dealing with Data Scarcity
- The Data Scarcity Problem
- Data Augmentation
- Transfer Learning
- No Data, No Problem
Module 15: Going Deeper Faster
- Batch Normalization
- Residual Networks
- Accelerators (CPU vs GPU, TPU)
- TPU Estimator
- Neural Architecture Search
Module 16: Pre-built ML Models for Image Classification
- Pre-Built ML Models
- Cloud Vision API
- AutoML Vision
- AutoML Architecture
Module 17: Working with Sequences
- Sequence Data and Models
- From Sequences to Inputs
- Modeling Sequences with Linear Models
- Modeling Sequences with DNNs
- Modeling Sequences with CNNs
- The Variable-Length problem
Module 18: Recurrent Neural Networks
- Introducing Recurrent Neural Networks
- How RNNs Represent the Past
- The Limits of What RNNs Can Represent
- The Vanishing Gradient Problem
Module 19: Dealing with Longer Sequences
- LSTMs and GRUs
- RNNs in TensorFlow
- Deep RNNs
- Improving our Loss Function
- Working with Real Data
Module 20: Text Classification
- Working with Text
- Text Classification
- Selecting a Model
- Python vs Native TensorFlow
Module 21: Reusable Embeddings
- Historical Methods of Making Word Embeddings
- Modern Methods of Making Word Embeddings
- Introducing TensorFlow Hub
- Using TensorFlow Hub Within an Estimator
Module 22: Recurrent Neural NetworksEncoder-Decoder Models
- Introducing Encoder-Decoder Networks
- Attention Networks
- Training Encoder-Decoder Models with TensorFlow
- Introducing Tensor2Tensor
- AutoML Translation
- Dialogflow
Module 23: Recommendation Systems Overview
- Types of Recommendation Systems
- Content-Based or Collaborative
- Recommendation System Pitfalls
Module 24: Content-Based Recommendation Systems
- Content-Based Recommendation Systems
- Similarity Measures
- Building a User Vector
- Making Recommendations Using a User Vector
- Making Recommendations for Many Users
- Using Neural Networks for Content-Based Recommendation Systems
Module 25: Collaborative Filtering Recommendation Systems
- Types of User Feedback Data
- Embedding Users and Items
- Factorization Approaches
- The ALS Algorithm
- Preparing Input Data for ALS
- Creating Sparse Tensors For Efficient WALS Input
- Instantiating a WALS Estimator: From Input to Estimator
- Instantiating a WAL Estimator: Decoding TFRecords
- Instantiating a WALS Estimator: Recovering Keys
- Instantiating a WALS Estimator: Training and Prediction
- Issues with Collaborative Filtering
- Cold Starts
Module 26: Neural Networks for Recommendation Systems
- Hybrid Recommendation System
- Context-Aware Recommendation Systems
- Context-Aware Algorithms
- Contextual Postfiltering
- Modeling Using Context-Aware Algorithms
Module 27: Building an End-to-End Recommendation System
- Architecture Overview
- Cloud Composer Overview
- Cloud Composer: DAGs
- Cloud Composer: Operators for ML9
- Cloud Composer: Scheduling
- Cloud Composer: Triggering Workflows with Cloud Functions
- Cloud Composer: Monitoring and Logging