[GCP ML engineer certification Day6]

Machine Learning with Tensorflow on Google Cloud Platorm (Coursera Lecture)

Lecture preview

Why Google

• There are over 4,000 Tensorflow machine learning modesls in production at Google

Why Google Cloud

• To be successful at ML, you need to think not just about creating models, but also about serving out ML predictions

• On GCP, the key services are all serverless and they’re all managed infrastructure
• By building data pipelines in Google Cloud, you essentially get to take advantage of the scalability, reliability, and sheer engineering progress that Google brings to running machine learning systems

Review Question

1. What is a common reason for an ML model that works well in training but fails in production?
   • The ML dataset was improperly created
2. Personalized Algorithms are often built using which type of ML model?
   • Recommendation systems
3. What is a key lesson Google has learned with regards to reducing the chance of failure in production ML models?
   • Process batch data and streaming data the same way

AI VS Machine Learning

• AI is a discipline, it has to do with the theory and methods to build machines think & act like humans
• Machine Learning is a toolset, so we can use machine learning to solve certain kinds of AI problems

Two stages of AI

1. Train an ML model with examples

2. Inference (Predict with a trained model)

ML in Google product

• ML is part of pretty much every Google product
• In practice, you will have to build many ML modles to solve one problem
• EX. How to forecast whether an item will go out of stock
  • First model to predict demand for the product at the store location
  • Second model to predict the inventory of this item at the supplier’s warehouses and nearby tores
  • Third model to predict how long it is going to take them to stock the product and use this to predict which supplier you will ask to refill, and when
  • etc…

Review Question

1. The main stages of Machine Learning models are?
   • Train an ML model
   • Predict with a trained model
2. What are common mathematical models used in Machine Learning?
   • Linear methods
   • Decision Trees
   • Radial basis functions
3. In the past, why did neural networks models have just a few layers?
   • Neural networks with lots of layers takes a lot of computing power
   • As you add more layers, there are more weights to adjust, and you need lots more data available to make those adjustments
   • If you just add layers, you may run into issues, for example some of the layers may become all zero or blow up and become NAM (not a number)
4. What are the models included in Google Translate app?
   • Find the sign
   • Read the sign
   • Detect the language
5. What is the smart reply feature of Inbox and Gmail?
   • The email program suggests three possible responses to received emails

Pre-trained models

• There are variety of domains where Google exposes ML services trained with their own data
• Two ways that Google Cloud Platform can help you add Machine learning to your applications
  • tools to help you build custom machine-learning models
    • Tensorflow lets you build and train your own machine learning models, using your own data
    • If you want to run tensorflow models on managed Google infrastructure, use Cloud Machine Learning Engine
  • friendly machine learning (pre-trained APIs)
    • give you access to pre-trained machine learning models, with a single REST API request

Vision API in action

• Cloud Vision is an API that lets you perform complex image detection with a single rest API request
• label and Web detection
  • tell you what is this a picture of
  • search for similar images across the web and extract content from the pages where those images are found to return additional details on your image
• OCR (Optical Character Recognition)
  • extract text from the images
  • tell you where that text was found
  • tell you what language that text is in
• Logo Detection
  • identify company logos in an image
• Landmark Detection
  • tell if an image contains a common landmark
  • provide the latitude-longitude coordinates of that landmark
• Crop hints
  • crop your photos to focus on a particular subject

Video Intelligence API

• Label detection
  • tells you what’s in this video
  • At a high level, it’ll tell you what is your video about
  • In a more granular level, it can tell you what’s happening in every scene of your video
• Shot change detection
  • if your video changes from a landscape pan to a close-up of a person doing an interview, it’ll give you the timestamps every time the camera changes shots
• Explicit content detection
  • identify inappropriate scenes in your video
• Regionalization
  • specify the region where your video API requests should be executed

Cloud Speech-to-Text API

• Cloud speech is an API that let’s you perform speech to text transcription in over 100 languages
• Speech-to-text transcription
  • lets you pass it in audio file and it returns a text transcription of that file
• Speech timestamps
  • it will return the start and end time for every word in your audio transcription, which makes it really easy to search within your audio
• Profanity filtering
• Perform either batch or streaming transcription

Cloud Translation API

• API that let’s you translate text into over 100 different languages
• Translate text
• Detect the language of your text

Cloud Natural Language API

• Cloud Natural Language is an API that let’s you understand texts with a single rest API request
• Extract entities
• Detect sentiment
• Analyze syntax
• Classify content

Cloud Text-to-Speech API

• Cloud Text-to-Speech is an API that lets you convert text into human-like speech
• pass it to text file and it returns raw audio data as a base64 encoded string
• You must decode this base64 encoded string to an audio file before an application can play it
• creates raw audio data of natural human speech
• access more than 180 voices across more than 30 languages in variants

DialogFlow

• used for building conversational interfaces
• analyze text or audio and respond to a human in a natural chatty way
• an authoring platform, not just an API, but it’s possible to manage your chat bot through its API

Review Question

1. Which of the following is not a pre-trained machine learning model on goole cloud?
   • Tensorflow
2. Which API lets you perform complex image detection with a single REST API request?
   • Cloud Vision API
3. Which API lets you understand your video’s entities at shot, frame or video level?
   • Cloud video intelligence API
4. What are the benefits of cloud speech-to-text API?
   • perform speech-to-text transcription
   • supports speech timestamps
   • supports profanity filtering
5. What type of actions can be done by Cloud Natural Language API?
   • Gives you the overall sentiment of a sentence or a text document

Manual data analysis

Typical customer journey involves going from manual data analysis to ML

• if you’re doing manual data analysis, you probably have the data already
• if you cannot analyze your data to get reasonable inputs towards making decisions, then there’s no point in doing machine learning
• to build a good machine learning model you have to know your data

Training and serving skew

• Problem
  • unless the model sees the exact same data in serving, and as it was used to seeing during training, the model predictions are going to be off
  • the result of stream processing and the result of branch processing have to be the same
• Solution
  • take the same code that was used to process historical data during training and reuse it during predictions
  • data pipelines have to process both Batch and Stream

ML phrases

• MLOps
  • How do you reduce the time between analyzing their problem, creating the models, and deploying the solution, while maintaining the quality of the output?
  • lifecycle management discipline for machine learning

Review Question

1. What would you use to replace user input by machine learning?
   • Pre-trined model
2. Which of the following refers to the type of data used in ML models?
   • Labeld data, Unlabeled data
3. Which of the following are best practices for Data preparation?
   • Avoid training-serving skew
   • Avoid target leakage
   • Provide a time signal
4. Which of the following is not part of the ML training phase?
   • Connecting Neural Networks
5. What’s the most efficient way to transcribe speech?
   • Use Speech API