This is my notes for the coursera course TensorFlow in Practice Specialization, which talks about building scalable AI-powered algorithms in TensorFlow.
This specialization will include four main contents:
- Develop an understanding of how to build and train neural networks.
- Improve a network’s performance using convolutions. Train it to identify real-world images.
- Teach machines to understand, analyze, and respond to human speech with natural language processing systems.
- Process text, represent sentences as vectors, and input data to a neural network. Train an AI to create original poetry.
There are actually four courses:
-
Introduction to TensorFlow for Artificial Intelligence, Machine Learning, and Deep Learning
From looking at the most basic of neural networks to building a basic computer vision neural network that classified clothing. Then take this a little further by using Convolutions that spot features in an image, and then classify and learn based on those features. Use ImageGenerator to deal with complex real data. Train the ConvNet with fit_generator, combining trainging and validation.
1.1 A new programming paradigm
Introduction of difference between ML and traditional programming.
A hello world neural network program using TensorFlow.
1.2 Introduction to Computer Vision
Introduction of dataset Fashion MNIST.
Use this dataset to train a neural network to recognize clothes. Classify a picture of one piece of clothes into 10 categorites.
Exercise 2 is handwriting recognition using dataset MNIST.
1.3 Enhancing Vision with Convolutional Neural Networks
Introduce convolution and pooling. And how to implement them in code (add few layers before flattening layer).
Convolution: A technique to extract features from an image.
Exercise 3 is handwriting recognition using dataset MNIST with CNN.
Load around 1000 images about human or horses and label them using ImageGenerator. Feel how the ImageGenerator is pulling the images from the file system and feeding them into the neural network for training
Defining a complex ConvNet to use complex images which are colorful and have red green blue 3 channels, 3 bytes per pixel.
Use sigmoid instead of softmax as activation function since this is a binary classification problem.
Do prediction using this model once the model is trained.
Adding automatic validation to test accuracy.
Exercise 4: a happy or sad dataset which contains 80 images, 40 happy and 40 sad. Create a convolutional neural network that trains to 100%.
-
Understanding how ML works
-
using a DNN to do basic computer vision
-
beyond into Convolutions.
-
how to extract features from an image
-
the tools in TensorFlow and Keras to build with Convolutions and Pooling as well as handling complex, multi-sized images.
-
how overfitting can have an impact on your classifiers
-
explored some strategies to avoid it, including Image Augmentation, Dropouts, Transfer Learning and more.
-
moving towards multi-class classification
2.1 Exploring a Larger Dataset
Kaggle dataset of 25,000 cats versus dogs images.
Exercise 5: Use the 25k images to train a model. Need to handle directories yourself.
2.2 OVERFITTING: Augmentation: A technique to avoid overfitting
Deal with Overfitting: With a smaller dataset, you are at great risk of overfitting; with a larger dataset, then you have less risk of over-fitting, but overfitting can still happen.
Data augmentation:
Image augmentation, is rotation, skewing偏移, flipping, moving it around the frame, those kind of things. You're not changing the dataset. It all just happens in memory.
Another strategy, of course for avoiding overfitting, is to use existing models, and to have transfer learning.
2.3 Transfer Learning & OVERFITTING:Dropout.
Transfer Learning, lets you download the neural network, that maybe someone else has trained on a million images, or even more than a million images. So take an inception network, that someone else has trained, download those parameters, and use that to bootstrap your own learning process, maybe with a smaller dataset.
You can use their model directly, or just use the features they extracted.
Dropout to overcome overfitting.
2.4 Multicast learning. Multiclass Classifications
Go to multiclass Classification: Rock, Papaer, Scissors.