I realise now that it was quite clever to have apples and oranges in the previous notebook: 2-Linear regression because I now can understand better why the shape is 10, 784. For every picture the weights are calculated ten times, one for each output case since they are the probability of the number being recognised correctly. Well done
recap of last lecture (lecture one)
instructor shows where lecture notebooks are in the forums, first one will be "03-logistic-regression"- instructor recommends you to fork the notebook to have easier, quicker access to working with the notebook- rather than Binder, Kaggle is better for working with larger datasets, so, instructor recommends you use Kaggle instead for working with this notebook
starts
about Kaggle platform, that you need a free account
RECIPE (first time only) how to open & run Jovian notebook on Kaggle- run > on kaggle- language > python- type > notebook- create- (first time only) upper right corner sidebar > internet > on- note: once your notebook fully loads up in Kaggle, it should now have the text "-live" appended onto its name
preface to lecture (comments on setup)
actual START of lecture two content with notebook "03-logistic-regression-live" and "image classification using logistic regression in pytorch"
instructor notes that we will use existing knowledge of pytorch and linear regression to solve a different problem (image classification)
about the training dataset MNIST (grayscale handwritten digit images, considered a good dataset with whihc to get started in machine learning)
instructor notes that "system setup" heading can be skipped for now since the instructor's notebook is already running
firstly, we begin by "exploring the data"
note that if you run this notebook locally, you may need to uncomment the "!conda install" or "!pip install" commands to get the necessary code libraries (these come already installed though on Kaggle)
today's lecture will include torchvision, which continues "utilities" for working with image data and "helper classes" to automatically download and import popular datasets, such as MNIST
instructor donwloads the training dataset
instructor recommends "you should read up about Yann Lecun" as "one of the forefathers of deep learning"
instructor notes that "a bunch of files were downloaded but we don't need to worry about them just yet", the important takeaway is that we now have a "dataset" object to work with
instructor inspects the dataset object alone, and then with `len(dataset)`
instructor shows how to get out the "test dataset" from the MNIST class [constructor] (which is needed to determine the accuracy of your ML model)
instructor inspects the first element of the PyTorch dataset
instructor breaks down PyTorch dataset element shape (structure) as a PIL.Image with size dimensions and the number five. the number five here indicates that this image is showing/displaying the number five. in other words, this is a "supervised learning problem" where the training data has already be labeled (or we can simply say it contains labels - there's not necessarily a guarentee that they are perfectly correct or complete).
instructor explains that PIL stands for the the Pillow Imaging Library
we can visualize images in our notebook directly using the "cell magic" `%matplotlib inline` (which probably requires an inport)
let's inspect a couple images from the dataset
instructor notes that what is rendered in not exactly an image per se but rather a plot of pixles (colored cells)
"pytorch does not how to work with images", so, we convert the images into Tensors using the `ToTensor()` transform
![⌨️ () Intro - Learn Python Backend Development by Building 3 Projects [Full Course]](https://img.youtube.com/vi/ftKiHCDVwfA/mqdefault.jpg)
