Lesson 1: Deep Learning 2019 - Image classification

The key outcome of lesson 1 is that we'll have trained an image classifier which can recognize pet breeds at state of the art accuracy. The key to this success is the use of *transfer learning*, which will be a key platform for much of this course. We'll also see how to analyze the model to understand its failure modes. In this case, we'll see that the places where the model is making mistakes is in the same areas that even breeding experts can make mistakes.

We also discuss how to set the most important *hyper-parameter* when training neural networks: the *learning rate*, using Leslie Smith's fantastic *learning rate finder* method. Finally, we'll look at the important but rarely discussed topic of *labeling*, and learn about some of the features that fastai provides for allowing you to easily add labels to your images.

[_luca_peric_]

Timestamps:

00:00 - 12:25 Introduction to Jupyter, Jeremy's background, world class applications after 7 weeks, false statements.
12:25 - 17:55 Jumping into the Jupyter notebook.
17:55 - 48:30 Oxford-IIT Pet Dataset, untar_data, get_image_files & regex, ImageDataBunch, image size, normalising images, inspecting data, downloading weights and validation sets.
48:30 - 1:05:50 Success stories about alumni & fast.ai.
1:05:50 - 1:14:45 Why resnet, saving a model, analysing results with a) top_losses, b) confusion_matrix & c) most_confused.
1:14:45 - 1:38:15 Finetuning with unfreezing, visualising CNN layers, learning rate, lr_find, max_lr, loading models, different kinds of datasets, fast.ai docs practical examples available as Jupyter notebooks.
1:38:15 - 1:40:11 More examples of interesting stuff that can be done.

Thanks again for sharing!

[StefanYouCan]

I just finished a Master's in Software Eng... I much prefer the way things are taught in this course and I am finally learning what I wanted to during my education! Following this course during the confinement!

[unoqualsiasi]

This level of education is a rare pearl. Thank you for the effort you put into it!

[gardnmi]

I found this video after spending a week reading through tensorflow. . I am someone who just wants to use machine learning as a tool like a hammer for some projects and fastai seems exactly that. I can't wait to see what this becomes

[SamimEkram]

The best course for learning Deep Learning free and fast thanks for this course Jeremy

[VijayV-zwtj]

Really good explanation. Better than V2. Sad that the audio quality at the end went bad.

[kevalan]

"all our doc is also code" -- f'ing awesome!

[privatecustomer]

I really appreciate the top-down approach as it enables you to see results faster which keeps you motivated. Just the perfect fit for people who are on the practitioner side of things. Keep it up!

[evgeniymamchenko]

Another simple way to try transfer learning for image classification on your task even without writing any code is an Android app called Pocket AutoML. It trains a model right on your phone so it can even work offline.

[lingobol]

For all those getting ConvLearner not defined error:
ConvLearner has been replaced by cnn_learner.

[dibyaranjanmishra]

I have not watched the 2018 one. Just wondering if anyone has seen both can give some insight on how much this one is different than the other one

[mukeshjha]

If someone encounters error like Name ConvLearner() not defined, then use cnn_leaner(). Earlier u could have also used create_cnn().

[gdoteof]

Been waiting for these. Thanks so much!

[aronhighgrove]

Is there a version with better audio? The compression creates many high pitch artifacts, and is pretty hard/uncomfortable to listen to in general.

[gdoteof]

Sadly the audio is pretty bad / hard to listen to.

[girish]

your course is also valuable course like sentdex !!

[kawingchan]

1:02:49 About Splunk, if a fraud detection problem is recast as an image recognition, could this suffer from adversarial attack? I.e. do whatever bad deeds bad guys do, then just do one little thing, that result in a few pixel or noise added, such that the network is fooled. Would be interested for security and adversarial researchers to find out.

[toequantumspace]

Jeremy I'm so happy you are back!

[Finite]

This is amazing. Thanks for sharing your knowledge

[girish]

great explanation of deep learning !!!!