** I have published the contents of the tutorial in this article so far, but since it will be large and it will be troublesome to update twice, please refer to the note on Github for details. I think that would be better because I think I can follow the latest content. In addition, we will continue to inform you about updates on Github on this page. ** **
The repository is here If you want to see only the contents of the tutorial, please click here If you find any errors in the tutorial content, please let us know in the comments on the article or on the issue page on Github.
--CIFAR-10 Tutorial
: Contents of this article
--CIFAR-10 Tutorial [Appendix]
: A little tip
--CIFAR-10 Tutorial [Transfer Learning]
: About transfer learning using CIFAR-10. I might write an article someday.
--How to use the dlt package /
: A brief explanation of how to use the dlipr package using Fashion-MNIST
--results /
: Save the output ofCIFAR-10 Tutorial
--ʻImages / : Save example images and other images --
-tltag:
CIFAR-10 tutorial [transfer learning]` results obtained
--2018/02/03: The tutorial has been arranged so that it can be viewed as a website without cloning from Github.
dlipr
to dlt
(Deep Learning Tools)
--2018/01/19: Fixed a bug related to confusion matrix.
--2018/01/18: Deleted the description about the output of confusion matrix because the output of confusion matrix has an error caused by np.histogram2d
of the dlipr package. I changed the note structure etc. on Github. The appendix has been made independent of this paper to make it easier to implement.
--2017/11/28: Changed from the method of reading Keras method from tensorflow to the form of importing directly from Keras.
--2017/07/19: The update of the note of github has been collected, so I reflected it in Qiita. I added an explanation that seems to be an explanation, but it may be a little confusing. I'll fix it someday.
--2017/07/17: Fixed the part that could be confusing when looking at the contents of the data.
--2017/06/20: In order to save the complicated preparation for the dlipr file, it has been newly released so that it can be executed under any environment (which meets the operating environment). Please click here.This tutorial deals with an image classification problem called CIFAR-10. The CIFAR-10 consists of 50,000 32x32 pixel color image data and 10 labels that classify it: airplanes, cars, birds, cats, deer, dogs, frogs, horses, boats, trucks. .. There is also a CIFAR-100, which is classified into more subdivided labels, and this one has 100 labels. Both are basically the same, so this time I will use CIFAR-10.
Also, use Keras for implementation. Keras cannot be used for unsupervised learning, but I think Keras is sufficient for dealing with typical machine learning problems that represent CIFAR-10 and MNIST. What's more, Keras is very easy to implement for simple networks (not just one-way networks, but branching things, RNNs, etc.), so even beginners (rather than raw Tensorflow) I think it's easy to understand what you're doing. This tutorial also explains how to implement Keras (although it's easy), so even if you don't know Keras, you can read it. If you get stuck, please refer to the Official Document.
This tutorial is based on what I learned in a university lecture, so I added extra explanations where it took me a long time to understand. I can't deny the overall redundant feeling, but I think that leaving it on purpose will be useful for people with similar problems.
--What not to do ――What is machine learning? --About Tensorflow --Improve accuracy (hyperparameter adjustment method, etc.)
Recommended Posts