Home>

I'm about to start studying machine learning with tensorflow2.
(I have more than 10 years of experience in C ++, but I have only been using python for about a year after a blank of 5 or 6 years.)

The following issues should be solved by machine learning.
・ Classification of image files (I want to divide them into two, OK and NG)
・ All image files are the same size
・ There are not enough image files to learn (there are dozens of images)
・ I want to learn while adding image files generated at any time.
・ Occurrence timing About 1 sheet every 1 to 24 hours
・ Correct label can be attached to the image file to be added.

In other words
The image generated once every few hours is tentatively judged as OK/NG by machine learning, and the accurate judgment can be found after about 10 minutes, so I want to make a flow to make use of that image for the next judgment by additional re-learning. ..
(I think that tentative judgment is meaningless for machine learning, so it would be good if an image file that occurs at an arbitrary timing could be additionally relearned with a correct label attached.)

I tried to start studying with such a task, but I don't know where to start.
I think that the basics are supervised learning, but I think that the following keywords will be used to proceed while re-learning with images that occur sequentially, but I do not know if that is correct.
・ Transfer learning
・ Fine tuning

Transfer learning is based on the premise that there is learning data in advance, so I think it is different in this case.

It is a situation that is troubled before starting, which is common for beginners.
Any advice would be appreciated.

  • Answer # 1

    It is a situation that is troubled before starting, which is common for beginners.
    Any advice would be helpful

    If you have noticed it yourself, I think you should do it immediately.

    I'm about to start studying machine learning with tensorflow2.
    (I have more than 10 years of experience in C ++, but I have only been using python for about a year after a blank of 5 or 6 years.)

    That means that you have never performed image recognition.
    Then obviously

    Lots of rudimentary tutorials on image recognition in tensorflow
    A lot of image recognition related books in the introductory book

    Probably from the point of doing it (actually, I think it is better to train the legs by performing discrimination/regression with numerical values ​​/ category data in Python before tensorflow or image recognition, but it may be done later).

    Without this there is no transfer learning or fine tuning.
    Double play before playing catch, feint with 0 lifting
    There is no point in thinking about that.

    I don't know unless I ask for the details of the conditions
    Rather than transfer learning or fine tuning
    Online learning
    I think it is in the category.

    However, you shouldn't worry about what online learning is and don't do anything.