Train an Image Classifier with TensorFlow for Poets – Machine Learning Recipes #6

Monet or Picasso? In this episode, we’ll train our own image classifier, using TensorFlow for Poets. Along the way, I’ll introduce Deep Learning, and add context and background on why the classifier works so well. Here are links to learn more, thanks for watching, and have fun!

TensorFlow for Poets Codelab:

Google’s Udacity class on Deep Learning:

TensorFlow tutorial:

Google Research blog on Inception:

You can follow me on Twitter at for updates on episodes, and of course – Google Developers.

Subscribe to Google Developers: –
Subscribe to the brand new Firebase Channel:
And here’s our playlist:


  1. Vic N says:

    Could you make tutorial how do i train natural language classifier to
    retrieve information: i have a bunch of banks products and want to
    categorize posts by problems with that products, i.e. bank credit 1
    problem, mobile app glitch 1, bad customer support etc. This involves
    manual ontology making, and then complex features engineering. How to avoid
    Also, i dont want involve people to manually add fetures to post, because
    they are multidimensional and for some text it can be 10-20 features. So
    this is hard and nearly the same as making rule-based approach

  2. Hitesh Vaidya says:

    is there a way through which, the more number of images the classifier
    comes across even after training, the more accurate it becomes. That means
    it should go on adding new data to its training data.

  3. Otto Fazzl says:

    I have a question. I checked out topology of Inception-v3 ANN and it looks
    rather complex with many layers and nodes of different types. So the
    question is: does this complex structure have a theoretical motivation of
    why it is what it is or Inception-v3 was constructed by trial-and-error or
    maybe other computational optimization techniques? If I am building my own
    ANN for a particular purpose, how do I know which topology will provide
    best results?

  4. vijayenthiran subramaniam says:

    Hey Josh, Great tutorial, 5:50 you told that if we give a picture if it is
    not a flower the confidence will be low. If I give a picture of a flower
    which is not among the five flowers which we have trained, the predicted
    confidence will be high right. How to know that the flower is not among the
    five flowers?

  5. Bishshoy Das says:

    For people who are saying he is overacting, I doubt you guys know how
    satisfying it is to work in Machine Learning. Especially writing code,
    making new discoveries and sometimes inventing some new stuff that goes
    into top journal publications. Then combine all these with working at
    Google. It is a miraculous feat to achieve and such achievers do smile from
    their heart.
    …from a developer at Nvidia.

  6. Enio F says:

    Few nuggets I found that might help:
    * Tensorflow Docker container works on windows (tested on windows 10)
    * You can use the original inception model to tell whats in an image it has
    1000 classes


  7. KengLeong Wong says:

    can you create a video for a simple content recommendation engine using
    tensor flow?

  8. Mohamed Ezzaouia says:

    Hi Josh,

    Congratulation for this wonderful series of videos!
    I would appreciate so much if you could create some episodes about ML
    classification for speech or signal processing.

    Thank you so much in advance and have a nice weekend :)

  9. Namai Toidi says:

    Be aware of the provided training data. I removed at least 10 images in the
    roses folder, since they cannot be considered as roses.


    Please make a video on installation of tensorflow….. i m suffering from
    installation…. somethings it shows it show error like incompatible or not
    supported….. please make a video on this

  11. M0481 says:

    I just wanna say thanks to Josh and his team for the effort they put into
    creating videos related to ML! What I really liked about this video is that
    you me the oppurtonity to continue learning by clicking on some of your
    referenced links.

    Quick question: do you know if a good book or maybe source that explains
    the basics and thought processes behind Machine Learning?


  12. TheZaezee says:

    Hi Josh! beginner question(probably a stupid one =P): You said that if
    you try to classify the roman colissium, it must say it is a flower, but
    hopefuly the confidence level would be low. If i am usying a softmax
    function in the outlayer it would give me the probabilities of the
    colissium in each flower class. How do i measure the confidence level?

  13. Niko Lim says:

    I tried to do this exercise again but unfortunately I am unable to complete
    because the git pull command in docker no longer works which makes it
    impossible to run now because you need the files from the git pull to make
    this classifier to work. At this point I’ve tried a couple different
    solutions and none of them have worked. Before docker had there toolbox I
    got tensorflow for poets to work. But now anytime docker tries to pull or
    clone from git hub it tosses back an error saying that it can’t find the
    directory. Even though I can ping the git hub directory from within the
    docker environment. If anyone has come up with a solution to this please
    let me know. but as of right now the Google Codelab tutorial exercise for
    Tensorflow classifiers currently does not work.

  14. randerson112358 says:

    Could you put a link on here to get/use the file that’s used
    to label the image please?

  15. Robert Shaver says:

    Is there any documentation on the categories which Inception is trained on?
    I imagine using Inception is limited to those categories. Is that true? (I
    can’t find much info about Inception.) I guess I’m looking for guidance as
    to whether or not to train from scratch or use Inception.

  16. chakree ten says:

    hello, I have seen a lot of videos on tensorflow and its possibilities now
    the point is I’ve been trying to install it on my windows laptop for almost
    a week, done some hacking at last no result, if the idea was to provide
    machine learning for everyone how come this installation is such pain

  17. Anuj Dutt says:

    Hi. Thanks for this great tutorial. I want to know one thing. Since, in
    this tutorial we train our own classifier on our own dataset, how can I
    download this classifier, say, as a pickle file to use it in a python code.
    Like I can write a code and call this trained classifier and use it to test
    for new test images. Thanks :)

Comments are closed.